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 . This action is made final.
This office action is in response to the amendments filed on January 31, 2026.
Claims 1, 2, 3, and 8 have been amended. Claims 4-7 have been cancelled.
Response to Amendment
The amendments filed on January 31, 2026 has been entered. Claims 1, 2, 3, and 8 remain pending in the application.
Priority
Acknowledgment is made of applicant’s claim for foreign priority under 35 U.S.C. 119 (a)-(d). The certified copy has been filed in parent Application No. CN2022105338057, filed on November 23, 2022.
Response to Arguments
Acknowledgement is made of applicant’s amendments to the previous objections. The objections have been withdrawn.
Applicant's arguments filed January 31, 2026. have been fully considered but they are not persuasive.
Response to 101 arguments
Applicants argues on page 11
Claims 1 - 8 have been rejected under 35 U.S.C. § 101 as allegedly being directed to an abstract idea without significantly more.
It is respectfully submitted that in view of the currently presented amendments, the rejection has been overcome. In particular, Applicant has cancelled Claims 4 - 7 without prejudice or disclaimer and amended Claims 1 - 3 and 8 to overcome over rejections and further specify the invention. The features of Claims 4 - 7 have been incorporated into Claim 1. The embodiments of the present invention as currently claimed are not directed at an abstract idea but are computer-implemented practical applications that significantly improve recommending user preferences based on improved neural networks as explained in the specification in details.
Amended Claims 1 - 7 define computer-implemented method. Claim 8 is directed as "[a] user preference recommendation device based on neural set operations, comprising a non-transitory memory, a processor, and a computer program stored in said memory and executable on said processor, characterized in that said memory stores the neural set operations model constructed by the user preference recommendation method based on neural set operations". The present invention solved a specific technical problem by proposing automated specific technical solution, including specific hardware and data transformation.
Examiners response
Under Step 2A, Prong one, Claim 1 recites abstract ideas in the form of mental, and mathematical concepts, including mathematical calculations and formulas/equations. The claim recites encoding positive and negative feedback interaction sequences into vectors, encoding user global preference information into a vector, performing recited set operations on those vectors, mapping representations using a multi-layer perception, calculating cosine similarity, and optimizing using a recited loss function and regularization expression. The MPEP identifies mathematical relationships, mathematical formulas or equations, and mathematical calculations as abstract ideas.
Applicants’ reliance on generic computer implementation is also not persuasive. Although the claims are framed as a computer-implemented and claim 8 recites a processor and non-transitory memory, merely implementing an abstract idea on a generic computer or limiting the idea to a particular technological environment foes not integrate the exception into a practical application.
Applicants argues on pages 11-12
First, Applicant would like to bring the Examiner's attention to the newly PTAB precedential opinion designated by USPTO Director on November 4, 2025 after the present Office Action. The PTO Director "designated the Desjardins decision as precedential to 1) ensure the case reasoning binds all examination and appeals activity, and 2) underscore that improvements in computational performance, learning, storage, data sets and structures, for example, can constitute patent-eligible technological advancements under the Alice framework." See USPTO Memorandum to Patent Examining Corp., Subject Matter Eligibility Declarations, Dated December 4, 2025.
Just like the application in the Appeals Review Panel (ARP) decision In re Deljardins, Appeal No. 2024-000567 (PTAB September 26, 2025, Appeals Review Panel Decision) vacating a PTAB sua sponte Section 101 rejection, the present application utilizes both "the positive feedback interaction sequence and negative feedback interaction sequence, and the information is not compressed to ensure the accuracy of the information, which can improve the accuracy of user preference recommendation." (see page 4, lines 25 - 29). Therefore, the claims were directed to using a neural network embedded with ensemble operations to learn the comprehensive preference representation of a user and perform preference recommendation based on such user preference representation, improving the accuracy of the recommendation, which properly integrated an otherwise abstract idea into a practical application.
Accordingly, the claims satisfied Step 2 of the framework set forth in Alice Corp. v. CLS Bank Int'l, 573 US. 208 (2014). We specifically credited the claims for improving the functioning of the neural networks themselves by using the neural set operations model including set operation and multi-layer perceptron, and by not compressing the information to ensure the accuracy of the information. recommendation, which can improve the accuracy of user preference recommendation. Just like In re Deljardins, the present invention also reduced storage requirements, lowered system complexity, and the improvement in accuracy.
Additionally, Desjardin reaffirmed that 35 U.S.C. §§ 102, 103, and 112- rather than§ 101- are the proper inquiries for defining the scope of patent protection. As stated below, the present application is also non-obvious over prior art under 35 U.S.C. § 103.
