Prosecution Insights
Last updated: November 04, 2026
Application No. 18/928,581

TARGET PREDICTION METHOD USING PRE-TRAINING AND TRANSFER LEARNING, AND TARGET PREDICTION FRAMEWORK FOR PERFORMING SAME

Non-Final OA §101§103§112
Filed
Oct 28, 2024
Priority
Dec 01, 2023 — RE 10-2023-0172099
Examiner
SITTNER, MATTHEW T
Art Unit
3629
Tech Center
3600 — Transportation & Electronic Commerce
Assignee
Impactive AI
OA Round
1 (Non-Final)
58%
Grant Probability
Moderate
1-2
OA Rounds
1y 0m
Est. Remaining
99%
With Interview

Examiner Intelligence

Grants 58% of resolved cases
58%
Career Allowance Rate
528 granted / 911 resolved
+6.0% vs TC avg
Strong +56% interview lift
Without
With
+56.0%
Interview Lift
resolved cases with interview
Typical timeline
3y 0m
Avg Prosecution
35 currently pending
Career history
953
Total Applications
across all art units

Statute-Specific Performance

§101
35.8%
-4.2% vs TC avg
§103
42.0%
+2.0% vs TC avg
§102
9.5%
-30.5% vs TC avg
§112
12.2%
-27.8% vs TC avg
Black line = Tech Center average estimate • Based on career data from 911 resolved cases

