Office Action Analysis: 18529300 — MULTIPLIER TUNING POSTPROCESSING FOR MACHINE LEARNING BIAS MITIGATION

Office Action

§103
DETAILED ACTION
The action is in response to the original filing on December 05, 2023 and the Remarks and Amendments filed on July 20, 2026. Claims 1-6, 9-18, 20-23 are pending and have been considered below. Claims 1, 15, 17 and 23 are independent claims. Claims 1, 9, 10, 15, 17, 20 and 23 are amended and Claims 7, 8 and 19 are canceled.

Notice of Pre-AIA  or AIA  Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .

Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA  35 U.S.C. 102 and 103 (or as subject to pre-AIA  35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA  to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.  
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.

The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
	
Claims 1, 2, 9, 10, 11, 12, 17, 18, 20, 21 are rejected under 35 U.S.C. 103 as being unpatentable over Bhaila et al. (Bhaila), Non-Patent Literature, “Fair Collective Classification in Networked Data”, published on 2022, Pages: 10 in view of Non-Patent Literature of Deb et al. (Deb), Non-Patent Literature listed in IDS filed on 02/09/2024, “A Fast and Elitist Multiobjective Genetic Algorithm: NSGA-II Fairness and Calibration”, published on Apr 2002, Pages: 16 and in further view of Cruz et al. (Cruz), Non-Patent Literature listed in IDS filed on 02/09/2024, “FairGBM: Gradient Boosting with Fairness Constraints”, published on March 2023, Pages: 23.

As to independent Claim 1,
Bhaila teaches a method comprising:
inferring by a classifier that already was trained, from an input that contains a value of a feature, a probability of a class (
Bhaila, Pg1417, Left Col, P2, Lines1-4, “The local classifier f_L estimates label probabilities P(y_i | x_i) using only node attributes x_i… The classifier is trained V^k”
Pg1420, Right Col ,Subsection b), P2, Lines1-3, “Both of these postprocessing techniques can be applied over final CC predictions without modifying the local classifier, the relational classifier, and the inference method”
Pg1416-1417, Right Column, Subsection A. Collective Classification, Lines4-9, “The input is an unweighted, undirected, and attributed graph G = (V,E,X,Y) where V is a set of N nodes, E is a set of edges that connect node pairs in V, X represents the node feature matrix, and Y denotes node labels. Each node vi is defined by its feature vector xi which includes a sensitive attribute s_i”
Wherein input node v_i with feature vetor x_i includes a sensitive feature s and its specific value s_i, and the trained local classifier f_L initially infers label probabilities P(y_i | x_i) over target class labels without modifying or retraining f_L)

	generating a plurality of multipliers for the feature (
Bhaila, Pg1419, Left Col, P1, Lines1-4, “We propose to incorporate this reweighting technique in the collective classification model during each iterative update to assign weights to nodes for each possible combination of s and y”
Pg1419, Right Column, Paragraph1, Lines9-10, “wc,si is the weight due to node vi’s own sensitive attribute and the label to be estimated”
wherein Bhaila discloses assigning feature-specific multipliers wc,si based on a sensitive feature s and its specific value s_i, rendering it functionally equivalent to the claimed invention)

	selecting based on the value of the feature in the input, from the plurality of multipliers of the feature, a multiplier that is specific to both of the feature and the value of the feature (
Bhaila, Pg1419, Right Column, Paragraph1, Lines4-5, “We also apply the weight based on the node’s own sensitive attribute and predicted node label”
Pg1419, Left Column, Paragraph1, Lines14-16, “In our setup with binary s and binary y, we derive these weights for four combinations of s and y”
Pg1419, Right Column, Paragraph1, Lines9-10, “wc,si is the weight due to node vi’s own sensitive attribute and the label to be estimated”
wherein Bhaila teaches multiple multipliers to the sensitive feature s such that the multiplier or the weight is specific to the feature value s_i, so selecting and using the weight wc,si based on node v_i’s sensitive attribute value s_i is equivalent to the claimed invention)

	classifying the input based on a multiplicative product of the probability of the class and the multiplier that is specific to both of the feature and the value of the feature (
Bhaila, Pg1419, Left Column, Paragraph1, Lines1-4, “We propose to incorporate this reweighting technique in the collective classification model during each iterative update to assign weights to nodes for each possible combination of s and y”
Pg1419, Algorithm2 Fair Collective Classification via Node Reweighting, Line6, 
    PNG
    media_image1.png
    24
    203
    media_image1.png
    Greyscale

Pg1419, Equation8, 
    PNG
    media_image2.png
    46
    350
    media_image2.png
    Greyscale

wherein the summation part in the equation 8 shows that the feature-specific weight wc,si is multiplied by the probability of a class c, which corresponds to the multiplicative product. Also, this multiplication is incorporated into the classification model such that the algorithm 2 shows that 
    PNG
    media_image3.png
    31
    54
    media_image3.png
    Greyscale
 is the value obtained from the multiplication to classify the input with the final label 
    PNG
    media_image4.png
    33
    23
    media_image4.png
    Greyscale
)

	adjusting the multiplier (
Bhaila, Pg1419, Algorithm2, Line4,

    PNG
    media_image5.png
    72
    625
    media_image5.png
    Greyscale

Pg1419, Equation 6 and 7

    PNG
    media_image6.png
    63
    200
    media_image6.png
    Greyscale


    PNG
    media_image7.png
    52
    469
    media_image7.png
    Greyscale

Wherein Eq7 and Algorithm2 show that the algorithm iteratively derives or adjusts the weight 
    PNG
    media_image8.png
    28
    37
    media_image8.png
    Greyscale
 to 
    PNG
    media_image9.png
    30
    52
    media_image9.png
    Greyscale
 for feature value s_i (s = a)); and

reclassifying the input (
Bhaila, Pg1419, Algorithm2, Line6

    PNG
    media_image10.png
    31
    245
    media_image10.png
    Greyscale

Pg1417, Table1, 

    PNG
    media_image11.png
    20
    438
    media_image11.png
    Greyscale


    PNG
    media_image12.png
    21
    418
    media_image12.png
    Greyscale

Pg1420, Right Col ,Subsection b), P2, Lines1-3, “Both of these postprocessing techniques can be applied over final CC predictions without modifying the local classifier, the relational classifier, and the inference method”
wherein Algorithm2 shows that after updating node weight 
    PNG
    media_image9.png
    30
    52
    media_image9.png
    Greyscale
 and recalculating label probability vector 
    PNG
    media_image13.png
    27
    50
    media_image13.png
    Greyscale
, the updated probability vector is used to obtain a new predicted label 
    PNG
    media_image14.png
    27
    55
    media_image14.png
    Greyscale
for node v_i, which updates and assigns the target label prediction in iteration (t+1) without retraining the classifier f_L, so obtaining the new predicted label 
    PNG
    media_image14.png
    27
    55
    media_image14.png
    Greyscale
 from the adjusted probability distribution is functionally equivalent to the claimed reclassifying the input);
	
	wherein the method is performed by one or more computers without retraining the classifier (
Bhaila, Pg1421, Left Column, Subsection B. Compared Methods, Lines14-15, “All methods evaluated in this section are implemented on the basis of AIF360”
Pg1420, Right Col ,Subsection b), P2, Lines1-3, “Both of these postprocessing techniques can be applied over final CC predictions without modifying the local classifier, the relational classifier, and the inference method”
Pg1416, Right Col, Lines3-6, “Collective classification is even simpler than the C&S method discussed in [22] as it does not require training to optimize any parameters and directly implements label propagation based on certain assumptions on the graph structure”
wherein Bhaila explicitly teaches that all methods are implemented on the basis of AIF360 (an open-source toolkit running code on a computer system), which satisfied being performed by one or more computers. And further teaches that collective classification and post-processing adjustments directly perform label propagation and reweighting over predictions without requiring training to optimize parameters or modifying the local classifier, so executing the recalibration and reclassification without modifying or retraining the local classifier is functionally equivalent to the claimed invention.)

	However, Bhaila does not explicitly teach (highlighted parts)
Performing by a multi-objective optimizer: generating a plurality of multipliers for the feature, including from a validation, receiving two validation scores that are based on the plurality of multipliers of the feature

In the same field of endeavor, Deb teaches performing by a multi-objective optimizer: generating a plurality of candidates/multipliers (
Deb, Pg182, Abstract, Lines9-14, “Specifically, a fast nondominated sorting approach with O(MN^2) computational complexity presented. Also, a selection operator is presented that creates a mating pool by combining the parent and offspring populations and selecting the best (with respect to fitness and spread) N solutions”
Pg188, Subsection B, "Unlike in single objective optimization, there are two goals in a multiobjective optimization: 1) convergence to the Pareto-optimal set and 2) maintenance of diversity in solutions of the Preto-optimal set”
wherein Deb discloses a multi-objective optimization mechanism (NSGA-II) operating on a population of N candidates solutions, thereby generating a plurality of multipliers (representing different candidate weight vectors across generations) to simultaneously optimize trade-ff among multiple competing objective functions and output a Pareto-optimal set of solutions when combined with Bhaila’s feature-specific multipliers.)
	Bhaila and Deb are analogous art to the claimed invention because both belong to the same field of endeavor of machine learning classification and multi-objective parameter optimization.
​Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date to combine the iterative feature-and-value-specific reweighting multiplier framework of Bhaila with the fast, elitist multi-objective evolutionary optimization engine (NSGA-II) of Deb.
​The motivation to combine is explicitly recited by Deb (
Deb, Pg. 182, Abstract, Lines 9–14, "Specifically, a fast nondominated sorting approach with O(MN^2) computational complexity is presented. Also, a selection operator is presented that creates a mating pool by combining the parent and offspring populations and selecting the best (with respect to fitness and spread) N solutions"
) such that incorporating Deb's O(MN^2) multi-objective evolutionary engine into Bhaila's feature-specific reweighting mechanism enables the system to dynamically generate, evaluate, and select a plurality of candidate multipliers based on two validation scores (such as fitness score and fairness/spread score) received from non-dominated validation fronts without incurring prohibitive O(MN^3) computational complexity or requiring manual parameter tuning.