Examiners response
This argument is not persuasive because the alleged improvement is directed to the abstract mathematical model for recommendation itself, rather than to an improvement in the functioning of a computer or another technology. The claim does not recite a specific improvement to computer architecture, memory operation, networking, database structure, or other computer functionality. Instead, the claim recites mathematical processing of user information to produce recommendation results. Under Step 2A, Prong Two, the question is whether the claim as a whole integrates the judicial exception into a practical application by imposing a meaningful limit on the exception. Here the additional elements do not do so.
Further, the asserted “accuracy”, “storage”, and “complexity” improvements are describing at a high level and are not tied in the claims to any specific technical mechanism, improving computer operation. The claim remains focused on collecting user interaction data, transforming it into mathematical representations, applying mathematical operations to those representations, and ranking/recommending items based on the resulting calculations.
Applicants argues on pages 11-12
As indicated previously, Applicant respectfully submits that the Examiner failed to look at the additional limitations as an ordered combination. In fact, the Examiner improperly focused on each additional limitation individually. The specific features of the neural set operations model including set operation and multi-layer perceptron, which are the great advantages of the present invention over prior art.
In addition, the additional limitations are absolutely NOT "merely add[ing] insignificant extra-solution activity to" the invention. The combination of limitations ARE the solution of the present invention, rather than EXTRA-solution activity. The additional limitations as defined by the pending claims, include using specific features of the neural set operations model including set operation and multi-layer perceptron, etc. The invention as a whole amount to significantly more than the judicial exception.
As explained in the specification, the present invention achieved significant more than abstract idea and prior art: "[c]ompared to the prior art, the present invention owns beneficial effects including at least:
On the basis of obtaining the positive feedback interaction sequence and negative feedback interaction sequence and the user global preference information, combining both the positive feedback interaction sequence and the negative feedback interaction vector sequence corresponding to the positive feedback interaction sequence and the negative feedback interaction vector sequence, and after using the set operation to obtain the positive feedback preference representation and the negative feedback preference representation, the multi-layer perceptron is used to map the positive feedback preference representation and negative feedback preference representation and the global user preference embeddings corresponding to the user global preference information to obtain the user integrated preference representation, so that the user integrated preference representation includes the positive feedback and negative feedback information, and the information is not compressed to ensure the accuracy of the information. recommendation, which can improve the accuracy of user preference recommendation." (see paragraphs from page 4 line 17 - 29).
Examiners response
The rejection considers the claims as a whole, including the recited mathematical limitations and the additional computer-imp[lamented elements together. The MPEP states that the Prong Two analysis considers the claim as a whole and evaluates how the limitations interact rather than analyzing additional elements in a vacuum.
Even when considered as an ordered combination, however, amended claim 1 remains directed to mathematical processing of information for recommendation. The claim as a whole reciters receiving/extracting interaction information, encoding the information into vectors, applying recited set operations and MLP mapping, calculating similarity, and using the result to rank/recommend items. These steps do not integrate the exception into a practical application, but instead use a generic computer as a tool to perform the abstract mathematical analysis. To the extent the claim recites recommending items based on the calculated score, such outputting of results is also insufficient to conger eligibility. The MPEP further explains that insignificant extra-solution activity, including mere data gathering or outputting, does not amount to eligibility-conferring integration.
Applicants’ arguments on pages 14-17
Furthermore, Claim 8 is directed at a device based on neural set operations with processors and a non-transitory computer-readable medium, respectively. The processors and computer-readable medium achieve more accurate user preference recommendation than the prior art in a non-routine manner as practical applications.
Applicant respectfully submits that the present invention is not a simple abstract concept such as mathematical algorithms or mental processes, but a specific application of real-time user interaction in large-scale recommendation systems to improve the accuracy of user preference recommendations in specific technical environments. The specific descriptions in the original manuscript include:
As described in the first sentence of the background art of this application, the method and apparatus of this application are a recommendation system, particularly for online service platforms to achieve personalized recommendations for users through preference recommendations.
For example, in the first sentence below step 1 in the specific implementation, the recommendation method or device is used in applications of interactive platforms with recommendation functions, such as shopping platforms or community platforms.
Positive feedback interaction items refer to items that users click or like; negative feedback interaction items refer to items that users do not click or dislike.
Moreover, the present invention as defined is not merely gathering data, but rather using the data to accurately reflect reality of all positive feedback interaction items and negative feedback interaction items, just like the automatic lip synchronization and facial expression animation using computer-implemented rules in McRo Inc. v. Bandai Namco Games America, Inc.. 837 F.3d 1299 (Fed. Cir., 2016).
The courts have also found that improvements in technology beyond computer functionality may demonstrate patent eligibility. In McRO, the Federal Circuit held claimed methods of automatic lip synchronization and facial expression animation using computer-implemented rules to be patent eligible under 35 U.S.C. 102, because they were not directed to an abstract idea. McRO, 837 F.3d at 1316, 120 USPQ2d at 1103. The basis for the McRO court's decision was that the claims were directed to an improvement in computer animation and thus did not recite a concept similar to previously identified abstract ideas. Id.