Office Action

§101 §103 §112
DETAILED ACTION 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 . Status of Claims This action is responsive to the application filed October 28, 2024. Claims 1-12 are pending. Claim 1 is independent in form. Claims 7-12 are addressed below under Claim Interpretation and 35 U.S.C. 112(d). PRELIMINARY MATTERS Specification Citation Basis Citations in this action to applicant's specification are to the paragraph numbers of the published application, US 2025/0181929 A1 (published June 5, 2025). The specification as filed on October 28, 2024 does not number its paragraphs; the paragraph numbers were assigned at publication. Foreign Priority Acknowledgment is made of applicant's claim for foreign priority under 35 U.S.C. 119(a)-(d) to Korean Patent Application No. 10-2023-0172099, filed December 1, 2023. Receipt is acknowledged of certified copies of papers required by 37 CFR 1.55. For purposes of this action, December 1, 2023 is treated as the effective filing date of the claimed invention. Information Disclosure Statements The information disclosure statements submitted on October 28, 2024 and July 10, 2026 were filed before the mailing of a first Office action on the merits and comply with 37 CFR 1.97 and 1.98. Accordingly, the information disclosure statements have been considered by the examiner (MPEP 609). Korean Patent No. 10-2050855, Korean Publication No. 10-2023-0081028, and Chinese Publication No. 116362799 were considered in full to the extent of the English-language abstracts and translations of record (MPEP 609.04(a)(III)). Drawings, Specification, Abstract, and Claim Form The drawings filed October 28, 2024 (three sheets, FIGS. 1-4) are acceptable. The title is descriptive (MPEP 606.01) and the abstract complies with 37 CFR 1.72(b). The claims carry no status identifiers, which is proper for the claims as originally filed, and no claim is an improper multiple dependent claim. No objection is made to the specification, the abstract, or the drawings. CLAIM INTERPRETATION During examination, claims are given their broadest reasonable interpretation consistent with the specification (MPEP 2111). 35 U.S.C. 112(f) — "framework" (Claims 7-12) The following is a quotation of 35 U.S.C. 112(f): (f) Element in Claim for a Combination. – An element in a claim for a combination may be expressed as a means or step for performing a specified function without the recital of structure, material, or acts in support thereof, and such claim shall be construed to cover the corresponding structure, material, or acts described in the specification and equivalents thereof. Claims 7-12 each recite "A target prediction framework for performing a target prediction method of claim [1-6]." The claim limitation is analyzed under the three-prong test of MPEP 2181(I). (A) The term "framework" is a generic placeholder: the specification states that the term includes an abstract structure of reusable software (¶ [0069]), and the term does not have a sufficiently definite meaning as the name for structure. (B) The term is modified by functional language, namely "for performing a target prediction method of claim [N]." (C) The term is not modified by sufficient structure, material, or acts for performing that function; the claims recite no processor, memory, or other structure. The limitation therefore invokes 35 U.S.C. 112(f). The corresponding structure is the computing device including at least one processor (for example, an AI server with a CPU and a GPU) and a storage medium storing the steps of the method (¶¶ [0067]-[0068]), programmed to perform the algorithm of steps S100-S400 described at ¶¶ [0027]-[0050] and shown in FIG. 1, and equivalents thereof (MPEP 2181(II)(B)). Because the specification discloses that structure and algorithm, no rejection under 35 U.S.C. 112(b) is made on this basis. Because this claim limitation is being interpreted under 35 U.S.C. 112(f), it is being interpreted to cover the corresponding structure described in the specification as performing the claimed function, and equivalents thereof. If applicant does not intend to have this limitation interpreted under 35 U.S.C. 112(f), applicant may (1) amend the claim limitation to avoid it being interpreted under 35 U.S.C. 112(f) (e.g., by reciting sufficient structure to perform the claimed function); or (2) present a sufficient showing that the claim limitation recites sufficient structure to perform the claimed function so as to avoid it being interpreted under 35 U.S.C. 112(f). Limitations That Do Not Invoke 35 U.S.C. 112(f) Claim 1 recites "a data input step," "a base model training step," "a cluster classification step," and "a transfer learning step." Each "step" is followed by the act that performs it ("of inputting," "of training," "of classifying," "of inputting ... and retraining"), so the steps recite acts and are not step-plus-function limitations (MPEP 2181(I)(A)). The "computing device comprising at least one processor" of claim 1 recites structure. Preamble and Descriptive Language In claim 1, the recitation "target prediction" states the purpose of the method; the body of the claim sets out the complete method, so the recitation does not further limit the steps (MPEP 2111.02(II)). The recitation that the method is "performed by a computing device comprising at least one processor" is given weight because it identifies what performs each step. The recitation in claim 6 that "the targets are demand" characterizes the subject of the predictions; it is given weight to the extent that it limits the content of the prediction datasets, and it is in any case taught by the applied art. Terms Described in the Specification The specification expressly defines "framework" (¶ [0069]); that definition is applied through the 35 U.S.C. 112(f) construction above. The specification describes a SHAP value as a value describing how each feature affects a predicted value in a machine learning model (¶ [0039]); the term "SHAP values" is given its ordinary meaning in the art consistent with that description, namely Shapley-value feature attributions for a model's prediction. The claims do not define "transfer learning"; consistent with ¶¶ [0046]-[0049], it is interpreted to include using the weights of an already-trained model as the starting point and retraining some or all of those weights on new data (fine-tuning). In claim 1, "the respective deep learning models" of the transfer learning step is read as referring back to the "deep learning models" trained in the base model training step, so that the plurality of clusters are input into the already-trained models and their weights are retrained. This reading follows the antecedent basis of the claim and ¶ [0047] of the published application. Claim 1 does not, however, require that any one of the deep learning models be trained on the data of all of the items, and it does not specify which weights are retrained or how. Claim 1 also does not specify the unit that is classified into clusters in the cluster classification step — an item's dataset as a whole, or the individual data instances of the datasets. Claim 3 refers to SHAP values "for each data of the prediction datasets," and the specification's averaging of SHAP values per item before clustering items (¶ [0042]) is not recited in claim 1. Classifying the data of the prediction datasets into clusters, whether by item or by instance, is therefore within the broadest reasonable interpretation of the step; limitations from the specification are not read into the claim (MPEP 2111.01(II)). 35 U.S.C. 101 — SUBJECT MATTER ELIGIBILITY 35 U.S.C. 101 reads as follows: "Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title." Claims 1-12 are rejected under 35 U.S.C. 101 because the claimed invention is directed to a judicial exception (an abstract idea) without significantly more. The claims have been analyzed under the two-part framework of Alice Corp. Pty. Ltd. v. CLS Bank Int'l, 573 U.S. 208, 217-18 (2014), as set out in MPEP 2106: Step 1 (statutory category), Step 2A Prong One (whether the claim recites a judicial exception), Step 2A Prong Two (whether the claim as a whole integrates the exception into a practical application), and Step 2B (whether the additional elements amount to significantly more than the exception). Claim 1, Step 1 Claim 1 recites a method comprising a series of steps and is directed to a process, a statutory category (MPEP 2106.03). Step 2A, Prong One Claim 1 recites a mathematical concept (MPEP 2106.04(a)(2)(I)). The limitation "a cluster classification step of classifying the prediction datasets into a plurality of clusters by using SHapley Additive exPlanations (SHAP) values" recites calculating SHAP values and grouping the datasets according to those values. A SHAP value is a Shapley value: a quantity defined by a formula from cooperative game theory that apportions a model's predicted value among its input features (see the published application, ¶ [0039], describing a SHAP value as a value describing how each feature affects a predicted value). Computing such values and partitioning data according to the computed numbers are mathematical calculations and mathematical relationships. The specification confirms the mathematical character of the step: the SHAP values are taken in absolute value, averaged into a vector for each item, and partitioned by K-means clustering on Euclidean distances to cluster centers (¶¶ [0042]-[0043]). Grouping data by numerical correlations of this kind is an abstract idea. Digitech Image Techs., LLC v. Elecs. for Imaging, Inc., 758 F.3d 1344, 1350-51 (Fed. Cir. 2014) (organizing information through mathematical correlations is abstract). The base model training step ("a base model training step of training deep learning models by using the prediction datasets input in the data input step") and the transfer learning step ("a transfer