However, both Bhaila and Deb do not teach the rest of the limitation: including from a validation, receiving two validation scores that are based on the plurality of multipliers of the feature. 
However, in the same field of endeavor, Cruz teaches this (
Cruz, Pg6, Section3, P3, Lines4-7, “Fairlearn GS (Agarwal et al., 2018) is a similar method that instead uses a grid search over the constraint multipliers, and outputs a single (deterministic) classifier that achieves the best fairness-performance trade-off. RS Reweighing is a variation of GS that instead casts the choice of multipliers as another model hyperparameter”
Pg6, section3, p4, lines9-10, “The best performing model (the maximizer of 
    PNG
    media_image15.png
    44
    495
    media_image15.png
    Greyscale
) of each trial on validation data is then separately selected”
wherein Cruz discloses evaluating multiplier configurations on validation data to receive two validation scores (a performance score and a fairness score) that are calculated based on the multipliers. Here, evaluating candidate multipliers on validation data to receive performance and fairness scores is performed to assess multiplier performance without modifying, updating, or retraining the already trained base classifier (f_L) mentioned by Bhaila.)
Bhaila, Deb, and Cruz are analogous art to the claimed invention because they all belong to the same field of endeavor of machine learning classification, multi-objective parameter optimization, and algorithmic bias mitigation under trade-off constraints. ​Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date to combine the iterative feature-and-value-specific reweighting multiplier framework of Bhaila with the fast, elitist multi-objective evolutionary optimization engine (NSGA-II) of Deb, and further with evaluating multiplier configurations on validation data to compute performance and fairness scores as disclosed by Cruz. 
The motivation to combine is explicitly recited by Cruz (Cruz, Pg. 1, Abstract, Lines 3–6, "While fairness in these domains is a foremost concern, existing in-processing Fair ML methods are either incompatible with GBDT, or incur in significant performance losses while taking considerably longer to train.") Because Cruz explicitly recognizes and highlights the severe computational overhead, predictive performance loss, and prolonged training times inherent to in-processing model retraining, a person having ordinary skill in the art (PHOSITA) seeking to optimize algorithmic fairness while avoiding such retraining penalties would be directly motivated to apply Cruz’s validation-based score evaluation mechanism in a post-processing capacity. Specifically, a PHOSITA would be motivated to integrate Cruz’s validation-based score evaluation (receiving a performance score and a fairness score calculated based on the multipliers from validation data) into Deb’s multi-objective evolutionary optimization engine applied to Bhaila’s feature-specific reweighting framework, thereby enabling the system to dynamically generate, evaluate, and select a plurality of candidate multipliers based on the two validation scores in a post-processing manner, achieving optimal fairness-performance trade-offs without incurring prohibitive computational complexity, severe predictive performance degradation, or requiring slow retraining of the classifier.

As to dependent Claim 2, 
The combination of Bhaila, Deb and Cruz teaches all the limitations of Claim 1 as set forth above. 
Bhaila further teaches the method of Claim 1 wherein:
said probability of the class is a first probability of the class (
Bhaila, Pg1416-1417, Right Column, Subsection A. Collective Classification, Lines4-9, “The input is an unweighted, undirected, and attributed graph G = (V,E,X,Y) where V is a set of N nodes, E is a set of edges that connect node pairs in V, X represents the node feature matrix, and Y denotes node labels. Each node vi is defined by its feature vector xi which includes a sensitive attribute si”
Pg1417, Left Column, Lines6-8, “The task then is to simultaneously infer the values yi for vi or a probability distribution over those label values”
wherein as mentioned above in Claim1, the probability y_i is one of the label values of the input vi such that it encompasses first probability of the label or the class for the first input)

said multiplier that is specific to both of the feature and the value of the feature is a first value-specific multiplier (
Bhaila, Pg1419, Right Column, Paragraph1, Lines9-10, “wc,si is the weight due to node vi’s own sensitive attribute and the label to be estimated”
wherein as mentioned in Claim 1 that the multiplier or the weight wc,si is specific to the feature and the value of the feature as ‘s_i’ represents the feature value of the feature s, meaning if s_i is the first value of the feature, then it is equivalent to the claimed invention)

the method further comprises:
inferring, from a second input that contains a second value of the feature, a second probability of the class (
Bhaila, Pg1416-1417, Right Column, Subsection A. Collective Classification, Lines4-9, “The input is an unweighted, undirected, and attributed graph G = (V,E,X,Y) where V is a set of N nodes, E is a set of edges that connect node pairs in V, X represents the node feature matrix, and Y denotes node labels. Each node vi is defined by its feature vector xi which includes a sensitive attribute si”
Pg1417, Left Column, Lines6-8, “The task then is to simultaneously infer the values yi for vi or a probability distribution over those label values”
Pg1419, Left Column, Paragraph1, Lines14-16, “In our setup with binary s and binary y, we derive these weights for four combinations of s and y”
wherein as mentioned in Claim 1, the input vi is not a single input but it comprises multiple inputs such as v1, v2, … and so on. Thus, each input vi will be inferred with its own unique probability of the class, y_i or y1 and y2 respect to the input, which is equivalent to the claimed invention)

selecting based on the second value of the feature in the second input, from the plurality of multipliers of the feature, a second multiplier that is specific to both of the feature and the second value of the feature (
Bhaila, Pg1419, Right Column, Paragraph1, Lines4-5, “We also apply the weight based on the node’s own sensitive attribute and predicted node label”
Pg1419, Left Column, Paragraph1, Lines14-16, “In our setup with binary s and binary y, we derive these weights for four combinations of s and y”
Pg1419, Right Column, Paragraph1, Lines9-10, “wc,si is the weight due to node vi’s own sensitive attribute and the label to be estimated”
wherein as mentioned in Claim 1, each weight or multiplier is the sensitive feature-specific and there is the plurality of multipliers which the feature s comprises, thus it discloses there is a second multiplier or a multiplier that is specific to the second value of the feature for the second input)

classifying the second input based on the second multiplier that is specific to both of the feature and the second value of the feature and not based on the first value-specific multiplier (
Bhaila, Pg1419, Left Column, Paragraph1, Lines1-4, “We propose to incorporate this reweighting technique in the collective classification model during each iterative update to assign weights to nodes for each possible combination of s and y”, 
Pg1419, Algorithm2 Fair Collective Classification via Node Reweighting, Line6
 
    PNG
    media_image1.png
    24
    203
    media_image1.png
    Greyscale


Pg1419, Equation8, 
    PNG
    media_image2.png
    46
    350
    media_image2.png
    Greyscale

wherein as mentioned above, the final classification is according to the weight or multiplier that is specific to the sensitive feature s_i from the plurality of multipliers of the feature which means this s_i could be either the first value or the second value of the feature not both, which is equivalent to the claimed invention.)

As to dependent Claim 9, 
The combination of Bhaila, Deb and Cruz teaches all the limitations of Claim 1 as set forth above. 
However, Bhaila and Deb do not teach
two validation scores that are based on the plurality of multipliers of the feature are a fitness score and a fairness score

In the same field of endeavor, Cruz teaches this (
Cruz, Pg. 6, Section 3, Paragraph 3, Lines 4–7, "Fairlearn GS (Agarwal et al., 2018) is a similar method that instead uses a grid search over the constraint multipliers, and outputs a single (deterministic) classifier that achieves the best fairness-performance trade-off. RS Reweighing is a variation of GS that instead casts the choice of multipliers as another model hyperparameter" 
Pg. 6, Section 3, Paragraph 4, Lines 9–10, "The best performing model (the maximizer of [\alpha \cdot \text{performance} + (1 - \alpha) \cdot \text{fairness}]) of each trial on validation data is then separately selected"
wherein as mentioned in Claim 1, these validation scores are calculated based on the multipliers, and Cruz explicitly describes these scores as a performance score (fitness score) and a fairness score to evaluate trade-offs, which is functionally equivalent to the claimed invention.)
Bhaila, Deb, and Cruz are analogous art to the claimed invention because they all belong to the same field of endeavor of machine learning classification, multi-objective parameter optimization, and algorithmic bias mitigation under trade-off constraints. ​Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date to combine the iterative feature-and-value-specific reweighting multiplier framework of Bhaila with the fast, elitist multi-objective evolutionary optimization engine (NSGA-II) of Deb, and further with evaluating multiplier configurations on validation data to compute performance and fairness scores as disclosed by Cruz. 
The motivation to combine is explicitly recited by Cruz (Cruz, Pg. 1, Abstract, Lines 3–6, "While fairness in these domains is a foremost concern, existing in-processing Fair ML methods are either incompatible with GBDT, or incur in significant performance losses while taking considerably longer to train.") Because Cruz explicitly recognizes and highlights the severe computational overhead, predictive performance loss, and prolonged training times inherent to in-processing model retraining, a person having ordinary skill in the art (PHOSITA) seeking to optimize algorithmic fairness while avoiding such retraining penalties would be directly motivated to apply Cruz’s validation-based score evaluation mechanism in a post-processing capacity. Specifically, a PHOSITA would be motivated to integre Cruz’s validation-based score evaluation (receiving a performance score and a fairness score calculated based on the multipliers from validation data) into Deb’s multi-objective evolutionary optimization engine applied to Bhaila’s feature-specific reweighting framework enables the system to dynamically generate, evaluate, and select a plurality of candidate multipliers based on the two validation scores in a post-processing manner, thereby achieving optimal fairness-performance trade-offs without incurring prohibitive computational complexity, severe predictive performance degradation, or requiring slow retraining of the classifier.