The amended claims included meaningful limitations that are significantly more than the alleged abstract idea. Applicant respectfully calls the Examiner's attention again to Example 42 of The Subject Matter Eligibility Examples (Method for Transmission of Notifications When Medical Records Are Updated). The amended claims of the present invention recite a combination of additional elements. The claims as a whole are practical applications including specific improvement over prior art. Thus, the claims are eligible for patent protection and are not directed to abstract ideas.
Please refer to the following analysis of Example 40 of The Subject Matter Eligibility Examples (Adaptive Monitoring of Network Traffic Data), wherein under step 2A analysis, the claim is integrated into a Practical Application because "[t]he claim recites the combination of additional elements of collecting at least one of network delay, packet loss, or jitter relating to the network traffic passing through the network appliance, and collecting additional Netflow protocol data relating to the network traffic when the collected network delay, packet loss, or jitter is greater than the predefined threshold."
Furthermore, as pointed out in MPEP2106.05(e), when evaluating whether additional elements meaningfully limit the judicial exception, it is particularly critical that examiners consider the additional elements both individually and as a combination. When an additional element is considered individually by an examiner, the additional element may be enough to qualify as "significantly more" if it meaningfully limits the judicial exception. However, even in the situation where the individually-viewed elements do not add significantly more, those additional elements when viewed in combination may amount to significantly more than the judicial exception by meaningfully limiting the exception. See Diamond v. Diehr, 450 U.S. 175, 188, 209 USPQ2d 1, 9 (1981) ("a new combination of steps in a process may be patentable even though all the constituents of the combination were well known and in common use before the combination was made"); BASCOM Global Internet Servs. V. AT&T Mobility LLC, 827 F.3d 1341, 1349, 119 USPQ2d 1236, 1242 (Fed. Cir. 2016). It is important to note that, when appropriate, an examiner may explain on the record why the additional elements meaningfully limit the judicial exception.
Moreover, in Thales Visionix, the particular configuration of inertial sensors and the particular method of using the raw data from the sensors was more than simply applying a law of nature. Thales Visionix, Inc. v. United States, 850 F.3d 1343, 1348-49, 121 USPQ2d 1898, 1902 (Fed. Cir. 2017). The court found that the claims provided a system and method that "eliminate[d] many 'complications' inherent in previous solutions for determining position and orientation of an object on a moving platform." In other words, the claim recited a technological solution to a technological problem. Id.
Examiners should keep in mind that the courts have held computer-implemented processes to be significantly more than an abstract idea (and thus eligible), where generic computer components are able in combination to perform functions that are not merely generic. DDR Holdings, LLC v. Hotels.com, LP, 773 F.3d 1245, 1258-59, 113 USPQ2d 1097,
1106-07 (Fed. Cir. 2014).
Finally, as pointed out in the September 26, 2025 decision in Ex parte Desjardins, Appeal 2024-000567, Application 16/319,040, ("[t]his case demonstrates that §§ 102, 103 and 112 are the traditional and appropriate tools to limit patent protection to its proper scope. These statutory provisions should be the focus of examination."). The present application s free of prior a1t, as recognized by the Examiner.
Therefore, the rejection under 35 U.S.C. § 101 has been overcome. Accordingly, withdrawal of the rejections under 35 U.S.C. § 101 is respectfully requested.
Examiners response
The present claims are unlike the claims found eligible in McRO, Thales, BASCOM, or DDR because the present claims do not recite a specific improvement in computer functionality or another technology. Rather, the claims recite mathematical manipulation of user preference data for recommendation. Stating that the claim is sued in a recommendation system, online platform, or other technological environment does not by itself integrate the exception into a practical application. The MPEP explains that merely linking the use of the judicial exception to a particular technological environment or using a computer as a tool is insufficient.
Applicants’ references to “real-time user interaction”, “large-scale recommendation systems’, and application contexts such as shopping or community platforms do not change the character of the claim. The claim language itself remains focused on abstract mathematical processing and result generation. Accordingly, the claim does not recite a technological solution comparable to the eligible claims in the cases and examples cited by the applicant.
Claim 8 merely recites generic computer components, namely a processor, non-transitory memory, and executable program instructions for carrying out the same abstract mathematical operations discussed above. Generic recitation of a processor and memory does not integrate the judicial exception into a practical application and does not render the claim patent eligible.