learning step of inputting the plurality of clusters into the respective deep learning models and retraining respective weights through transfer learning after the cluster classification step") are recited only as the computation of numerical weights of a model from data. To the extent these steps are understood to recite the calculation of those weights, they are part of the recited mathematical concept as well. Compare Example 47, claim 2, of the July 2024 Subject Matter Eligibility Update, in which a step of training an artificial neural network with backpropagation and gradient descent algorithms was identified as a mathematical concept. Claim 1 does not name the algorithms used, so these steps are treated below as additional elements; the result is the same either way. Accordingly, claim 1 recites a judicial exception. Step 2A, Prong Two The additional elements recited in claim 1 are: (a) "a computing device comprising at least one processor"; (b) "a data input step of inputting prediction datasets related to targets for prediction"; (c) the base model training step; and (d) the transfer learning step. Considered individually and in combination, these elements do not integrate the judicial exception into a practical application. Element (a) is a generic computer recited at a high level of generality as the tool that performs the steps. The specification describes it only as a computing device with at least one processor, such as a server with a CPU or GPU (¶¶ [0067]-[0068]). Using a generic computer to perform an abstract idea is mere instructions to apply the exception (MPEP 2106.05(f)). Element (b) gathers the data on which the mathematical analysis is performed. It is insignificant extra-solution activity in the form of mere data gathering (MPEP 2106.05(g)). Element (c) recites "training deep learning models by using the prediction datasets" with no model architecture, training objective, loss function, or training procedure. It invokes a generic machine learning model as a tool to process the data. Claims that rely on generic machine learning, trained in the ordinary way, to analyze data in a particular environment do no more than apply the abstract idea. Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205, 1212-13 (Fed. Cir. 2025) (iterative training of a machine learning model using selected training material is incident to the very nature of machine learning and is not a technological improvement), cert. denied, No. 25-505, 2025 WL 3507020 (U.S. 2025); MPEP 2106.05(f)(2). Element (d) recites "retraining respective weights through transfer learning" on the clusters produced by the mathematical step. It identifies the result to be obtained — retrained weights — but not any particular way of obtaining it: the claim does not state which weights are retrained, from what starting values, against what objective, or how knowledge from the earlier training is retained. Reciting only the idea of a solution or outcome is mere instructions to apply the exception (MPEP 2106.05(f)(1)). The step uses the output of the mathematical grouping (the clusters) as the data for further generic training, which is applying the result of the abstract idea with a generic tool. The phrase "after the cluster classification step" only sequences the steps. Improvement consideration. The specification asserts two benefits. First, it asserts more accurate target (demand) predictions across items with diverse patterns (¶¶ [0009], [0019], [0058]-[0063], Table 1). An improvement in the accuracy of a forecast produced by mathematical analysis is an improvement in the abstract idea itself, not in the functioning of a computer or other technology (MPEP 2106.05(a)(II)). SAP Am., Inc. v. InvestPic, LLC, 898 F.3d 1161, 1163 (Fed. Cir. 2018) (where the claimed advance is a series of mathematical calculations on selected information, the advance lies entirely in the realm of abstract ideas, however great it is); Recentive, 134 F.4th at 1216 (claims that apply generic machine learning to a new data environment, without disclosing improvements to the machine learning models to be applied, are not eligible). Second, the specification asserts that the approach avoids underfitting caused by the reduced training data of each cluster (¶¶ [0010], [0020], [0064]). The specification attributes that result to a specific arrangement: a single model trained on all of the items is used as the pre-trained model, and the weights of the entire layer are retrained by fine-tuning for each cluster (¶ [0047]; see ¶¶ [0048]-[0049]). Claim 1 does not recite that arrangement. Claim 1 requires "training deep learning models by using the prediction datasets," which does not require that any model be trained on the data of all of the items; it reads on training several models, each on a different portion of the datasets. Claim 1 then requires only "retraining respective weights through transfer learning," which, as the specification describes transfer learning, includes reusing "a part or all" of a trained model and retraining "some or all of the layers" (¶ [0048]). The claim therefore covers configurations in which no model has learned from all of the data and only some weights are retrained — the configurations that do not provide the asserted benefit. For an improvement to be considered, the claim must include the components or steps of the invention that provide the improvement described in the specification (MPEP 2106.04(d)(1), 2106.05(a)). Claim 1 does not. Like the claims in Recentive, claim 1 does not delineate the steps through which the machine learning technology achieves the asserted improvement. Recentive, 134 F.4th at 1213. The claims differ from those in Ex parte Desjardins, Appeal No. 2024-000567 (PTAB Sept. 26, 2025) (Appeals Review Panel Decision) (precedential). There, the claims recited a particular training procedure — adjusting the model's parameters to optimize performance on a second task while protecting performance on a first task — and the specification explained how that procedure improved the model. Here, claim 1 recites no training objective or procedure; it recites generic training, a mathematical grouping, and generic retraining. The claims likewise differ from those in Enfish, LLC v. Microsoft Corp., 822 F.3d 1327, 1335-36 (Fed. Cir. 2016) (a specific self-referential table that improved the way the computer stores and retrieves data) and McRO, Inc. v. Bandai Namco Games Am. Inc., 837 F.3d 1299, 1313-16 (Fed. Cir. 2016) (specific rules that improved a technological process), because claim 1 does not recite a specific technical means but only the use of generic models on mathematically selected data. Example 39 of the 2019 subject matter eligibility examples is not analogous: that claim recited no judicial exception, whereas claim 1 recites clustering by SHAP values. Claim 1 is closer to Example 47, claim 2, in which the use of a trained neural network, without limits on how the network functions, was mere instructions to apply the exception and did not integrate it. Considered as a whole, claim 1 amounts to training generic models on data, grouping the data mathematically according to the models' feature attributions, and training generic models further on each group. The claim is directed to the abstract idea. Step 2B The additional elements, considered individually and in combination, do not amount to significantly more than the judicial exception. The generic processor (element (a)) and the generic training and retraining steps (elements (c) and (d)) are mere instructions to apply the exception for the reasons given at Prong Two, and that analysis is the same at Step 2B (MPEP 2106.05(f)). Inputting data (element (b)) is receiving data, a well-understood, routine, and conventional computer function (MPEP 2106.05(d)(II)). Training a neural network on a dataset, and transfer learning by fine-tuning a pre-trained network on a subset of data, were well-understood, routine, and conventional before the effective filing date. The specification describes them only in generic terms, which indicates that they were known to the person of ordinary skill (¶¶ [0035]-[0037], [0048]; MPEP 2106.05(d)(I)(2)(a)). The publications of record confirm it: Tzortzis describes warm-start initialization followed by fine-tuning as "a widely used and streamlined TL technique" (Tzortzis, § 4) and notes that transfer learning has been extensively studied, particularly in computer vision and image recognition (§ 1.1); Rathgamage Don builds its baseline by fine-tuning a publicly available pre-trained model on its own data (Rathgamage Don, § 3.1). See Berkheimer v. HP Inc., 881 F.3d 1360, 1368 (Fed. Cir. 2018); MPEP 2106.05(d)(I)(2)(c). The ordered combination — a model trained on pooled data, a grouping of the data, and a model trained or fine-tuned for each group — was itself a conventional arrangement, described across fields in Coletta (K-SHAP, ICML 2023, of record: §§ 3.1, 3.3 and § 4.2 — a model trained on all observations, K-Means clustering of the observations by their SHAP values, and a network trained on each cluster), Tzortzis (§§ 1.1, 2.5: pre-training on pooled data and transfer within clusters), Rathgamage Don (§ 3.2: fine-tuning a global model separately on each cluster), Lomont (US 2023/0267177 A1, ¶¶ [0029]-[0031]: a holistic model on all nodes and a model for each cluster), and Bandara (arXiv:1710.03222v1, Abstract: networks trained across all series and on clusters of similar series). The combination adds nothing beyond the sum of its generic parts and is not a non-conventional arrangement of the kind found in BASCOM Global Internet Servs., Inc. v. AT&T Mobility LLC, 827 F.3d 1341, 1350 (Fed. Cir. 2016). Claim 1 is not patent eligible. Dependent Claims 2-6 Each dependent claim incorporates the limitations of claim 1 and is analyzed individually below. Claim 2 adds "wherein each deep learning model