As to dependent Claim 10, 
The combination of Bhaila, Deb and Cruz teaches all the limitations of Claim 1 as set forth above. 
Bhaila further teaches
said adjusting the multiplier is specific to both of the feature and the value of the feature (
Bhaila, Pg1419, Left Column, Algorithm2, Line4 and Euqation7

    PNG
    media_image16.png
    591
    613
    media_image16.png
    Greyscale


    PNG
    media_image17.png
    73
    671
    media_image17.png
    Greyscale

wherein as mentioned in Claim1, Bhaila discloses a post-hoc, post-processing framework that iteratively adjusts multipliers applied to model outputs based on a sensitive attribute s taking a specific value a (s=a) and class label c. Bhaila also derives and updates these feature-and-value-specific multipliers (w_{c,a}) across iterations t directly on the pre-trained predictions without modifying or retraining the parameters of the base classifier (f_L), which is functionally equivalent to adjusting the feature-and-value-specific multiplier without retraining as claimed.)

As to dependent Claim 11, 
The combination of Bhaila, Deb and Cruz teaches all the limitations of Claim 10 as set forth above. 
Bhaila further teaches
the method of Claim 10 wherein said adjusting and said reclassifying do not use the classifier (
Bhaila, Pg1417, Tabl1, Notation

    PNG
    media_image18.png
    35
    470
    media_image18.png
    Greyscale

Pg1417, Left Col, 1) Local Classifier, Lines1-2, “The local classifier f_L estimates label probabilities P(y_i | x_i) using only node attributes x_i”
Pg1419, Algorithm2

    PNG
    media_image19.png
    621
    644
    media_image19.png
    Greyscale

wherein the local classifier f_L is only used in Line 1 to initially infer class probabilities, and is no longer used during the weight adjusting step of Line 4 or the reclassifying step of Line 6, which is functionally equivalent to the claimed invention.)

As to dependent Claim 12, 
The combination of Bhaila, Deb and Cruz teaches all the limitations of Claim 1 as set forth above. 
Bhaila in view of Deb further teaches 
generating multiple pluralities of multipliers of the feature (
Bhaila, Pg1419, Left Col, P1, Lines1-4, “We propose to incorporate this reweighting technique in the collective classification model during each iterative update to assign weights to nodes for each possible combination of s and y”
Pg1419, Right Column, Paragraph1, Lines9-10, “wc,si is the weight due to node vi’s own sensitive attribute and the label to be estimated”
wherein as mentioned in Claim1, Bhaila discloses assigning feature-specific multipliers wc,si based on a sensitive feature s and its specific value s_i (a plurality of multipliers).
While Deb further discloses generating multiple populations of candidate solutions across generations during multi-objective optimization as mentioned in Claim 1 (
Deb, Pg182, Abstract, Lines9-14, “Specifically, a fast nondominated sorting approach with O(MN^2) computational complexity presented. Also, a selection operator is presented that creates a mating pool by combining the parent and offspring populations and selecting the best (with respect to fitness and spread) N solutions”
Pg188, Subsection B, "Unlike in single objective optimization, there are two goals in a multiobjective optimization: 1) convergence to the Pareto-optimal set and 2) maintenance of diversity in solutions of the Preto-optimal set”
) rendering it functionally equivalent to the claimed generating multiple pluralities of multipliers of the feature when combined.)
	However, Bhaila does not explicitly teach
detecting a subset of the multiple pluralities of multipliers of the feature that are on a bi-objective Pareto frontier
	In the same field of endeavor, Deb further teaches this (
Deb, Pg182, Abstract, Lines20-25, "Moreover, we modify the definition of dominance in order to solve constrained multiobjective problems efficiently. Simulation results of the constrained NSGA-II on a number of test problems, including a five-objective seven-constraint nonlinear problem, are compared with another constrained multiobjective optimizer and much better performance of NSGA-II is observed"
Pg182, Abstract, Lines11-14, "Also, a selection operator is presented that creates a mating pool by combining the parent and offspring populations and selecting the best (with respect to fitness and spread) N solutions"
Deb, Pg188, Subsection B, "Unlike in single objective optimization, there are two goals in a multiobjective optimization: 1) convergence to the Pareto-optimal set and 2) maintenance of diversity in solutions of the Preto-optimal set"
Deb, Pg186, Paragraph1, Lines6-14, "Now, solutions belonging to the best nondominated set F1 are of best solutions in the combined population and must be emphasized more than any other solution in the combined population. If the size of F1 is smaller than N, we definitely choose all members of the set F1 for the new population P_t+1. The remaining members of the population P_t+1 are chosen from subsequent nondominated fronts in the order of their ranking. Thus, solutions from the set F2 are chosen next, followed by solutions from the set F3, and so on"
wherein as the population is a set of candidate solutions, which a population of N Pareto-optimal solutions such that NGSA-II optimizer or the multi-objective optimizer will continuously selects the best subset of multipliers that fits the object thus where those subsets lie is the Pareto frontier which is functionally equivalent to the claimed invention.)

	Bhaila and Deb are analogous art to the claimed invention because both belong to the same field of endeavor of machine learning classification and multi-objective parameter optimization.
​Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date to combine the iterative feature-and-value-specific reweighting multiplier framework of Bhaila with the fast, elitist multi-objective evolutionary optimization engine (NSGA-II) of Deb.
​The motivation to combine is explicitly recited by Deb (
Deb, Pg. 182, Abstract, Lines 9–14, "Specifically, a fast nondominated sorting approach with O(MN^2) computational complexity is presented. Also, a selection operator is presented that creates a mating pool by combining the parent and offspring populations and selecting the best (with respect to fitness and spread) N solutions"
) such that incorporating Deb's O(MN^2) multi-objective evolutionary engine into Bhaila's feature-specific reweighting mechanism enables the system to dynamically generate, evaluate, and select a plurality of candidate multipliers based on two validation scores (such as fitness score and fairness/spread score) received from non-dominated validation fronts without incurring prohibitive O(MN^3) computational complexity or requiring manual parameter tuning.

As to independent Claim 17, 
it is a non-transitory computer-readable media claim that contains similar limitations of Claim 1 and thus rejected under the same rationale.

As to dependent Claim 18, 
it is a non-transitory computer-readable media claim that contains similar limitations of Claim 2 and thus rejected under the same rationale.

As to dependent Claim 20, 
it is a non-transitory computer-readable media claim that contains similar limitations of Claim 10 and thus rejected under the same rationale.

As to dependent Claim 21, 
it is a non-transitory computer-readable media claim that contains similar limitations of Claim 12 and thus rejected under the same rationale.


Claim 3 is rejected under 35 U.S.C. 103 as being unpatentable over Bhaila, Deb and Cruz, as mentioned in Claim 2, in further view of Pranto et al. (Pranto), Non-Patent Literature, “Performance Analysis of Ensemble Based Approaches to Mitigate Class Imbalance Problem After Applying Normalization”, published on July 2021, Pages: 5.

As to dependent Claim 3, 
The combination of Bhaila, Deb and Cruz teaches all the limitations of Claim 2 as set forth above. 
Bhaila further teaches the method of Claim 2 further comprising:
inferring, from a third input that contains the second value of the feature, a third probability of the class (
Bhaila, Pg1416-1417, Right Column, Subsection A. Collective Classification, Lines4-9, “The input is an unweighted, undirected, and attributed graph G = (V,E,X,Y) where V is a set of N nodes, E is a set of edges that connect node pairs in V, X represents the node feature matrix, and Y denotes node labels. Each node vi is defined by its feature vector xi which includes a sensitive attribute si”
Pg1417, Left Column, Lines6-8, “The task then is to simultaneously infer the values yi for vi or a probability distribution over those label values”
Pg1419, Left Column, Paragraph1, Lines14-16, “In our setup with binary s and binary y, we derive these weights for four combinations of s and y”
wherein as mentioned in Claim 2, the input vi is not a single input but it comprises multiple inputs such as v1, v2, … and so on. Thus, each input vi will be inferred with its own unique probability of the class, yi or y1 and y2 respect to the input, which is equivalent to the claimed invention)

inferring, from a fourth input that contains the second value of the feature, a fourth probability of the class (
Bhaila, Pg1416-1417, Right Column, Subsection A. Collective Classification, Lines4-9, “The input is an unweighted, undirected, and attributed graph G = (V,E,X,Y) where V is a set of N nodes, E is a set of edges that connect node pairs in V, X represents the node feature matrix, and Y denotes node labels. Each node vi is defined by its feature vector xi which includes a sensitive attribute si”
Pg1417, Left Column, Lines6-8, “The task then is to simultaneously infer the values yi for vi or a probability distribution over those label values”
Pg1419, Left Column, Paragraph1, Lines14-16, “In our setup with binary s and binary y, we derive these weights for four combinations of s and y”, 
wherein as mentioned above, the input vi is not a single input but it comprises multiple inputs such as v1, v2, … and so on. Thus, each input vi will be inferred with its own unique probability of the class, yi or y1 and y2 respect to the input, which is equivalent to the claimed invention)