Regarding the 103 arguments
Applicant argues on page 17-18
Claims 1 - 4 and 8 have been rejected under 35 U.S.C. §103(a) as allegedly being obvious over Chenzhong et al. (CN110555112) in view of Ren et al. ("Beta Embeddings for Multi-Hop Logical Reasoning in Knowledge Graphs"). Claim 5 has been rejected under 35
U.S.C. §103(a) as allegedly being obvious over Chenzhong in view of Ren, and Sarwar et al ("Item-based Collaborative Filtering Recommendation Algorithms"). Claim 6 has been rejected under 35 U.S.C. §103(a) as allegedly being obvious over Chenzhong in view of Ren, and Rendle et al. (BPR: Bayesian Personalized Ranking from Implicit Feedback). Claim 7 has been rejected under 35 U.S.C. §103(a) as allegedly being obvious over Chenzhong in view of Ren, and Rendle et al. and He et al. (Neural Collaborative Filtering).
Applicant traverses the rejection and respectfully submits that the presently claimed invention is not obvious over the cited references.
The claimed subject matter is not disclosed nor suggested by prior art including Chenzhong, Ren, Sarwar, Rendle and He. There are significant differences between the present invention as claimed and the cited references. Even if these references are combined, they do not teach or suggest the present invention. There is no predictable success in the combination of references for a person of ordinary skill in the art.
More specifically, in Chenzhong: the "neural assembly operation model" and its specific structure as required by the present invention have not been disclosed. In Ren, although it involves set operations, it is not used for time series modeling of positive and negative user feedback. In Sarwar, only involves traditional collaborative filtering, without combining neural ensemble operations. Rendle and He involves loss function, but does not combine the set operation and multi-source preference fusion mechanism of the present invention.
Examiners Response
The rejection does not rely on ant one reference to teach the entire claimed invention. Chenzhong teaches positive and negative feedback interaction sequences, user global preference information, corresponding vector representations, and ranking/recommendation framework. Ren teaches the neural set-operations model over embeddings, including set-operator transformations used to obtain the claimed positive and negative feedback preference representations. Sarwar teaches cosine-similarity scoring, Rendle teaches Bayesian personalized ranking loss with regularized optimization, and He teaches neural embedding interaction and scoring through multi-layer perception-based architectures. Accordingly, the rejection relies on the combined teachings of the references for their respective limitans, and Applicant’s attack on the references individually is therefore not persuasive.
Applicant argues on page 18-21
Combining neural network ensemble operations with user positive and negative feedback sequences is not a simple patchwork, but a deep coupling at the mechanistic level to solve the inherent challenges of adaptability and reliability in dynamic intelligent systems. This combination is non-obvious:
The divergence between technical objectives and solutions
The traditional core purpose of neural network ensembles is to improve the accuracy and stability of static tasks (such as classification and prediction), and their optimization goal is static "model averaging" or "variance reduction".
The core of user feedback sequence processing is to capture dynamic, temporal preference evolution, and its mainstream technologies are sequence modeling such as recurrent neural networks and Transformers.
Before their integration, the two belonged to different technological paradigms. Those skilled in the art lack the motivation to actively adapt tools (model sets) primarily used for static robustness to solve dynamic evolutionary problems (feedback sequences), because the objective functions and data types optimized by the two are fundamentally different.
Existing technologies lack integration with teaching methods.
In typical application areas such as recommender systems and dialogue systems, existing literature shows that researchers either focus on improving the ensemble algorithm itself (such as new Bagging/Boosting variants) or on more refined modeling of user sequences (such as using attention mechanisms).
While a few studies have mentioned utilizing feedback, they typically involve interacting the output of a single, aggregated model with the feedback, rather than allowing the feedback sequence to dynamically guide and adjust the collaborative relationships and weight allocation
among the multiple models within the ensemble. This deep, mechanistic interaction has not been fully revealed or taught before integration.
3. There are technical obstacles to integration
Simple combination leads to significant computational complexity and decision latency, which contradicts the requirements of online systems. Designing a lightweight mechanism that allows feedback sequences to influence ensemble decisions in real time and efficiently is itself a technical problem requiring creative solutions. This suggests that combination is not obvious, otherwise, the approach would have already been widely adopted.
The synergistic effect of neural network ensemble operations and user positive and negative feedback sequences is reflected in:
Neural network ensemble operations can handle the "ensemble" semantics of user historical behavior, while sequence modeling can capture the temporal dependencies of behavior. In recommender systems, user behavior sequences provide context, and ensemble modeling can capture the diversity of user preferences (e.g., a user likes both movies A and B, but watched A first and then B in the sequence). The fusion of these two approaches can improve the understanding of multi-dimensional user preferences.
Neural network ensemble operations excel at handling sparse data (such as a collection of items that a user has only viewed a small number of), while sequence modeling is adept at capturing long-term dependencies and interest shifts. Combining the two can improve recommendation accuracy while maintaining diversity.
By analyzing the differences in responses of different models in the set to users' historical feedback sequences, the basis for recommendation decisions can be traced, providing a certain degree of interpretability.
This combination brings overall advantages that go beyond the performance of a single component. In short, this combination is not a simple addition, but a deep coupling of the two through mechanism design, creating a smarter system that can dynamically adapt, assess its own
credibility, and understand the evolution of user preferences.