comprises a Multi-Layer Perceptron (MLP) model or a Feed Forward Neural Network (FFNN)". This is an additional element. It names generic, well-known types of neural network without any particular configuration. Prong Two: it limits the generic tool used to apply the exception (MPEP 2106.05(f)) and does not integrate. Step 2B: MLP and feed-forward networks were well-understood, routine, and conventional forecasting models, as the specification's generic description (¶¶ [0035]-[0036]) and Tzortzis (§ 1.2: MLP architectures are frequently employed for forecasting) show. Not significantly more. Claim 3 adds "wherein the cluster classification step performs clustering and classification according to influence of variables corresponding to the SHAP values for each data of the prediction datasets". This limitation further describes the mathematical grouping of claim 1 (grouping according to the Shapley-value attribution of each variable) and is part of the judicial exception. It adds no additional element. Prong Two and Step 2B: no change from claim 1. Not eligible. Claim 4 adds "wherein the cluster-specific classification step performs the clustering by using the SHAP values, and classifies the prediction datasets for each of the plurality of clusters by applying a K-means clustering method". K-means clustering is a mathematical algorithm (assigning points to the nearest of K centers by Euclidean distance and recomputing centers as averages, ¶ [0043]); this limitation is part of the judicial exception and adds no additional element. Prong Two and Step 2B: no change from claim 1. Not eligible. Claim 5 adds "wherein the transfer learning step uses a neural network model as a pre-trained model, and performs fine tuning for each of the plurality of clusters". This is an additional element and is the limitation closest to integrating the exception, because it narrows "transfer learning" to fine-tuning a pre-trained network for each cluster. It does not cross the line. It does not require that the pre-trained model be the model trained on all of the data ("a neural network model", not the models of the base model training step), and it does not specify which weights are fine-tuned or how, so it still does not reflect the arrangement the specification credits with avoiding underfitting (¶ [0047]). Prong Two: generic fine-tuning is mere instructions to apply the exception (MPEP 2106.05(f)). Step 2B: fine-tuning a pre-trained network for a subset of data was well-understood, routine, and conventional (Tzortzis, §§ 1.1, 4; Rathgamage Don, §§ 3.1-3.2). Not eligible. Claim 6 adds "wherein the targets are demand". This limitation confines the predictions to a particular field of use (demand forecasting) and is an additional element that does not integrate the exception or add significantly more (MPEP 2106.05(h)). Predicting demand from historical sales data is, moreover, the kind of commercial forecasting that the specification describes as long performed with statistical models (¶¶ [0005]-[0006]). Not eligible. Claims 7-12 Under the 35 U.S.C. 112(f) construction set out above, each of claims 7-12 is directed to a computing device including a processor and a storage medium programmed to perform the method of the referenced claim and is therefore directed to a machine (Step 1). The processor and storage medium are generic computer components (¶¶ [0067]-[0069]) that add nothing beyond mere instructions to apply the exception on a computer (MPEP 2106.05(f)) and are well-understood, routine, and conventional (MPEP 2106.05(d)). Each claim is otherwise analyzed as its referenced method claim: Claim 7 (framework performing the method of claim 1): recites the abstract idea identified for claim 1; not integrated; not significantly more, for the reasons given for claim 1. Claim 8 (method of claim 2): adds the MLP/FFNN limitation; not eligible for the reasons given for claim 2. Claim 9 (method of claim 3): adds the grouping-by-influence limitation; part of the exception; not eligible for the reasons given for claim 3. Claim 10 (method of claim 4): adds K-means clustering; part of the exception; not eligible for the reasons given for claim 4. Claim 11 (method of claim 5): adds per-cluster fine-tuning of a pre-trained neural network; not eligible for the reasons given for claim 5. Claim 12 (method of claim 6): adds the demand field of use; not eligible for the reasons given for claim 6. Suggested Amendment The rejection may be overcome by amending claim 1 to recite the training arrangement that the specification credits with the asserted improvement, for example: (1) training a base deep learning model using the prediction datasets of all of the items (¶ [0047]); (2) calculating the SHAP values with that trained base model and clustering the items on per-item vectors of averaged absolute SHAP values (¶ [0042]); and (3) for each cluster, generating a cluster model by initializing it with the weights of the base model trained on all of the items and fine-tuning the weights of all layers using only the prediction datasets of that cluster (¶¶ [0047], [0049]), so that each cluster model is trained without the underfitting that results from training on the cluster data alone (¶ [0064]). Of the present claims, claim 5 comes closest; combining its fine-tuning limitation with items (1) and (3) would reflect the improvement described at ¶¶ [0047] and [0064]. Any such amendment would be evaluated under MPEP 2106.04(d)(1) and 2106.05(a). Applicant may also submit a subject matter eligibility declaration under 37 CFR 1.132 explaining how one of ordinary skill would understand the technical improvement described in the specification. 35 U.S.C. 112 35 U.S.C. 112(b) — Claims 4 and 10 The following is a quotation of 35 U.S.C. 112(b): "(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention." Claims 4 and 10 (and all claims depending from 4 and 10) are rejected under 35 U.S.C. 112(b) as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor regards as the invention. Claim 4 recites "the cluster-specific classification step." There is insufficient antecedent basis for this limitation. Claims 1-3 recite "a cluster classification step" but no "cluster-specific classification step," and it is unclear whether claim 4 refers to the cluster classification step of claim 1 or to a separate step. Claim 10 incorporates claim 4 and is rejected for the same reason. For examination, "the cluster-specific classification step" is interpreted as the cluster classification step of claim 1. Suggested fix: amend claim 4 to read "wherein the cluster classification step performs the clustering by using the SHAP values ...." 35 U.S.C. 112(d) — Claims 7-12 The following is a quotation of 35 U.S.C. 112(d): "(d) REFERENCE IN DEPENDENT FORMS.—Subject to subsection (e), a claim in dependent form shall contain a reference to a claim previously set forth and then specify a further limitation of the subject matter claimed. A claim in dependent form shall be construed to incorporate by reference all the limitations of the claim to which it refers." Claims 7-12 are rejected under 35 U.S.C. 112(d) as being of improper dependent form for failing to include all the limitations of the claim upon which they depend. Each of claims 7-12 refers to a method claim (claims 1-6, respectively) but claims a "framework" for performing that method. A claim in dependent form must include every limitation of the claim from which it depends, and it is improper if it could be infringed without infringing the claim from which it depends (MPEP 608.01(n)(III)). A framework that is capable of performing the method of claim 1 may be made, used, or sold without anyone performing the steps of claim 1; claim 7 may therefore be infringed without infringement of claim 1, and the same is true of claims 8-12 with respect to claims 2-6. Applicant may cancel the claims, amend the claims to place them in proper dependent form, rewrite the claims in independent form, or present a sufficient showing that the dependent claims comply with the statutory requirements. Suggested fix: replace claims 7-12 with an independent system claim, for example, "A target prediction system comprising: at least one processor; and a storage medium storing instructions that, when executed by the at least one processor, cause the at least one processor to perform: [the steps of claim 1]," with dependent claims corresponding to claims 2-6. Support appears at ¶¶ [0067]-[0069]. Rewriting in independent form will affect the claim fee calculation (37 CFR 1.16(h)). Other 35 U.S.C. 112 Matters Considered The following were considered and resolved without rejection. (1) Written description and enablement (35 U.S.C. 112(a)): claims 1-12 are supported by ¶¶ [0027]-[0050] and FIGS. 1-2, and the specification describes a working example with implementation details (¶¶ [0051]-[0064]); no rejection is made. (2) In claim 3, "for each data of the prediction datasets" is understood, in light of ¶ [0040], to mean for each data instance of the prediction datasets; the phrase is not indefinite. (3) In claim 5, "a neural network model as a pre-trained model" is broad but clear; it reads on any neural network used as the pre-trained model, including the models trained in the base model training step (¶¶ [0038], [0047]). (4) Claims 7-12 do not claim both an apparatus and a method of using it; the method is recited only as the function of the framework, so no rejection under IPXL Holdings, L.L.C. v. Amazon.com, Inc., 430 F.3d 1377, 1384 (Fed. Cir. 2005), is made. 35 U.S.C. 103 — OBVIOUSNESS 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 of Graham v. John Deere Co., 383 U.S. 1, 17-18 (1966), are applied below: the scope and content of the prior art, the differences between the prior art and the claims, the level of ordinary skill in the pertinent art, and any objective evidence of nonobviousness (MPEP 2141). No objective evidence