However, Bhaila, Deb and Cruz do not explicitly teach
detecting, not based on the first probability of the class, a minimum probability for the second value of the feature and a maximum probability for the second value of the feature based on: the second probability of the class, the third probability of the class, and the fourth probability of the class

While Bhaila teaches about each input vi or v1, v2, …, vn has its own unique probability, yi, for the class and the three inputs (second, third and fourth) have the second value of the feature, in the same field of endeavor, Pranto teaches this limitation (
Pranto, Pg2, Section III, normalization technique 1, "1) min-max: For the minimum value min(xi) and the maximum value max(xi) of the data set, min-max normalization can be represented as Eqn. 1

    PNG
    media_image20.png
    88
    928
    media_image20.png
    Greyscale
"
wherein as mentioned above, there are 3 inputs where they share the feature or the second value of the feature and applying the min-max normalization technique of choosing the maximum and the minimum values, or the probabilities, which is functionally equivalent to the claimed invention.)
Bhaila, Deb, Cruz, and Pranto are analogous art to the claimed invention because they all belong to the same field of endeavor of machine learning classification, multi-objective parameter optimization, statistical normalization, and algorithmic bias mitigation under trade-off constraints. Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date to combine the iterative feature-and-value-specific reweighting multiplier framework of Bhaila, the fast elitist multi-objective evolutionary optimization engine (NSGA-II) of Deb, and the validation-based score evaluation of Cruz, further with the min-max normalization and bounding technique of Pranto. The motivation to combine is explicitly recited by Pranto (Pranto, Pg. 1, Right Column, Lines 7–9, "But these methods are hindered by several drawbacks such as sampling methods may lose information or cause over-fitting") such that incorporating Pranto's min-max bounding mechanism to detect the minimum and maximum probabilities for the second value of the feature into the combined post-processing multi-objective optimization framework of Bhaila, Deb, and Cruz establishes operational boundary margins to limit computational requirements and prevent over-fitting without losing critical feature distribution information.



Claim 4 is rejected under 35 U.S.C. 103 as being unpatentable over Bhaila, Deb, Cruz and Pranto as mentioned in Claim 3, in further view of Gidaris et al. (Gidaris), Non-Patent Literature, “Generating Classification Weights with GNN Denoising Autoencoders for Few-Shot Learning”, published on July 2019, Pages:10.

As to dependent Claim 4, 
The combination of Bhaila, Deb, Cruz and Pranto teaches all the limitations of Claim 3 as set forth above. 
Bhaila, Deb, Cruz and Pranto, however, do not teach
generating the second multiplier that is specific to both of the feature and the second value of the feature based on the minimum probability for the second value of the feature and the maximum probability for the second value of the feature

In the same field of endeavor, Gidaris teaches this (
Gidaris, Pg22, Subsection Our approach, Lines20-29, “In this context, in order to be able to recognize novel classes one must be able to generate classification weight vectors for them. So, the goal of our work is to learn a meta model that fulfills exactly this task: i.e., given a set of novel classes with few training examples for each of them, as well as the classification weights of the base classes, it learns to output a new set of classification weight vectors (both for the base and novel classes) that can then be used from the feature classifier in order to classify in a unified way both types of classes”
wherein the combination of the three will functionally match the limitation of generating a weight that is specific to the second value of the feature (corresponding second multiplier) between the range of the minimum probability and the maximum probability for efficient pruning procedure.)
	Bhaila, Deb, Cruz, Pranto, and Gidaris are analogous art to the claimed invention because they all belong to the same field of endeavor of machine learning classification, multi-objective parameter optimization, statistical normalization, and dynamic classification weight generation.
​Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date to combine the iterative feature-and-value-specific reweighting multiplier framework of Bhaila, the fast elitist multi-objective evolutionary optimization engine (NSGA-II) of Deb, the validation-based score evaluation of Cruz, and the min-max normalization bounding technique of Pranto, further with the dynamic weight generation mechanism of Gidaris.
​The motivation to combine is explicitly recited by Gidaris (Gidaris, Pg. 22, Left Column, Subsection DAE based model parameters generation, Lines 1–6, "Learning to perform such a meta-learning task, i.e., inferring the classification weights of a set of classes, is a difficult meta-problem that requires plenty of training data in order to be reliably solved. However, having access to such a large pool of data is not always possible; or otherwise stated the training data available for learning such meta-tasks might never be enough") such that incorporating Gidaris's parameter-generating mechanism into the bounded multi-objective optimization framework of Bhaila, Deb, Cruz, and Pranto enables the system to reliably generate a feature-and-value-specific second multiplier constrained within the min-max probability boundary without requiring large pools of training data or full retraining of the classifier.

Claims 5 and 6 are rejected under 35 U.S.C. 103 as being unpatentable over Bhaila, Deb, Cruz, Pranto and Gidaris as mentioned in Claim 4, in further view of Pleiss et al. (Pleiss), Non-Patent Literature listed in IDS filed on 02/09/2024, “On Fairness and Calibration”, published on Nov 2017, Pages: 19.

As to dependent Claim 5, 
The combination of Bhaila, Deb, Cruz, Pranto and Gidaris teaches all the limitations of Claim 4 as set forth above. 
However, Bhaila, Deb, Cruz, Pranto and Gidaris do not teach
said generating the second multiplier that is specific to both of the feature and the second value of the feature is based on a first ratio of the minimum probability for the second value of the feature over the maximum probability for the second value of the feature and a second ratio of the maximum probability for the second value of the feature over the minimum probability for the second value of the feature

In the same field of endeavor, Pleiss teaches this (
Pleiss, Pg6, Paragraph2, Lines6-8, "In Lemma 4 in Section S2, we show that if 
    PNG
    media_image21.png
    84
    945
    media_image21.png
    Greyscale
 then the cost of h~_2 is a linear interpolation between the costs of h_2 and h^u2 (Figure 2d). More formally, we have that g2(h~_2) = (1-alpha)g2(h2) + alpha g2 (h^u2)), and thus setting 
    PNG
    media_image22.png
    58
    276
    media_image22.png
    Greyscale
 ensures that g2(g2(h~_2) = g1(h1) as desired” 
Pg14, Lemma 4, 
    PNG
    media_image23.png
    90
    946
    media_image23.png
    Greyscale

wherein the two ratios are simply any continuous number between 0 and 1 which acts as an upper bound and a lower bound of the multiplier's selection range as mentioned in the claimed invention's specification paragraph [0036], "Multiplier 161 is a positive real number that , if less that one, generates a multiplicative product that is less than probability M1", thus these ratios must be a number that is also between 0 and 1. The alpha introduced by Pleiss is an interpolation parameter which operates as a functional mathematical equivalent to a multiplier's selection fractional range. Mathematically, a multiplier is defined as any scaling factor or coefficient that is multiplied by a variable to alter its magnitude or weight. The parameter alpha is explicitly utilized in a combinatorial formulation to determine the operational cost of the post-processed classifier, expressed in g2(h~_2) equation. In this algebraic structure, alpha acts directly as a multiplicate weight applied to the upper bound constraint, while (1-alpha) acts as the corresponding multiplier for the lower bound constraint. Therefore, generating a multiplier as mentioned in Claim 4 using the ratios is functionally equivalent to the claimed invention)
Bhaila, Deb, Cruz, Pranto, Gidaris, and Pleiss are analogous art to the claimed invention because they all belong to the same field of endeavor of machine learning classification, multi-objective parameter optimization, statistical normalization, dynamic classification weight generation, and error-cost balancing under calibration constraints. ​Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date to combine the iterative feature-and-value-specific reweighting multiplier framework of Bhaila, the fast elitist multi-objective evolutionary optimization engine (NSGA-II) of Deb, the validation-based score evaluation of Cruz, the min-max normalization bounding technique of Pranto, and the dynamic weight generation mechanism of Gidaris, further with the error-cost balancing using an interpolation parameter under calibration constraints of Pleiss.
​The motivation to combine is explicitly recited by Pleiss (Pleiss, Pg. 1, Abstract, Lines 4–5, "In this paper, we investigate the tension between minimizing error disparity across different population groups while maintaining calibrated probability estimates") such that incorporating Pleiss's interpolation-based balancing mechanism into the bounded multi-objective optimization framework of Bhaila, Deb, Cruz, Pranto, and Gidaris enables the system to generate the second value-specific multiplier based on a first ratio (min/max) and a second ratio (max/min) as operational boundary scaling factors, thereby mitigating group error disparities while maintaining calibrated probability estimates without requiring full classifier retraining. 