The present invention proposes a user preference modeling method based on neural set operations. Its core lies in first dynamically selecting the k most recent items from the user's positive and negative feedback interactions to construct a basic preference representation, rather than uniformly sampling the entire interaction history. Based on this, it innovatively transforms set operations into a function form that can be learned by a neural network. Utilizing the distributed representation capability and universal approximation theorem of neural networks, it effectively distinguishes between user likes and dislikes. A deep neural network architecture built using multilayer perceptron’s is employed to perform nonlinear transformation and fusion of the preference representation using fully connected mappings between layers, ultimately generating a comprehensive user preference representation. This neural set operation model uses a constructed composite loss function and performs end-to-end parameter optimization through backpropagation, thereby continuously improving the accuracy and personalization of the recommendation system.
The algorithms and block diagrams described in this paper were refined and abstracted after being deployed and verified in actual computer systems, and thus possess inventiveness in the sense of patentability.
According to MPEP 2143.01, the mere fact that references can be combined or modified does not render the resultant combination obvious unless the results would have been predictable to one of ordinary skill in the art. KSR International Co. v. Teleflex Inc., 82 USPQ2d 1385, 1396 (2007). "Where the prior art, at best gives only general guidance as to the particular form of the claimed invention or how to achieve it, relying on an obvious-to-try theory to support an obviousness finding is impermissible," In re Cyclobenzaprine Hydrochloride Extended-Release Capsule Patent Litigation, 676 F.3d 1063, 1073 (Fed. Cir. 2012). Further, "KSR did not create a presumption that all experimentation in fields where there is already a background of useful
knowledge is 'obvious to try,' without considering the nature of the science or technology," Abbot Labs. v. Sandoz, Inc., 544 F.3d 1341, 1352 (Fed. Cir. 2008).
In summary, the cited references fail to disclose or suggest the features of the presently amended claims. In addition, the cited references fail to provide motivation to modify or combine them to solve the problem of the present invention. Even if they are modified or combined, it will not render the present claimed invention obvious. One of ordinary skill in the art would not discern the present invention as claimed at the time of its invention.
Therefore, the newly presented claims are not obvious over the cited references and the rejection under 35 U.S.C. §103(a) has been overcome. Accordingly, withdrawal of the rejection under 35 U.S.C. § 103(a) is respectfully requested.
Examiners Response
The rejection relies on known techniques being applied according to their established functions in the recommender-system context. Chenzhong teaches positive and negative feedback sequence modeling and user preference representations, Ren teaches the neural set-operations over embeddings, Sarwar teaches cosine-similarity scoring, Rendle teaches regularized ranking optimization for implicit-feedback recommendation, and He teaches neural embedding interaction and scoring. Combining these teachings would have been predictably yielded a recommender system that generates learned preference representations from positive, negative, and global preference signals and optimizes those representations for ranking. Applicant’s assertions regarding “deep coupling,” “different paradigms,” and alleged technical obstacles are attorney argument unsupported by evidence and do not outweigh the express teachings of the references and the articulated rationale for combination.
Applicant’s additional assertions that the3 cited references fail to disclose or suggest the amended claims, fail to provide motivation to combine, and would not have rendered the claimed invention obvious likewise are not persuasive for the reasons discussed above, because the rejection relies on the combined teachings of the references for their respective limitations and on an articulated rationale for combining them to achieve predictable results.
Claim Rejections - 35 USC § 101
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.
The text of those sections of Title 35, U.S. Code not included in this action can be found in a prior Office action.
To determine if a claim is directed to patent ineligible subject matter, the Court has guided the Office to apply the Alice/Mayo test, which requires:
Step 1: Determining if the claim falls within a statutory category.
Step 2A: Determining if the claim is directed to a patent ineligible judicial exception consisting of a law of nature, a natural phenomenon, or abstract idea; and Step 2A is a two prong inquiry. MPEP 2106.04(II)(A). Under the first prong, examiners evaluate whether a law of nature, natural phenomenon, or abstract idea is set forth or described in the claim. Abstract ideas include mathematical concepts, certain methods of organizing human activity, and mental processes. MPEP 2104.04(a)(2). The second prong is an inquiry into whether the claim integrates a judicial exception into a practical application. MPEP 2106.04(d).
Step 2B: If the claim is directed to a judicial exception, determining if the claim recites limitations or elements that amount to significantly more than the judicial exception. (See MPEP 2106).
Claims 1, 2, 3, and 8 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Step 1: Claims 1, 2, and 3 are directed to a method (a process), Claim 8 is directed to a device comprising a memory, a processor, and a computer program (a machine). Therefore, Claims 1, 2, 3, and 8 are directed to a process, machine or manufacture or composition of matter.