of nonobviousness is of record. This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention. Prior Art of Record Applied in This Rejection Coletta — A. COLETTA, S. VYETRENKO, and T. BALCH, "K-SHAP: Policy Clustering Algorithm for Anonymous Multi-Agent State-Action Pairs," Proceedings of the 40th International Conference on Machine Learning (ICML 2023), Honolulu, Hawaii, PMLR vol. 202 (2023), arXiv:2302.11996v5 [cs.LG] (June 26, 2023) (21 pages) (of record, information disclosure statement of July 10, 2026). Prior art under 35 U.S.C. 102(a)(1): the version of record was publicly available by June 26, 2023, and the original version (arXiv:2302.11996v1) by February 23, 2023, both before December 1, 2023. Coletta trains a model on all of a set of observations, computes SHAP values for each observation from that model, groups the observations with K-Means applied to the SHAP values and trains a neural network on each resulting cluster. Cited by section, appendix, and page number of the 21-page version of record. Rathgamage Don — D. RATHGAMAGE DON, Y. XIE, L. YU, S. HUGHES, and Y. ZHU, "Cluster Language Model for Improved E-Commerce Retrieval and Ranking: Leveraging Query Similarity and Fine-Tuning for Personalized Results," arXiv:2309.14323v1 [cs.IR] (submitted Sept. 25, 2023) (accepted at the 6th Workshop on e-Commerce and NLP (ECNLP 6), KDD 2023). Prior art under 35 U.S.C. 102(a)(1). Rathgamage Don trains a baseline neural network model on all of the training data, clusters the data with K-Means, and fine-tunes the baseline model separately on each cluster to produce one model per cluster. Cited by section number. Tzortzis — A.-M. TZORTZIS, S. PELEKIS, E. SPILIOTIS, E. KARAKOLIS, S. MOUZAKITIS, J. PSARRAS, and D. ASKOUNIS, "Transfer Learning for Day-Ahead Load Forecasting: A Case Study on European National Electricity Demand Time Series," arXiv:2310.15555v1 [cs.LG] (submitted Oct. 24, 2023), National Technical University of Athens. Prior art under 35 U.S.C. 102(a)(1). Tzortzis forecasts national electricity demand with multi-layer perceptron (MLP) models and transfers the weights of models pre-trained on pooled data to target models that are then fine-tuned. Applied to claims 6 and 12 only. Cited by section number. Designation of References and Level of Ordinary Skill Coletta is the Primary reference. It teaches the data input, the training of a neural network on all of the data, the classification of the data into clusters by K-Means applied to SHAP values, and the training of a neural network on each cluster after the clustering; it teaches claims 2, 3, and 4 in full. Rathgamage Don is the first Secondary reference; it fills the single gap in claim 1, namely that each cluster's network is obtained by retraining the weights of the network trained on all of the data (transfer learning) rather than by training a network for the cluster on its own. Tzortzis is the second Secondary reference, applied only for the demand target of claims 6 and 12. A person of ordinary skill in the art would have had at least a bachelor's degree in computer science, electrical engineering, or a related field and about two years of experience training neural network models for prediction or forecasting, including transfer learning, model-explanation methods such as SHAP, and clustering. The applied references themselves reflect that level of skill. Okajima v. Bourdeau, 261 F.3d 1350, 1355 (Fed. Cir. 2001). Claims 1-5 and 7-11 are rejected under 35 U.S.C. 103 as being unpatentable over Coletta in view of Rathgamage Don. Claim 1 rejected under 35 U.S.C. 103 as being unpatentable over Coletta in view of Rathgamage Don. Claim 1 Claim 1 recites "A target prediction method performed by a computing device comprising at least one processor, the method comprising". Coletta teaches this feature. Coletta's method is computer-implemented: its models, SHAP computation, and clustering are implemented with the scikit-learn and PyTorch software libraries (Coletta, App. C, p. 15; App. D.2, p. 17), which execute on a computer having at least one processor. Rathgamage Don likewise performs its training and clustering on a cloud computing platform using an NVIDIA A100 GPU with Python and PyTorch (Rathgamage Don, § 4.2). Claim 1 recites "a data input step of inputting prediction datasets related to targets for prediction". Coletta teaches this feature. The input to Coletta's K-SHAP algorithm is a set of observations D, each a state-action pair (s_i, a_i) (Coletta, Algorithm 1, p. 5), drawn from multi-agent market data — simulated market data and historical market data for several NASDAQ stocks (§ 4, p. 6; § 4.2, p. 8). In each observation, the state features are the inputs and the action is the value to be predicted; Coletta's networks predict the next action given the current market state (§ 4.2, p. 8). The observations are thus prediction datasets related to targets for prediction (the agents' actions). Claim 1 recites "a base model training step of training deep learning models by using the prediction datasets input in the data input step". Coletta teaches this feature. Coletta trains a neural network over all of the state-action pairs (Coletta, § 4.2, p. 8), using a feedforward neural network with two hidden layers of 64 and 32 neurons (App. C, p. 15). K-SHAP also first learns a single world-policy model using all of the available observations (§ 3.1, p. 4; § 3.3 and Algorithm 1, line 1, p. 5), and Coletta discloses that the world-policy may itself be that feedforward neural network (App. C, p. 15; App. D.2, p. 17). A multi-layer feedforward neural network is a deep learning model under the broadest reasonable interpretation; claim 2 itself identifies an MLP or FFNN as a deep learning model. The networks trained on all of the observations are deep learning models trained using the input prediction datasets. Claim 1 recites "a cluster classification step of classifying the prediction datasets into a plurality of clusters by using SHapley Additive exPlanations (SHAP) values". Coletta teaches this feature. For each observation, Coletta computes a vector of SHAP (SHapley Additive exPlanations) values, each element being the average marginal contribution of a state feature to the trained model's prediction (Coletta, § 3.2, p. 5; Algorithm 1, lines 3-7). Coletta then states that "we adopt an existing clustering approach, namely K-Means, to group observations according to their SHAP values" (§ 3.3, p. 5; Algorithm 1, line 8), producing k clusters of observations. As interpreted above, claim 1 does not limit the unit that is clustered, and grouping the observations into k clusters by their SHAP values is classifying the prediction datasets into a plurality of clusters by using SHAP values. Claim 1 recites "a transfer learning step of inputting the plurality of clusters into the respective deep learning models and retraining respective weights through transfer learning after the cluster classification step". Coletta teaches, after the clustering, inputting each identified cluster into a respective neural network and training that network's weights: Coletta trains k neural networks over the identified clusters, one per cluster, all having the same architecture as the network trained over all of the state-action pairs, and measures the improvement in predicting the next action that the cluster networks achieve over the all-data network (Coletta, § 4.2, p. 8; § 4, p. 7 (Utility metric)). Coletta does not explicitly teach that each cluster network is obtained by retraining, through transfer learning, the weights of the network trained in the base model training step; Coletta does not say how the cluster networks are initialized. Primary prior-art reference Coletta does not explicitly teach "a transfer learning step of inputting the plurality of clusters into the respective deep learning models and retraining respective weights through transfer learning after the cluster classification step", in the sense of retraining the weights of the network trained on all of the data for each cluster through transfer learning; however, secondary prior-art reference Rathgamage Don teaches this feature as follows. Rathgamage Don describes "fine-tuning a global language model into cluster language models on individual clusters" (Rathgamage Don, Abstract). Rathgamage Don first trains a baseline model on all of the training queries and products (§ 3.1), then splits the queries into clusters with the K-Means algorithm, using embeddings produced by the baseline model (§ 3.2, "Generating Query Clusters"), and then, for each cluster, fine-tunes the baseline model on that cluster's data to generate a separate cluster model (§ 3.2, "Training Cluster Language Model on Individual Clusters"). Each cluster's data is thus input into a copy of the model trained on all of the data, and that copy's weights are retrained through transfer learning after the clustering step. Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art to modify Coletta so that each of its k cluster networks (Coletta, § 4.2, p. 8) is initialized with the trained weights of the network trained over all of the state-action pairs and then fine-tuned on the observations of its cluster, as Rathgamage Don fine-tunes its baseline model separately on each K-Means cluster (Rathgamage Don, § 3.2), so that the plurality of clusters are input into the respective trained deep learning models and the respective weights are retrained through transfer learning after the cluster classification step. The reason to do so comes from the references. Coletta trains a network on each cluster to obtain models that predict the action of each cluster more accurately than a single network trained on all of the observations (Coletta, § 4, p. 7; § 4.2, p. 8). Rathgamage Don teaches how to obtain such cluster models: it explains that a baseline model tuned on all of the data captures general relationships but may not be sensitive to the specific characteristics of each cluster, and that because the baseline parameters have already been tuned on all of the data, fine-tuning them on each cluster lets the model learn the local structure of that cluster efficiently and improves accuracy (Rathgamage Don, §§ 1, 3.2, 4.4). Rathgamage Don is reasonably pertinent to the problem the claims address — adapting a model trained on heterogeneous pooled data to groups of similar data without losing what the pooled model learned — and is in the same field of endeavor of training neural network models on clustered data (MPEP 2141.01(a)). One of ordinary skill would have had a reasonable expectation of success because Coletta's all-data network and cluster networks have the same architecture (Coletta, § 4.2, p. 8), so the trained weights of the one can be copied directly into each of the others; both references implement their networks in PyTorch (Coletta, App. D.2, p. 17; Rathgamage Don, § 4.2); and Rathgamage Don reports that the per-cluster fine-tuned models outperform the baseline (Rathgamage Don, § 4.4) (MPEP 2143.02). The rationale is MPEP 2143(I)(D), with (G): Coletta's method of training a network for each SHAP-derived cluster is a known method ready for improvement, and Rathgamage Don's fine-tuning of a model trained on all of the data separately for each cluster was a known technique that improved such methods in the same way, for a reason Rathgamage Don itself states. The modification does not change Coletta's principle of operation — Coletta still clusters the observations by their SHAP values and still trains a network for each cluster — or make it unsatisfactory for its intended purpose; it changes only the starting weights of each cluster network. The predictable result is one fine-tuned network per SHAP-derived cluster. KSR Int'l Co. v. Teleflex Inc., 550 U.S. 398, 416-17 (2007). This reason to combine comes from Coletta and Rathgamage Don themselves, not from applicant's disclosure. Coletta reports that, in its main experiments, a random-forest world-policy produced the best clustering, and that its feedforward network world-policy performed somewhat lower on its synthetic data (Coletta, § 4, p. 6; App. D.2, p. 17). This does not teach away from the combination. Coletta discloses the network world-policy as a working alternative that outperformed the random forest for some stocks (App. D.2, p. 17), and a reference's disclosure of a preferred embodiment does not teach away from its other disclosed embodiments (MPEP 2123(II)). Further, claim 1 does not require the SHAP values to be computed with the deep learning models, and the networks that are trained on all of the data and then retrained for each cluster in the combination are Coletta's neural networks of § 4.2, whatever model is used as the world-policy. The combination of Coletta and Rathgamage Don therefore teaches every limitation of claim 1. Claim 2 rejected under 35 U.S.C. 103 as being unpatentable over Coletta in view of Rathgamage Don. Claim 2 Claim 2 recites "wherein each deep learning model comprises a Multi-Layer Perceptron (MLP) model or a Feed Forward Neural Network (FFNN)". Coletta teaches this feature. Coletta's networks — the network trained over all of the state-action pairs, the cluster networks, and the network world-policy — are each a feedforward neural network with two linear hidden layers of 64 and 32 neurons and Leaky ReLU activations (Coletta, App. C, p. 15; App. D.2, p. 17; § 4.2, p. 8). A feedforward network with multiple fully connected hidden layers is both an FFNN and a Multi-Layer Perceptron. In the combination given for claim 1, the network trained on all of the data and its per-cluster fine-tuned copies keep that architecture. No further modification is needed; the combination and reasons are those given for claim 1. Claim 3 rejected under 35 U.S.C. 103 as being unpatentable over Coletta in view of Rathgamage Don. Claim 3 Claim 3 recites "wherein the cluster classification step performs clustering and classification according to influence of variables corresponding to the SHAP values for each data of the prediction datasets". Coletta teaches this feature. Coletta computes the SHAP values vector separately for each observation, each element describing how much a state feature (input variable) drives the model's predicted action for that observation (Coletta, § 3.2, p. 5; Algorithm 1, lines 3-7). Coletta explains that observations whose actions respond similarly to the input features have similar SHAP vectors and lie close together in the SHAP values space, while dissimilar behaviors lie in different regions, and it clusters the observations in that space (§ 3.3, pp. 5-6; Fig. 2, p. 6). Clustering each observation by its per-feature SHAP values is clustering and classification according to the influence of the variables corresponding to the SHAP values for each data of the prediction datasets. No further modification is needed; the combination and reasons are those given for claim 1. Claim 4 rejected under 35 U.S.C. 103 as being unpatentable over Coletta in view of Rathgamage Don. Claim 4 Claim 4 recites "wherein the cluster-specific classification step performs the clustering by using the SHAP values, and classifies the prediction datasets for each of the plurality of clusters by applying a K-means clustering method" (the "cluster-specific classification step" is interpreted as the cluster classification step; see the 35 U.S.C. 112(b) rejection above). Coletta teaches this feature. Coletta clusters the observations by applying K-Means to their SHAP values, assigning each observation to one of k clusters by minimizing the within-cluster sum of squares of the SHAP values (Coletta, § 3.3, Algorithm 1, line 8, and Eq. (5), p. 5), and selects k by the Elbow method when it is unknown (§ 3.3, p. 6; App. D.1, p. 16). Rathgamage Don likewise forms its clusters with K-Means (Rathgamage Don, § 3.2). No further modification is needed; the combination and reasons are those given for claim 1. Claim 5 rejected under 35 U.S.C. 103 as being unpatentable over Coletta in view of Rathgamage Don. Claim 5 Claim 5 recites "wherein the transfer learning step uses a neural network model as a pre-trained model, and performs fine tuning for each of the plurality of clusters". The combination given for claim 1 teaches this feature. In that combination, Coletta's feedforward neural network trained over all of the state-action pairs (Coletta, § 4.2, p. 8; App. C, p. 15) is the pre-trained model, and it is fine-tuned separately on each of the plurality of clusters, as Rathgamage Don fine-tunes its trained baseline model on each cluster (Rathgamage Don, Abstract; § 3.2). No further modification is needed; the combination and reasons are those given for claim 1. Claims 7-11 are rejected under 35 U.S.C. 103 as being unpatentable over Coletta in view of Rathgamage Don. Claims 7-11 Claims 7, 8, 9, 10, and 11 each recite "A target prediction framework for performing a target prediction method of claim [1, 2, 3, 4, and 5, respectively]." Under the 35 U.S.C. 112(f) construction above, the framework is a computing device including at least one processor and a storage medium programmed to perform the method of the referenced claim. Coletta implements its method in software (scikit-learn and PyTorch) executed on a computer (Coletta, App. C, p. 15; App. D.2, p. 17), and Rathgamage Don performs its training on an NVIDIA A100 GPU with Python and PyTorch and stores the trained cluster models in a model registry (Rathgamage Don, §§ 3.2, 4.2). The method steps of claims 1-5 are met for the reasons given above for claims 1-5, respectively, and the combination and motivation are the same. Claims 6 and 12 are rejected under 35 U.S.C. 103 as being unpatentable over Coletta in view of Rathgamage Don and Tzortzis. Claim 6 Claim 6 recites "wherein the targets are demand". Coletta's targets are the actions of agents in a financial market (Coletta, § 4, p. 6), and Rathgamage Don's are product purchases in e-commerce search (Rathgamage Don, § 3.2). Neither teaches demand as the target. Primary prior-art reference Coletta does not explicitly teach "wherein the targets are demand"; however, secondary prior-art reference Tzortzis teaches this feature. Tzortzis forecasts the day-ahead national electricity demand of 27 European countries (Tzortzis, Abstract; § 2.1) with MLP models (§ 2.4) and, to improve those forecasts, transfers the weights of MLP models pre-trained on the pooled demand data of other countries to target models that are then fine-tuned (§§ 1.1, 2.5). Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art to apply the method of Coletta as modified by Rathgamage Don to demand data, such as the national electricity demand data of Tzortzis (Tzortzis, § 2.1), so that the observations input to the method are demand observations and the targets predicted by the base and cluster networks are demand. The reason comes from the references. Coletta states that its SHAP-based approach can be used with any representation of the input state, including tabular data (Coletta, § 3.2, p. 5), so it is not limited to market data. Tzortzis teaches that demand forecasting with neural networks benefits from starting from models trained on pooled data of many similar series and fine-tuning them, especially where the target data are limited (Tzortzis, §§ 1.1, 1.3), and that grouping similar demand series by clustering improves that transfer on average (§ 3) — the same benefits the Coletta-Rathgamage Don combination provides by fine-tuning a network trained on all of the data for each cluster. Tzortzis is in the same field of endeavor as the claims, demand forecasting with neural networks (MPEP 2141.01(a)). One of ordinary skill would have had a reasonable expectation of success because Tzortzis's MLP demand models are the same kind of feedforward network that Coletta uses (Tzortzis, § 2.4; Coletta, App. C, p. 15), and SHAP values can be computed for any trained predictive model (Coletta, § 3.2, p. 5) (MPEP 2143.02). The rationale is MPEP 2143(I)(C), with (G): the clustering and per-cluster fine-tuning technique of Coletta and Rathgamage Don was known to improve the accuracy of neural network predictions for groups of similar data, one of ordinary skill would have recognized that it would improve Tzortzis's neural demand forecasts in the same way, and Tzortzis supplies the reason to seek that improvement. The predictable result is a per-cluster fine-tuned demand forecasting network. KSR, 550 U.S. at 417. This reason to combine comes from Coletta and Tzortzis themselves, not from applicant's disclosure. Claim 12 rejected under 35 U.S.C. 103 as being unpatentable over Coletta in view of Rathgamage Don and Tzortzis. Claim 12 Claim 12 recites "A target prediction framework for performing a target prediction method of claim 6." The framework is met for the reasons given for claims 7-11; Tzortzis further teaches executing its demand forecasting pipeline on a virtual machine with a GPU and CPU cores (Tzortzis, § 2.6). The method of claim 6 is met for the reasons given above for claim 6, and the combination and motivation are the same. Suggestions for Overcoming the Art The following amendments, each supported by the specification, would likely distinguish the applied art, subject to an updated search: • Recite how the SHAP values are aggregated and what is clustered: that the SHAP values are calculated with the base deep learning model for every instance of each item, that the absolute values are averaged per item to form a per-item SHAP vector, and that the items (rather than individual instances) are clustered with K-means on those per-item vectors (¶ [0042]). Coletta clusters individual observations, each by its own SHAP vector, and does not aggregate SHAP values over the instances of an item; Coletta would remain the closest art. • Recite that each cluster model is initialized with the weights of the base model trained on all items and that fine-tuning is applied to all layers using only that cluster's data (¶¶ [0047], [0049]), together with the per-item SHAP aggregation above. Rathgamage Don would remain the closest art for the per-cluster fine-tuning. Applicant is reminded that any amendment must be supported by the specification as filed (MPEP 2163.06). PERTINENT PRIOR ART The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. Lomont — US 2023/0267177 A1, published Aug. 24, 2023, Thomas LOMONT et al., assignee Mastercard International Incorporated (35 U.S.C. 102(a)(1)). Lomont identifies the same problem as applicant: a holistic prediction model generated from all nodes is not accurate for particular segments because it captures the overall trend rather than each segment's characteristics, while a model per node is computationally costly (Lomont, ¶¶ [0003]-[0004]). Lomont generates a primary prediction model from data aggregated over all nodes, clusters the nodes, and generates a corrective prediction model for each cluster that adjusts the primary model's output (Lomont, ¶¶ [0008], [0027]-[0031]). Lomont is pertinent to claim 1's "base model training step" and the per-cluster models of the "transfer learning step." It does not teach clustering by SHAP values, and its per-cluster models are separate corrective models rather than the primary model retrained through transfer learning. Koo — KR 10-2023-0081028 A, published June 7, 2023, Korea Institute of Industrial Technology (inventors KOO Jeong-in, KIM Bo-hyun, and JEONG Hong-jin) (35 U.S.C. 102(a)(1); of record in the IDS of July 10, 2026). Koo forecasts product demand. It first generates a first demand forecasting model by sequentially training on the data of all of the modeling-target products (Koo, ¶¶ [0060], [0064], [0066]); uses the permutation importance of the trend, seasonality, and residual components computed from that model to weight correlations between products (¶¶ [0060], [0062]); classifies the products into clusters by their average weighted correlations (¶¶ [0070]-[0073]); and generates a demand forecasting model for each cluster by sequentially transfer-learning the data of that cluster's products (¶ [0076]), on the stated basis that a cluster of similar products supplies more data than a single product group and so improves accuracy (¶ [0030]). Koo is pertinent to claims 1, 3, and 6: a model trained over all products, clustering informed by the influence of the input variables on that model, per-cluster models built by transfer learning, and demand as the target. It does not use SHAP values or K-means, its models are random-forest based rather than deep learning models (¶¶ [0065], [0076]), and it does not describe initializing each cluster's model with the weights of the first model. Desai — US 2024/0346338 A1, published Oct. 17, 2024, Sneha DESAI et al., assignee The Toronto-Dominion Bank, Appl. No. 18/301,413, filed Apr. 17, 2023 (35 U.S.C. 102(a)(2)). Desai computes, for each user, Shapley values representing the degree to which each feature of a trained predictive model contributed to the model output (Desai, ¶¶ [0030], [0063], [0067]), and assigns the users to clusters with a k-means clustering model using those Shapley values (¶¶ [0037], [0074], [0084], [0093]). Desai is pertinent to the "cluster classification step" of claim 1 and to claims 3 and 4. It does not train or retrain a model for each cluster, and it uses the clusters to customize recommendations rather than to train prediction models. Bandara — K. BANDARA, C. BERGMEIR, and S. SMYL, "Forecasting Across Time Series Databases using Long Short-Term Memory Networks on Groups of Similar Series," arXiv:1710.03222v1 [cs.LG] (submitted Oct. 9, 2017) (35 U.S.C. 102(a)(1)). Bandara reports that neural networks trained across all available time series can outperform univariate methods, that accuracy may degrade when the database is heterogeneous, and that training LSTM models on subgroups of similar series identified by time series clustering consistently outperforms the baseline model trained on all series (Bandara, Abstract). Bandara is pertinent to the "cluster classification step" and per-cluster models of claim 1 and to the problem stated at ¶ [0007] of applicant's specification. It does not use SHAP values and trains its per-cluster models without transfer learning from a model trained on all series — the configuration that applicant reports underfits (¶ [0064]). OTHER REJECTIONS No restriction or election requirement is made; claims 1-12 are drawn to a single inventive concept. No other ground of rejection is made. ALLOWABLE SUBJECT MATTER No claim is indicated as containing allowable subject matter at this time. The amendments identified under Suggestions for Overcoming the Art would be considered after an updated search. SUGGESTED INTERVIEW The examiner invites applicant's representative to schedule a telephone or video interview before filing a response. An interview may be requested using the USPTO Automated Interview Request (AIR) form. Suggested topics: Subject matter eligibility (claims 1-12). The Suggested Amendment in the 35 U.S.C. 101 rejection: reciting a base model trained on the prediction datasets of all items, SHAP values calculated with that model and clustered per item (¶ [0042]), and cluster models initialized with the base model's weights and fine-tuned in all layers on each cluster's data only (¶¶ [0047], [0049], [0064]). Claim 5 is the closest present claim. A subject matter eligibility declaration may also be discussed. Prior art (claims 1-12). Amendments reciting the per-item aggregation of absolute SHAP values computed with the base model and K-means clustering of items on the resulting per-item vectors (¶ [0042]), together with initializing each cluster model with the base model's weights and fine-tuning all layers on that cluster's data (¶¶ [0047], [0049]). Coletta and Rathgamage Don would remain the closest art. Claims 7-12 under 35 U.S.C. 112(d) and 112(f). Replacing claims 7-12 with an independent system claim reciting at least one processor and a storage medium storing instructions for the method steps (¶¶ [0067]-[0069]), which would resolve the 35 U.S.C. 112(d) rejection and avoid the 35 U.S.C. 112(f) interpretation of "framework." Claim 4 under 35 U.S.C. 112(b). Changing "the cluster-specific classification step" to "the cluster classification step." Applicant is requested to send any proposed claim amendments to the examiner before the interview, so that the examiner can state for each proposed amendment whether it would overcome each outstanding rejection and why (MPEP 713.01, 713.04). CONCLUSION Contact Information Any inquiry concerning this communication or earlier communications from the examiner should be directed to MATTHEW T. SITTNER whose telephone number is (571) 270-7137 and email: matthew.sittner@uspto.gov. The examiner can normally be reached on Monday-Friday, 8:00am - 5:00pm (Mountain Time Zone). Please schedule interview requests via email: matthew.sittner@uspto.gov If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Sarah M. Monfeldt can be reached on (571) 270-1833. Information regarding the status of an application may be obtained from the Patent Application Information Retrieval (PAIR) system. Status information for published applications may be obtained from either Private PAIR or Public PAIR. Status information for unpublished applications is available through Private PAIR only. For more information about the PAIR system, see http://pair-direct.uspto.gov. Should you have questions on access to the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative or access to the automated information system, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000. /MATTHEW T SITTNER/ Primary Examiner, Art Unit 3629b
Read full office action