As to dependent Claim 6, 
The combination of Bhaila, Deb, Cruz, Pranto, Gidaris and Pleiss teaches all the limitations of Claim 5 as set forth above. 
However, Bhaila, Deb, Cruz, Pranto and Gidaris do not teach
the second multiplier that is specific to both of the feature and the second value of the feature is generated in a range from the first ratio to the second ratio

In the same field of endeavor, Pleiss teaches this (
Pleiss, Pg6, Paragraph2, Lines6-8, "In Lemma 4 in Section S2, we show that if 
    PNG
    media_image21.png
    84
    945
    media_image21.png
    Greyscale
 then the cost of h~_2 is a linear interpolation between the costs of h_2 and h^u2 (Figure 2d). More formally, we have that g2(h~_2) = (1-alpha)g2(h2) + alpha g2 (h^u2)), and thus setting 
    PNG
    media_image22.png
    58
    276
    media_image22.png
    Greyscale
 ensures that g2(g2(h~_2) = g1(h1) as desired”
Pg14, Lemma 4, 
    PNG
    media_image23.png
    90
    946
    media_image23.png
    Greyscale

wherein the two ratios are simply any continuous number between 0 and 1 which acts as an upper bound and a lower bound of the multiplier's selection range as mentioned in the claimed invention's specification paragraph [0036], "Multiplier 161 is a positive real number that , if less that one, generates a multiplicative product that is less than probability M1", thus these ratios must be a number that is also between 0 and 1. The alpha introduced by Pleiss is an interpolation parameter which operates as a functional mathematical equivalent to a multiplier's selection fractional range. Mathematically, a multiplier is defined as any scaling factor or coefficient that is multiplied by a variable to alter its magnitude or weight. The parameter alpha is explicitly utilized in a combinatorial formulation to determine the operational cost of the post-processed classifier, expressed in g2(h~_2) equation. In this algebraic structure, alpha acts directly as a multiplicate weight applied to the upper bound constraint, while (1-alpha) acts as the corresponding multiplier for the lower bound constraint. Therefore, generating a multiplier as mentioned in Claim 5 using the ratios is functionally equivalent to the claimed invention)
	Bhaila, Deb, Cruz, Pranto, Gidaris, and Pleiss are analogous art to the claimed invention because they all belong to the same field of endeavor of machine learning classification, multi-objective parameter optimization, statistical normalization, dynamic classification weight generation, and error-cost balancing under calibration constraints. ​Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date to combine the iterative feature-and-value-specific reweighting multiplier framework of Bhaila, the fast elitist multi-objective evolutionary optimization engine (NSGA-II) of Deb, the validation-based score evaluation of Cruz, the min-max normalization bounding technique of Pranto, and the dynamic weight generation mechanism of Gidaris, further with the linear interpolation parameter framework under calibration constraints of Pleiss.
The motivation to combine is explicitly recited by Pleiss (
Pleiss, Pg. 1, Abstract, Lines 4–5, "In this paper, we investigate the tension between minimizing error disparity across different population groups while maintaining calibrated probability estimates"
Pleiss, Pg6, Paragraph2, Lines6-8, "In Lemma 4 in Section S2, we show that if 
    PNG
    media_image21.png
    84
    945
    media_image21.png
    Greyscale
 then the cost of h~_2 is a linear interpolation between the costs of h_2 and h^u2 (Figure 2d). More formally, we have that g2(h~_2) = (1-alpha)g2(h2) + alpha g2 (h^u2)), and thus setting 
    PNG
    media_image22.png
    58
    276
    media_image22.png
    Greyscale
 ensures that g2(g2(h~_2) = g1(h1) as desired”
) such that incorporating Pleiss's interpolation parameter alpha into the bounded multi-objective optimization framework of Bhaila, Deb, Cruz, Pranto, and Gidaris enables the system to generate the second value-specific multiplier in a range from the first ratio to the second ratio as operational scaling bounds, thereby dynamically balancing group error disparities while maintaining calibrated probability estimates without requiring full retraining of the classifier. 

Claims 13, 14 and 22 are rejected under 35 U.S.C. 103 as being unpatentable over Bhaila, Deb and Cruz as mentioned in Claim 1 in further view of Grandini et al. (Grandini), Non-Patent Literature, “METRICS FOR MULTI-CLASS CLASSIFICATION: AN OVERVIEW”, published on Aug 2020, Pages: 17 in further view of Kriegman et al. (Kriegman), US-Patent, US-10,255,527-B2, Patented on Apr 9, 2019


As to dependent Claim 13, 
	The combination of Bhaila, Deb and Cruz teaches all the limitations of Claim 1 as set forth above. 
	Bhaila further teaches
rescaling, based on said multiplicative product of the probability of the class and the multiplier that is specific to both of the feature and the value of the feature (
Bhaila, Pg1419, Algorithm2

    PNG
    media_image19.png
    621
    644
    media_image19.png
    Greyscale

Pg1419, Equation8

    PNG
    media_image24.png
    46
    350
    media_image24.png
    Greyscale

wherein Bhaila teaches applying a feature-and-value-specific weight (w_{c,s_i}) as a multiplicative product to a target class probability, and applying a normalizing factor Z such that upon adjusting the target class probability via the multiplicative product, the remaining class probabilities (e.g., the second probability of the second class and the third probability of the third class) are proportionally rescaled by Z based on said multiplicative product in a post-processing manner without retraining the base classifier, which is functionally equivalent to the claimed rescaling)

	However, Bhaila, Deb and Cruz do not explicitly teach
inferring, from the input that contains the value of the feature, a second probability of a second class and a third probability of a third class

	In the same field of endeavor, Grandini teaches this (
Grandini, Pg1, Introduction, Paragraph6, Lines1-4, "A classification model gives us the probability of belonging to a specific class for each possible units. Starting from the probability assigned by the model, in the two-class classification problem a threshold is usually applied to decide which class has to be predicted for each unit. While in the multi-class case, there are various possibilities; among them, the highest probability value and the softmax are the most employed techniques"
wherein each unit or input in the multi-class model is assigned a unique probability for each class, such that if there are three classes, each class of the input is inferred with a probability, which is functionally equivalent to the claimed inferring a second probability of a second class and a third probability of a third class).
​Bhaila, Deb, Cruz, and Grandini are analogous art to the claimed invention because they all belong to the same field of endeavor of machine learning classification, multi-objective parameter optimization, and multi-class probabilistic model evaluation. Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date to combine the post-process iterative feature-and-value-specific reweighting multiplier framework of Bhaila with the fast, elitist multi-objective evolutionary optimization engine (NSGA-II) of Deb, and evaluating multiplier configurations on validation data to compute performance and fairness scores as disclosed by Cruz with the multi-class probability inference mechanism of Grandini. The motivation to combine is explicitly recited by Grandini (
Grandini, Pg. 2, Paragraph 1, Lines 1–3, "There are many metrics that come in handy to test the ability of any multi-class classifier and they turn out to be useful for: i) comparing the performance of two different models, ii) analysing the behaviour of the same model by tuning different parameters"
) such that extending Bhaila's binary reweighting framework to Grandini's multi-class environment enables the classification model to infer multiple class probabilities (such as a second probability for a second class and a third probability for a third class) for a given input, thereby allowing comprehensive performance evaluation and bias mitigation across multi-class predictions without retraining the base classifier.

However, Bhaila, Deb, Cruz and Grandini do not explicitly teach
said classifying the input is not based on said rescaling the second probability of the second class and the third probability of the third class

 	In the same field of endeavor, Kriegman teaches this (
Kriegman, Pg29, Column 32, Paragraph1, Lines1-13, “Notably, P(A|X) and P(I|X) have the same denominator, p(s). Moreover, the probabilities P(A|X), P(B|X), P(C|X) . . . P(N|X), P(I|X) that the Person X corresponds to person A, B, C, . . . N or an imposter sums to 1 (i.e., person X must either be an imposter or a tagged person). Accordingly, in one or more embodiments, digital classification system 100 ignores p(s) and simply scales calculated probabilities to 1 after calculating the numerator of the equations outlined above. For instance, one or more embodiments of digital classification system 100 utilizes the following equation to scale the probabilities: P(A|X)+P(B|X)+ . . . +P(N|X)+P(I|X)=1”
wherein Kriegman discloses that scaling or normalizing calculated class probabilities so that they sum to 1 is a post-processing operation that preserves original score proportions and does not alter or dictate the classification determination, such that classifying the input is not based on said rescaling of the remaining non-target class probabilities, which is functionally equivalent to the claimed limitation.)
	Bhaila, Deb, Cruz, Grandini, and Kriegman are analogous art to the claimed invention because they all belong to the same field of endeavor of machine learning statistical classification, predictive score processing, and probability distribution normalization. Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date to combine the post-process iterative feature-and-value-specific reweighting multiplier framework of Bhaila with the fast, elitist multi-objective evolutionary optimization engine (NSGA-II) of Deb, and evaluating multiplier configurations on validation data to compute performance and fairness scores as disclosed by Cruz, the multi-class probability inference mechanism of Grandini with the probability normalization and rescaling mechanism of Kriegman.
The motivation to combine is explicitly recited by Kriegman (
Kriegman, Pg. 14, Column 1, Lines 58–61, "conventional classification systems often have difficulty accurately classifying digital objects with regard to multiple possible classifications and multiple corresponding classification scores"
) such that incorporating Kriegman's post-processing score scaling mechanism into the multi-class framework of Bhaila, Deb, Cruz, and Grandini rescales the remaining non-target probabilities (the second and third probabilities) to ensure all class probabilities are mathematically calibrated, while ensuring that the final classification decision remains dictated by the primary target multiplicative product rather than being altered by the rescaling of the non-target probabilities.