Regarding claim 1
Step 2A Prong 1
Claim 1 recites the following mental processes, that in each case under the broadest reasonable interpretation, covers performance of the limitation in the mind (including observation, evaluation, judgement, opinion) or with the aid of pencil and paper but for recitation of generic computer components (e.g., “computing system”, “predictive model”, and “supervised learning algorithm”) [see MPEP 2106.04(a)(2)(III)].
“extracting all positive feedback interaction items and negative feedback interaction items from user interaction history information, and constructing positive feedback interaction sequences and negative feedback interaction sequences respectively by arranging all positive feedback interaction items and negative feedback interaction items in the order of interaction time” (e.g., a human can compare and identify positive and negative feedback and order them based on date)
“encoding positive and negative feedback interaction sequences into a sequence of positive and negative feedback interaction vectors, respectively, and encoding user global preference information into a user global preference vector”(e.g., a human can sort the positive and negative feedback into respective positive and negative tables)
“calculating similarity between the comprehensive user preference representation and the corresponding vector of multiple candidate interaction items and the similarity is used as the recommendation score to rank the candidate interaction items, and recommending the user preference based on the ranking result” (e.g., a human can compare data from a model output, then rank it based on preferred outcomes)
Claim 1 further recites the following mathematical process, that in each case under the broadest reasonable interpretation, involves mathematical relationships, formulas, calculations, or algorithms implemented using generic computer components (e.g., “processor”, “ conditional generative model”, “graph neural network”) [see MPEP 2106.04(a)(2)(I)].
“first connecting the positive feedback preference representation, the negative feedback preference representation, and the global user preference embeddings, and then computing the mapping of the concatenated results using a multi-layer perceptron in order to obtain the comprehensive user preference representation, denoted as:
e
u
i
=MLP(concat(
e
u
i
+
,
e
u
i
-
,mi)) (5)
wherein
e
u
i
denotes the comprehensive user preference representation, and ui denotes the i-th user, and
e
u
i
+
denotes the positive feedback preference representation, and
e
u
i
-
denotes the negative feedback preference representation, and mi denotes the global user preference embeddings for the ith user, and concat( ) denotes the tandem operation, and MLP( ) denotes the mapping computation corresponding to the multi-layer perceptron” (e.g., tensor concatenation, linear algebra and multivariate calculus)
“calculating the cosine similarity between the comprehensive user preference representation and the corresponding vector of multiple candidate interaction items respectively, and using the cosine similarity as the recommendation score” (e.g., scalar similarity metric, sorting/ranking to determine a comparison)
“characterized in that said the neural set operations model is subject to parameter optimization before being applied, the loss function used for parameter optimization being: L=LSNOM+λr L SetReg (6)
wherein L denotes the loss function, LSNOM denotes the Bayesian personalized ranking loss corresponding to the recommendation, LSetReg denotes the regularization constraint loss, λr denotes the regularization constraint loss and LSetReg denotes the weights of the recommendations;
Bayesian personalized ranking loss is denoted as:
LSNOM =
-
∑
m
i
ϵ
M
∑
x
u
i
j
+
ϵ
X
u
i
+
∑
x
u
i
j
'
-
ϵ
X
u
i
-
ln
σ
y
^
u
i
j
-
y
^
u
i
j
'
)
+
λ
θ
∥θ∥2 (7)
wherein mi denotes the global user preference embedding for the i-th user, M denotes the sequence of global user preference embeddings,
x
u
i
j
+
denotes the j-th positive feedback interaction vector,
X
u
i
+
denotes the i-th user ui corresponding sequence of positive feedback interaction vectors,
x
u
i
j
'
-
denotes the j′ negative feedback interaction vector,
x
u
i
-
denotes the i-th user ui corresponding negative feedback interaction vector sequence, σ( ) denotes the sigmoid function,
y
^
u
i
j
denotes the recommendation score calculated from the comprehensive user preference representation and the positive feedback interaction vector in the training set,
y
^
u
i
j
'
denotes the recommendation score computed from the comprehensive user preference representation and the negative feedback interaction vector in the training set, θ denotes the model parameters, ∥θ∥2 denotes the regularization of the model parameters L2 regularization, and λθ denotes the weights” (e.g., optimization objective, picking parameters θ to minimize a loss)
“characterized in that said regularization constraint loss LSetReg is expressed as:
LSetReg=
Σ
f
z
∈
R
1
z
∑
z
∈
Z
f
z
(8)
wherein z is a set of vectors including a sequence of positive feedback interaction vectors corresponding to a single input sample, a sequence of negative feedback interaction vectors, and a vector of global user preferences, the positive feedback preference representation and the negative feedback preference representation obtained by the set operation, and the comprehensive user preference representation computed by the multi-layer perceptron mapping, Z denotes the set of z corresponding to multiple input samples, ƒ(z) denotes the single regularization constraint function corresponding to the ensemble operation, and R is the set of regularization constraint functions including the ƒ(z) as: ƒ(z)=1−DEC(UNION(z,EMPTY),EMPTY) ƒ(z)=1−DEC(UNION(z,z),z) ƒ(z)=1−DEC(INTERSECTION(z,EMPTY),EMPTY) ƒ(z)=1−DEC(INTERSECTION(z,z),z) ƒ(z)=1−DEC(DIFFERENCE(z,EMPTY),z) ƒ(z)=1−DEC(DIFFERENCE(EMPTY,z),EMPTY)
wherein EMPTY is a randomly initialized vector, UNION( ) denotes the union operation, DEC( ) denotes the similarity function, DIFFERENCE( ) denotes the difference operation, and DIFFERENCE( ) denotes the intersection operation” (e.g., penalties to male sets behave like real set algebra when combined with empty sets, set theory axioms enforced via vector-similarity penalties, linear algebra and metric geometry)
Accordingly, at Step 2A, prong one, the claim recites an abstract idea.