Prosecution Timeline

Oct 28, 2024
Application Filed
Oct 29, 2026
Non-Final Rejection mailed — §101, §103, §112 (current)

Precedent Cases

Applications granted by this same examiner with similar technology

Patent 12784118
MACHINE EVALUATION OF CONTRACT TERMS
1y 10m to grant Granted Nov 03, 2026
Patent 12755245
DISTRIBUTED ORGANIZATIONAL DATA CONSTRUCTION METHOD AND ORGANIZATIONAL DATA CONSTRUCTION BACKEND
2y 0m to grant Granted Oct 06, 2026
Patent 12743703
SYSTEMS AND METHODS FOR DIGITAL ONBOARDING USING ERP DATA
3y 3m to grant Granted Sep 22, 2026
Patent 12737731
COMMERCIAL VEHICLES ROTOR CRACKING PREDICTION USING RECURRENT NEURAL NETWORK
3y 6m to grant Granted Sep 15, 2026
Patent 12718127
METHOD AND SYSTEM FOR RECOMMENDING OPTIMUM COMBINATION OF QUANTUM CIRCUITS
3y 1m to grant Granted Aug 25, 2026
Study what changed to get past this examiner. Based on 5 most recent grants.

Strategy Recommendation AI-generated — please review before filing

Get a prosecution strategy drawn from examiner precedents, rejection analysis, and claim mapping.
Typically takes 5-10 seconds — AI-generated, attorney review required before filing

Prosecution Projections

1-2
Expected OA Rounds
58%
Grant Probability
99%
With Interview (+56.0%)
3y 0m (~1y 0m remaining)
Median Time to Grant
Low
PTA Risk
Based on 911 resolved cases by this examiner. Grant probability derived from career allowance rate.

Sign in with your work email

Enter your email to receive a magic link. No password needed.

Personal email addresses (Gmail, Yahoo, etc.) are not accepted.

Free tier: 3 strategy analyses per month