As to dependent Claim 14, 
	The combination of Bhaila, Deb and Cruz teaches all the limitations of Claim 1 as set forth above. 
	Bhaila further teaches the method of Claim 1 wherein:
	the multiplier that is specific to both of the feature and the value of the feature (
Bhaila, Pg1419, Right Column, Paragraph1, Lines4-5, “We also apply the weight based on the node’s own sensitive attribute and predicted node label”
Pg1419, Left Column, Paragraph1, Lines14-16, “In our setup with binary s and binary y, we derive these weights for four combinations of s and y”
Pg1419, Right Column, Paragraph1, Lines9-10, “wc,si is the weight due to node vi’s own sensitive attribute and the label to be estimated”
wherein as mentioned in Claim 1, Bhaila teaches multiple multipliers to the sensitive feature s such that the multiplier or the weight is specific to the feature value s_i, so selecting and using the weight wc,si based on node v_i’s sensitive attribute value s_i is equivalent to the claimed invention)

said probability of the class is a probability of a first class (
Bhaila, Pg1416-1417, Right Column, Subsection A. Collective Classification, Lines4-9, “The input is an unweighted, undirected, and attributed graph G = (V,E,X,Y) where V is a set of N nodes, E is a set of edges that connect node pairs in V, X represents the node feature matrix, and Y denotes node labels. Each node vi is defined by its feature vector xi which includes a sensitive attribute si”
Pg1417, Left Column, Lines6-8, “The task then is to simultaneously infer the values yi for vi or a probability distribution over those label values”
Pg1419, Left Column, Paragraph1, Lines14-16, “In our setup with binary s and binary y, we derive these weights for four combinations of s and y”
wherein the target class probability inferred by the local classifier or relational classifier corresponds to a first probability of a first class)

	Bhaila further teaches the method further comprises from the input that contains the value of the feature (
Bhaila, Pg1419, Right Column, Paragraph1, Lines4-5, “We also apply the weight based on the node’s own sensitive attribute and predicted node label”
Pg1419, Left Column, Paragraph1, Lines14-16, “In our setup with binary s and binary y, we derive these weights for four combinations of s and y”
Pg1419, Right Column, Paragraph1, Lines9-10, “wc,si is the weight due to node vi’s own sensitive attribute and the label to be estimated”
wherein Bhaila teaches multiple multipliers to the sensitive feature s such that the multiplier or the weight is specific to the feature value s_i, so selecting and using the weight wc,si based on node v_i’s sensitive attribute value s_i is equivalent to the claimed invention.)
	
	However, Bhaila does not explicitly teach (highlighted part)
the multiplier that is specific to both of the feature and the value of the feature is less than one

In the same field of endeavor, Pleiss teaches this (
Pleiss, Pg6, Paragraph2, Lines6-8, "In Lemma 4 in Section S2, we show that if 
    PNG
    media_image25.png
    57
    613
    media_image25.png
    Greyscale
 then the cost of h~_2 is a linear interpolation between the costs of h_2 and h^u2 (Figure 2d). More formally, we have that g2(h~_2) = (1-alpha)g2(h2) + alpha g2 (h^u2)), and thus setting 
    PNG
    media_image26.png
    35
    167
    media_image26.png
    Greyscale
 ensures that g2(g2(h~_2) = g1(h1) as desired”
Pg14, Lemma 4, 
    PNG
    media_image27.png
    78
    813
    media_image27.png
    Greyscale

wherein the alpha here or the interpolation parameter that directly acts like a multiplier that is directly multiplied to the probabilities of the classes which is a number between 0 and 1. As a scaler factor, it tries to scale down the majority group’s values such that the minority can be selected which means it must be a value between 0 and 1 which is equivalent to the claimed invention.)
	Bhaila, Deb, Cruz, and Pleiss are analogous art to the claimed invention because they all belong to the same field of endeavor of machine learning classification, multi-objective parameter optimization, algorithmic bias mitigation, and error-cost balancing under calibration constraints. Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date to combine the post-processing multi-objective reweighting framework of Bhaila, Deb, and Cruz with the error-cost balancing using a linear interpolation parameter under calibration constraints disclosed by Pleiss.
​The motivation to combine is explicitly recited by Pleiss (Pleiss, Pg. 1, Abstract, Lines 4–5, "In this paper, we investigate the tension between minimizing error disparity across different population groups while maintaining calibrated probability estimates") such that incorporating Pleiss's scaling factor and interpolation parameter framework (alpha < 1) into the combined multi-objective optimization framework of Bhaila, Deb, and Cruz enables the system to generate and apply a feature-and-value-specific multiplier that is less than one, thereby mitigating group error disparities while maintaining calibrated probability estimates without requiring slow retraining of the base classifier.


However, Bhaila, Deb, Cruz and Pleiss do not teach
inferring a second probability of a second class that is less than the probability of the first class and a third probability of a third class
said classifying the input comprises classifying the input as the second class

	In the same field of endeavor, Grandini teaches limitation 1 (
Grandini, Pg1, Introduction, Paragraph6, Lines1-4, "A classification model gives us the probability of belonging to a specific class for each possible units. Starting from the probability assigned by the model, in the two-class classification problem a threshold is usually applied to decide which class has to be predicted for each unit. While in the multi-class case, there are various possibilities; among them, the highest probability value and the softmax are the most employed techniques”
Page4, Last paragraph, Lines 2-4, “This also means that Balanced Accuracy is insensitive to imbalanced class distribution and it gives more weight to the instances coming from minority classes. On the other hand, Accuracy treats all instances alike and usually favors the majority class”
Pg3, Section 2, Paragraph5, Lines-1-2, “When we think about classes instead of individuals, there will be classes with a high number of units and others with just few ones. In this situation, highly populated classes will have higher weight compared to the smallest ones”
 wherein Grandini discloses multi-class environment and its classes’ (corresponding second and third class) probabilities. Grandini also shows there will be imbalances between the classes such that one class might be bigger or smaller than the other classes in which is equivalent to the claimed invention’s the second class has the smallest probability compared to the first and the third class.)
	Gidaris further teaches said classifying the input comprises classifying the input as the second class (
Gidaris, Pg. 22, Left Column, Paragraph1, Lines 16–20, "a feature classifier, which (given as input the feature representation of an image) classifies it to one of the available classes by applying a set of classification weight vectors (one per class) to the input feature"
Pg. 23, Right Column, Section 2, Lines 8–12, "Given an image x, the feature extractor will output a d-dimensional feature z = F(x|theta) and then the classifier will compute the classification scores [s_1, ..., s_N] = [z^T w_1, ..., z^T w_N] := C(z|w) of the N classes"
wherein Gidaris discloses a feature classifier C(z\vert{}w) that computes classification scores [s_1, ..., s_N] across target classes and assigns the input to the class corresponding to the evaluated score. When combined with Bhaila and Pleiss, where a multiplier less than one (w < 1) is multiplied by the first class probability/score to scale down the primary class output below the second class probability, Gidaris's classification decision mechanism selects the resulting highest score, thereby classifying the input as the second class in a post-processing inference step without retraining the base classifier, which is functionally equivalent to the claimed invention.)

	Bhaila, Deb, Cruz, Pleiss, and Grandini are analogous art to the claimed invention because they all belong to the same field of endeavor of machine learning classification, multi-objective parameter optimization, probability calibration, and multi-class probabilistic model evaluation. ​Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date to combine the post-processing reweighting and calibrated optimization framework of Bhaila, Deb, Cruz, and Pleiss with the multi-class probability inference and class imbalance evaluation mechanism disclosed by Grandini.
​	The motivation to combine is explicitly recited by Grandini (
Grandini, Pg. 2, Paragraph 1, Lines 1–3, "There are many metrics that come in handy to test the ability of any multi-class classifier and they turn out to be useful for: i) comparing the performance of two different models, ii) analysing the behaviour of the same model by tuning different parameters"
) such that extending the post-processing reweighting and calibration framework to Grandini's multi-class environment enables the classification model to infer multiple class probabilities (including a second probability for a second class that is less than the first class probability, and a third probability for a third class) for a given input, thereby allowing comprehensive performance analysis and algorithmic bias mitigation across competing multi-class predictions without retraining the base classifier.

As to dependent Claim 22, 
it is a non-transitory computer-readable media claim that contains similar limitations of Claim 14 and thus rejected under the same rationale.


Claims 15 and 23 are rejected under 35 U.S.C. 103 as being unpatentable over Garg et al. (Garg), Non-Patent Literature, “A Unified View of Label Shift Estimation”, published on 2020, Pages: 11 in view of Non-Patent Literature of Deb et al. (Deb), Non-Patent Literature listed in IDS filed on 02/09/2024, “A Fast and Elitist Multiobjective Genetic Algorithm: NSGA-II Fairness and Calibration”, published on Apr 2002, Pages: 16.

As to independent Claim 15,
​Garg teaches a method comprising:
a first multiplier for a first class of a plurality of mutually exclusive classes and a second multiplier for a second class of the plurality of mutually exclusive classes (
Garg, Pg2, Section 2, Paragraph1, Lines 1-8, 

    PNG
    media_image28.png
    168
    764
    media_image28.png
    Greyscale

Pg2, Section 2, Lines 15-17, “This paper focuses on goal (ii), estimating importance weights w(y)=p_t(y)/p_s(y) for all y in Y”
wherein in multiclass classification over output space Y = {1, ..., k}, the target classes are mutually exclusive categorical labels, and Garg derives class-specific importance weights w_y for each class y, corresponding to a first multiplier for a first class and a second multiplier for a second class)

generating by a classifier that already was trained, from an input, an inference that contains a probability of the first class and a probability of the second class (
Garg, Pg2, Section 2, Paragraph1, Lines 1-8, 

    PNG
    media_image28.png
    168
    764
    media_image28.png
    Greyscale

Pg4, Section 4.3, Paragraph2, Lines5-6, “Absent knowledge of the ground truth p_s(y|x), we can plug in any approximate predictor f and optimize the following objective”
Pg3, Definition 1

    PNG
    media_image29.png
    44
    761
    media_image29.png
    Greyscale

Pg5, Lemma2

    PNG
    media_image30.png
    46
    763
    media_image30.png
    Greyscale

wherein input x is evaluated by an already trained black-box classifier f to output a probability vector f(x) 
    PNG
    media_image31.png
    15
    10
    media_image31.png
    Greyscale
 