Step 2A Prong 2
The judicial exception is not integrated into a practical application. In particular, the claim recites the additional elements of “neural set operations model”, “set operation”, “multi-layer perceptron”, “union operation”, and “difference operation” which are recited at a high-level of generality such that they amount to no more than mere instructions to apply the exception using generic computer components (See MPEP 2106.05(f)). The Examiner notes that this is used throughout the claim limitations, and is rejected thusly for each claim which recites the same language.
Regarding the “obtaining user global preference information” limitation, the additional element is recited at a high-level of generality and amounts to extra-solution activity of obtaining data to input for a model, i.e., pre-solution activity of data gathering (e.g., obtaining information for processing in a computer system (see MPEP 2106.05(g)).
Regarding the “learning comprehensive preference representation with the neural set operations model including set operation and multi-layer perceptron, comprising performing a union operation on a time-nearest plurality of positive feedback interaction vectors drawn from a sequence of positive feedback interaction vectors followed by a difference operation with a time-nearest plurality of negative feedback interaction vectors drawn from a sequence of negative feedback interaction vectors to obtain a positive feedback preference representation”, and “after performing a union operation on a time-nearest plurality of negative feedback interaction vectors extracted from the negative feedback interaction vector sequence, performing the difference set operation on a time-nearest plurality of positive feedback interaction vectors from the positive feedback interaction vector sequence to obtain negative feedback preference representation” limitations, which are recited at a high-level of generality such that they amount to no more than mere instructions to apply the exception using generic computer components (See MPEP 2106.05(f)). In particular, these are merely inputting data into a model to learn positive and negative feedback to generate a positive and negative feedback score/representation.
Regarding the “the positive feedback preference representation, the negative feedback preference representation and the global user preference embeddings are mapped using a multi-layer perceptron to obtain the comprehensive user preference representation” limitation, which is recited at a high-level of generality such that it amounts to no more than mere instructions to apply the exception using generic computer components (See MPEP 2106.05(f)). In particular, this is merely combining the results of a model to generate a combined dataset.
Regarding the “training, by a computing system, a plurality of predictive models to identify a set of object attributes from input data, the plurality of predictive models individually being trained based at least in part on a supervised learning algorithm and a training data set including one or more examples for which a corresponding set of object features is known” limitations, which are recited at a high-level of generality such that they amount to no more than generally linking the use of abstract idea to a particular technological environment or field of use using a generic computer component (See MPEP 2106.05(h)). In particular it is merely describing how the data is labeled for use in the claimed mental process.
Regarding the “based at least in part on providing the input data to a first predictive model of the subset of the plurality of predictive models”, “based at least in part on the first predicted attribute identified, obtaining, from a second predictive model of the subset of the plurality of predictive models,”, and “based at least in part on the first predicted attribute identified, obtaining, from a third predictive model of the subset of the plurality of predictive models,” limitations, which are recited at a high-level of generality such that they amount to no more than mere instructions to apply the exception using generic computer components (See MPEP 2106.05(f)). In particular, they are merely inputting data into different models to identify desired data.
Accordingly, at Step 2A, prong two, the additional elements individually or in combination do not integrate the judicial exception into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more that the judicial exception. As discussed above, the additional element of a “neural set operations model”, “set operation”, “multi-layer perceptron”, “union operation”, and “difference operation” which are recited at a high-level of generality such that they amount to no more than mere instructions to apply the exception using a generic computer component (See MPEP 2106.05(f)).
Regarding the “obtaining user global preference information” limitation, the additional element is recited at a high-level of generality and amounts to extra-solution activity of obtaining data to input for a model, i.e., pre-solution activity of data gathering. The courts have found limitations directed to obtaining information electronically, recited at a high-level of generality, to be well-understood, routine, and conventional (see MPEP 2106.05(d)(II), “receiving or transmitting data over a network”, "electronic record keeping," and "storing and retrieving information in memory").