    PNG
    media_image32.png
    22
    45
    media_image32.png
    Greyscale
. As set forth in Definition1 and Lemma2, for a calibrated predictor f, the y-th vector index component f_y(x) directly corresponds to the conditional probability P_s(y|f(x)) for target class y. Thus, evaluating input x via f infers probabilities across mutually exclusive target classes including a probability of a first class (f_1(x)) and a probability of a second class (f_2(x)). 
​	
classifying the input based on: a first multiplicative product of the probability of the first class and the first multiplier, and a second multiplicative product of the probability of the second class and the second multiplier (
Garg, Pg2, Section 2, Lines 17-20, “Given w, we can update our classifiers on the fly, either by retraining... or by applying an analytic correction”
Pg4, Section 4.3, Equation 3 and 4

    PNG
    media_image33.png
    50
    329
    media_image33.png
    Greyscale


    PNG
    media_image34.png
    35
    414
    media_image34.png
    Greyscale

Pg4, Section4.1, Line1-2, “Under label shift, for any (possibly randomized) mapping from X to Z…”
Pg3, Definition 1

    PNG
    media_image29.png
    44
    761
    media_image29.png
    Greyscale

Pg5, Lemma2

    PNG
    media_image30.png
    46
    763
    media_image30.png
    Greyscale

wherein Garg discloses updating classifier predictions on the fly via post-hoc analytic correction involving the objective formulation 
    PNG
    media_image35.png
    29
    63
    media_image35.png
    Greyscale
 in Eq4. Under Definition1 and Lemma2, for a calibrated predictor f(x), the inner product term 
    PNG
    media_image35.png
    29
    63
    media_image35.png
    Greyscale
 expands algebraically to 
    PNG
    media_image36.png
    42
    141
    media_image36.png
    Greyscale
, where p_s(y|z) is equivalent to f_y(x), which explicitly forms individual multiplicative products between each class probability f_y(x) and its corresponding class multiplier w_y (specifically, a first multiplicative product f_1(x)*w_1 for the first class and a second class multiplicative product f_2(x)*w_2 for the second class), which entails classifying or assigning input x into a target class based on these resulting multiplicative products across the mutually exclusive classes)
​
wherein the method is performed by one or more computers without retraining the classifier (
Garg, Pg2, Section 2, Lines 17-20, “Given w, we can update our classifiers on the fly, either by retraining in an importance-weighted ERM framework... or by applying an analytic correction [1, 18].”
Pg3, Section 2, Lines 1-3, “calibrating a trained classifier on held-out data, post-hoc.”
Pg8, Section 6, Lines 8-12, “We experimentally illustrate the performance of MLLS on synthetic data, MNIST, and CIFAR10...”
wherein Garg explicitly teaches applying post-hoc analytic corrections directly to predictions without retraining the black-box classifier f, and implementing the algorithm on computer systems evaluating datasets such as MNIST and CIFAR-10)

​However, Garg does not explicitly teach (Highlighted Part)
Performing by a bi-objective optimizer: generating a first multiplier for a first class of a plurality of mutually exclusive classes and a second multiplier for a second class of the plurality of mutually exclusive classes
​
In the same field of endeavor, Deb teaches this (
Deb, Pg182, Abstract, Lines 9-14, “Specifically, a fast nondominated sorting approach with O(MN^2) computational complexity is presented. Also, a selection operator is presented that creates a mating pool by combining the parent and offspring populations and selecting the best (with respect to fitness and spread) N solutions”
Deb, Pg188, Subsection B, Lines 1-3, “Unlike in single objective optimization, there are two goals in a multiobjective optimization: 1) convergence to the Pareto-optimal set and 2) maintenance of diversity in solutions of the Pareto-optimal set”
wherein Deb discloses a bi-objective optimization mechanism (NSGA-II) operating on candidate parameter vectors, thereby generating candidate solution vectors comprising multipliers (w_1, w_2) for mutually exclusive classes when combined with Garg’s class-specific weight framework)
​	Garg and Deb are analogous art to the claimed invention because both belong to the same field of endeavor of machine learning classification, post-processing parameter adaptation, and multi-objective optimization. Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date to combine the post-hoc analytic correction and class-specific importance weight framework of Garg with the bi-objective evolutionary optimization engine (NSGA-II) of Deb. The motivation to combine is explicitly recited by Deb (
Deb, Pg. 182, Abstract, Lines 9–14, "Specifically, a fast nondominated sorting approach with O(MN^2) computational complexity is presented. Also, a selection operator is presented that creates a mating pool by combining the parent and offspring populations and selecting the best (with respect to fitness and spread) N solutions"
) such that incorporating Deb's bi-objective optimization engine into Garg's post-processing weight estimation framework enables the system to dynamically generate, evaluate, and select a plurality of candidate class multipliers (first and second multipliers) to solve trade-offs among competing objectives (such as convergence/accuracy and diversity/fairness) without requiring prohibitive computational complexity or slow retraining of the base classifier.

As to dependent Claim 23, 
it is a non-transitory computer-readable media claim that contains similar limitations of Claim 15 and thus rejected under the same rationale.


Claims 16 is rejected under 35 U.S.C. 103 as being unpatentable over Garg and Deb as mentioned in Claim 15 in further view of over Fernando et al. (Fernando), Non-Patent Literature, “Dynamically Weighted Balanced Loss: Class Imbalanced Learning and Confidence Calibration of Deep Neural Networks”, published on July 2022, Pages: 12

As to dependent Claim 16,
The combination of Garg and Deb teaches all the limitations of Claim 15 as set forth above. 
Garg further teaches
​said classifying the input uses a plurality of multipliers that contains the first multiplier and the second multiplier (
Garg, Pg2, Section 2, Paragraph3, Lines1-4

    PNG
    media_image37.png
    82
    756
    media_image37.png
    Greyscale

Garg, Pg2, Section 2, Lines 17-20, “Given w, we can update our classifiers on the fly, either by retraining... or by applying an analytic correction”
Pg4, Section 4.3, Equation 3 and 4

    PNG
    media_image33.png
    50
    329
    media_image33.png
    Greyscale


    PNG
    media_image34.png
    35
    414
    media_image34.png
    Greyscale

Pg4, Section4.1, Line1-2, “Under label shift, for any (possibly randomized) mapping from X to Z…”
Pg3, Definition 1

    PNG
    media_image29.png
    44
    761
    media_image29.png
    Greyscale

Pg5, Lemma2

    PNG
    media_image30.png
    46
    763
    media_image30.png
    Greyscale

wherein as mentioned in Claim 15, the combination of Garg and Deb generated plurality of multipliers in which Garg applies a vector/plurality of class importance weights w_y (multipliers as w_1 and w_2) containing a first weight for a first class and a second weight for a second class during input classification)

the inference contains a plurality of probabilities that contains the probability of the first class and the probability of the second class (
Garg, Pg2, Section 2, Paragraph1, Lines 1-8, 

    PNG
    media_image28.png
    168
    764
    media_image28.png
    Greyscale

wherein the trained classifier outputs an inference vector 
    PNG
    media_image38.png
    31
    21
    media_image38.png
    Greyscale
 (x) containing predicted probabilities across multiple target classes, including a first probability for a first class and a second probability for a second class)

	​a count of the plurality of probabilities equals a count of the plurality of mutually exclusive classes (
Garg, Pg2, Section 2, Paragraph1, Lines 1-8, 

    PNG
    media_image28.png
    168
    764
    media_image28.png
    Greyscale

wherein the output space Y consists of k mutually exclusive target classes, and the classifier infers a probability vector 
    PNG
    media_image38.png
    31
    21
    media_image38.png
    Greyscale
 (x) 
    PNG
    media_image31.png
    15
    10
    media_image31.png
    Greyscale
 
    PNG
    media_image32.png
    22
    45
    media_image32.png
    Greyscale
 containing k probability entries, such that the count of probabilities equals the count k of mutually exclusive classes)

	However, both Garg and Deb do not teach
a count of the plurality of multipliers is less than the count of the plurality of probabilities

	In the same field of endeavor, Fernando teaches this (
Fernando, Pg2943, Right Column, Equation 8, paragraph4, Lines1-3, 
    PNG
    media_image39.png
    59
    392
    media_image39.png
    Greyscale
 “As such, misclassification errors for a class j with classwise cost wj will have wj-times more penalty than misclassification errors for the majority class with weight equals 1”
Pg2944, Left Column, Paragraph2, Lines3-5, “the weighting factor for well-classified instance is close to 1, and hence, the loss is unaffected and equivalent to CE”
wherein the equation shows that the majority class will be assigned a weight in such that it will always be 1 that is unaffected. This is functionally equivalent to the claimed invention of a class not being assigned with a weight or assigned a weight that is 1 which is identical as multiplying a number by 1 yields the same number, meaning there is actually one less multiplier compared to the number of the mutually exclusive classes.)

	Garg, Deb and Fernando are analogous to the claimed invention as they are from the same field of endeavor of machine learning classification, post-processing parameter adaptation, multi-objective optimization, and class-imbalanced learning. Therefore, it would have been obvious to a person having ordinary skill in the art before the effective filing date to combine the post-hoc analytic correction and class-specific importance weight framework of Garg with the bi-objective evolutionary optimization engine (NSGA-II) of Deb, and further with assigning the majority class a neutral baseline weight of 1 while reweighting minority classes as disclosed by Fernando. The motivation to combine is explicitly recited by Fernando (
Fernando, Pg2943, Right Column, Equation 8 and following 3 lines, 

    PNG
    media_image40.png
    54
    214
    media_image40.png
    Greyscale

"As such, misclassification errors for a class j with classwise cost w_j will have w_j-times more penalty than misclassification errors for the majority class with weight equals 1"
) such that it would have been obvious to a person having ordinary skill in the art (PHOSITA) at the time of the invention to incorporate Fernando’s teaching of fixing the majority class multiplier to 1 into Deb’s multi-objective optimization engine as implemented within Garg’s post-processing framework. A PHOSITA would be motivated to fix the majority class multiplier to an identity value in order to reduce the dimensional search space of the optimization algorithm, thereby significantly improving computational efficiency.






Response to Arguments

35 USC § 101
In review of Applicant’s amendments, Applicant’s arguments and amendments, filed July 20, 2026, regarding the abstract idea rejections from the previous office action made under 35 U.S.C. 101 have been fully considered and they are persuasive.  Therefore, the rejections under 35 U.S.C. 101 have been withdrawn.

35 USC § 103
Applicant’s arguments and amendments, filed on July 20, 2026., regarding the rejections from the previous office action made under 35 U.S.C. 103 have been fully considered and see below for the details.


In the Remarks, Applicant argues in substance that
Claim 1
Applicant’s arguments with respect to claim 1 have been considered but are moot in view of the new ground of rejection under 35 U.S.C. 103 over Bhaila in view of Deb and Cruz, necessitated by Applicant’s amendment.
(A) – Applicant argues that Cruz is directed to an in-processing training method and is therefore incompatible with Claim 1, which requires performing operations “without retraining the classifier”. This argument has been carefully considered but is unpersuasive.
Applicant argues that there is no teaching, suggestion or motivation to combine the references (see 2nd paragraph of page 10 of Remarks). In response to applicant’s argument that there is no teaching, suggestion, or motivation to combine the references, the examiner recognizes that obviousness may be established by combining or modifying the teachings of the prior art to produce the claimed invention where there is some teaching, suggestion, or motivation to do so found either in the references themselves or in the knowledge generally available to one of ordinary skill in the art.  See In re Fine, 837 F.2d 1071, 5 USPQ2d 1596 (Fed. Cir. 1988), In re Jones, 958 F.2d 347, 21 USPQ2d 1941 (Fed. Cir. 1992), and KSR International Co. v. Teleflex, Inc., 550 U.S. 398, 82 USPQ2d 1385 (2007).  In this case, 
the Office Action under the new ground of rejection does not rely on Cruz’s underlying decision-tree training mechanism. Rather, Cruz is relied upon specifically for its teachings of evaluating candidate multiplier configurations on validation data to obtain two validation scores (a performance score and a fairness score). As cited in the Office Action, Cruz explicitly discloses evaluating multiplier choices on validation data to assess fairness-performance trade-offs. Evaluating candidate multipliers on validation data using pre-calculated model outputs is a post-hoc evaluation step that operates without modifying, updating, or retraining the parameters of the base classifier (f_L). Therefore, applying Cruz’s validation-based score evaluation to candidate multipliers generated by Deb’s multi-objective optimizer within Bhaila’s post-processing reweighting framework directly yields the claimed limitation without retraining the base classifier.
(B) – Applicant argues that it is unclear whether “a classifier” in Claim 1 maps to the local classifier (f_L), relational classifier (f_R), or the overall collective classification model in Bhaila. This argument has been carefully considered but is unpersuasive.
The Office Action under the new ground of rejection explicitly mapped “a classifier that already was trained” to Bhaila’s local classifier (f_L). Because the rejection explicitly identifies f_L as the baseline trained classifier that remains untrained during multiplier adjustment, there is no more ambiguity regarding it. Please refer to the rejection above for further details.

Claims 13-14
Applicant’s arguments with respect to claims 13-14 have been considered but are moot in view of the new ground of rejection under 35 U.S.C. 103 over Bhaila in view of Deb, Cruz, Grandini and Kriegman, necessitated by Applicant’s amendment to Claim 1.
Applicant argues in substance that Claim13 and 14 are non-obvious because Cruz relies on training loss optimization (e.g., Equations 7-8) and in-processing gradient boosting, which Applicant asserts is incompatible with the non-training limitation of Claim 1 (“without retraining the classifier”), which have been fully considered but Applicant’s arguments are unpersuasive.
Applicant argues that there is no teaching, suggestion or motivation to combine the references (see Last paragraph of page 11 of Remarks). In response to applicant’s argument that there is no teaching, suggestion, or motivation to combine the references, the examiner recognizes that obviousness may be established by combining or modifying the teachings of the prior art to produce the claimed invention where there is some teaching, suggestion, or motivation to do so found either in the references themselves or in the knowledge generally available to one of ordinary skill in the art.  See In re Fine, 837 F.2d 1071, 5 USPQ2d 1596 (Fed. Cir. 1988), In re Jones, 958 F.2d 347, 21 USPQ2d 1941 (Fed. Cir. 1992), and KSR International Co. v. Teleflex, Inc., 550 U.S. 398, 82 USPQ2d 1385 (2007).  In this case, the Office Action’s Cruz is not cited for in-processing model training or loss optimization. Instead, Cruz is relied upon in combination with Bhaila and Deb to disclose evaluating multiplier configurations on validation data to receive two validation scores (a performance score and a fairness score) in a post-processing capacity without modifying or retraining the trained base classifier. Furthermore, Cruz explicitly highlights the severe computational overhead and prolonged training times inherent to in-processing model retraining. This teaching provides explicit motivation for a PHOSITA seeking to avoid such retraining penalties to apply Cruz’s validation-based score evaluation mechanism in a post-processing framework as taught by Bhaila and Deb.
Regarding Claim 13, the combination of Bhaila, Deb and Cruz in view of Grandini and Kriegman demonstrates that inferring multi-class probabilities, applying post-processing multiplicative reweighting, and normalizing/rescaling non-target probabilities are performed entirely in a post-processing manner without retraining the base classifier. Please refer to the rejection above for further details.
Regarding Claim 14, the combination of Bhaila, Deb and Cruz in view of Grandini and Kriegman demonstrates that applying scaling multiplier less than one to adjust primary class scores and reclassifying the input as the second class is executed during post-processing inference without retraining the classifier. Please refer to the rejection above for further details.

Claim 15
Applicant argues that Claim 15 is nonobvious over Fernando because Fernando relies on training loss and training activities, which are incompatible with the claim limitation “without retraining the classifier”, which have been considered but are moot in view of the new ground of rejection under 35 U.S.C. 103 over Garg in view of Deb, necessitated by Applicant’s amendment, thus Applicant’s arguments are unpersuasive.
Garg explicitly teaches generating an inference using a classifier that was already trained and classifying the input based on multiplicative products via post-hoc analytic correction on the fly without retraining the black-box classifier f. Combined with Deb’s bi-objective optimization mechanism, the combination teaches all limitations of Claim 15. Please refer to the rejection above for further details.

Claim 16
Applicant argues that Claim 16 is nonobvious over Fernando because Fernando relies on training loss and training activities, which are incompatible with the claim limitation “without retraining the classifier”, which have been considered but are moot in view of the new ground of rejection under 35 U.S.C. 103 over Garg in view of Deb and Fernando, necessitated by Applicant’s amendment in Claim 15.
Firstly, as the rejection of the independent Claim 15 has been modified to rely on Garg in view of Deb, which Garg explicitly teaches a post-hoc, black-box analytic correction mechanism performed directly on inference outputs without retraining the base classifier. Thus, the foundational framework of the applied art fully satisfies the “without retraining the classifier” requirement.
Secondly, Claim 16’s Fernando is applied in further view of Garg and Deb strictly for its teachings of assigning a neural baseline weight of 1 to the majority class, which results in “a count of the plurality of multipliers is less than the count of the plurality of probabilities”. A PHOSITA implementing Garg’s post-processing inference framework would be directly motivated by Fernando to fix the majority class multiplier to an identify value (1) in order to reduce optimization search space and computational overhead. Applying a multiplier of 1 to an inferred class probability during post-processing classification operates purely on the predicted probability values and requires zero parameter modification or retraining of the base classifier. Therefore, Fernando’s disclosure of majority-class weight assignment is structurally and operationally compatible with the non-retraining framework of Garg and Deb, and Claim 16 would have been obvious under 35 U.S.C. 103. Please refer to the rejection above for further details.

Claim 17, 18, 20-23
These Claims have similar limitations of Claims 1, 2, 10, 12, 14 and 15, respectively, thus rejected under the same rationale. Likewise, please refer to the responses of Claims 1, 2, 10, 12, 14 and 15, respectively.
Conclusion
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to DONG YOON JUNG whose telephone number is (571)270-0198. The examiner can normally be reached 8am-5pm.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Cesar Paula can be reached at (571) 272-4128. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.





/DONG YOON JUNG/Examiner, Art Unit 2145         


/CESAR B PAULA/Supervisory Patent Examiner, Art Unit 2145
Read full office action
MULTIPLIER TUNING POSTPROCESSING FOR MACHINE LEARNING BIAS MITIGATION

Examiner Intelligence

Office Action

Prosecution Timeline

Strategy Recommendation AI-generated — please review before filing

Prosecution Projections

Ready to respond to this office action?

MULTIPLIER TUNING POSTPROCESSING FOR MACHINE LEARNING BIAS MITIGATION

Examiner Intelligence

Office Action

Prosecution Timeline

Strategy Recommendation AI-generated — please review before filing

Prosecution Projections

Ready to respond to this office action?

Sign in with your work email