Regarding the “learning comprehensive preference representation with the neural set operations model including set operation and multi-layer perceptron, comprising performing a union operation on a time-nearest plurality of positive feedback interaction vectors drawn from a sequence of positive feedback interaction vectors followed by a difference operation with a time-nearest plurality of negative feedback interaction vectors drawn from a sequence of negative feedback interaction vectors to obtain a positive feedback preference representation”, and “after performing a union operation on a time-nearest plurality of negative feedback interaction vectors extracted from the negative feedback interaction vector sequence, performing the difference set operation on a time-nearest plurality of positive feedback interaction vectors from the positive feedback interaction vector sequence to obtain negative feedback preference representation” limitations, which are recited at a high-level of generality such that they amount to no more than mere instructions to apply the exception using generic computer components (See MPEP 2106.05(f)).
Regarding the “the positive feedback preference representation, the negative feedback preference representation and the global user preference embeddings are mapped using a multi-layer perceptron to obtain the comprehensive user preference representation” limitation, which is recited at a high-level of generality such that it amounts to no more than mere instructions to apply the exception using generic computer components (See MPEP 2106.05(f)).
Accordingly, at Step 2B, the additional element individually or in combination does not amount to significantly more than the judicial exception.
Regarding claim 2
Step 2A Prong 1
Claim 2 recites the following mathematical process, that in each case under the broadest reasonable interpretation, involves mathematical relationships, formulas, calculations, or algorithms implemented using generic computer components (e.g., “processor”, “ conditional generative model”, “graph neural network”) [see MPEP 2106.04(a)(2)(I)].
first, drawing the k time-nearest positive feedback interaction vectors from the sequence of positive feedback interaction vectors, and the union operation is performed on these k positive feedback interaction vectors in turn to obtain the joint embedding vector of the k positive feedback interaction vectors
h
U
i
˙
+
, denoted as,
h
U
i
˙
+
=UNION
x
u
i
j
+
,
x
u
i
j
+
∈
X
u
i
+
(1)
where UNION( ) denotes the concatenation operation, and
x
u
i
t
j
denotes the j-th positive feedback interaction vector, and
X
u
i
+
denotes the i-th user ui corresponding sequence of positive feedback interaction vectors;
wherein the equation (1) is understood as follows: for the j-th positive feedback interaction vector
x
u
i
j
+
with the j+1-th positive feedback
x
u
i
j
+
1
+
interaction vector
x
u
i
j
+
1
+
, the result of the union operation
x
u
i
j
+
∪
x
u
i
j
+
1
+
is obtained, then using the result of the union operation
x
u
i
j
+
∪
x
u
i
j
+
1
+
with the j+2th positive feedback interaction vector
x
u
i
j
+
2
+
to obtain the result of the concatenation operation
x
u
i
j
+
∪
x
u
i
j
+
1
+
∪
x
u
i
j
+
2
+
, until the k positive feedback interaction vectors are all concatenated to obtain the joint embedding vector of the k positive feedback interaction vectors
h
U
i
˙
+
” (e.g., a concatenation of the vectors, linear algebra operation )
“then, drawing he k time-nearest negative feedback interaction vectors drawn from the sequence of negative feedback interaction vectors, the joint embedding vector
h
U
i
˙
+
with the k negative feedback interaction vectors in order to perform the difference set operation to obtain the positive feedback preference representation
e
U
i
˙
+
, denoted as:
e
U
i
˙
+
=DIFFERENCE(
h
U
i
˙
+
,
x
u
i
j
-
),
x
u
i
j
-
∈
X
u
i
-
(2)
where DIFFERENCE ( ) denotes the difference set operation, and
x
u
i
j
-
denotes the first j′ negative feedback interaction vector, and
X
u
i
-
denotes the i-th user's corresponding sequence of negative feedback interaction vectors;
wherein the equation (2) is understood as follows: for the joint embedding vector
h
U
i
˙
+
with the first j′ negative feedback interaction vector
x
u
i
j
-
, and the difference set operation is performed to obtain
h
U
i
˙
+
\
x
u
i
j
-
−, and the difference operation is performed to obtain the result of the difference operation
h
U
i
˙
+
\
x
u
i
j
-
,\
x
u
i
j
'
+
1
-
with
h
U
i
˙
+
\
x
u
i
j
'
-
and the j′+1 negative feedback interaction vector
x
u
i
j
'
+
1
-
, until the k negative feedback interaction vectors are differenced to obtain a positive feedback preference representation
e
U
i
˙
+
” (e.g., a set subtraction over index sets, linear algebra of vector subtraction)
Accordingly, at Step 2A, prong one, the claim recites an abstract idea.
Step 2A Prong 2
In accordance with Step 2A, Prong 2, the claim does not include any additional elements and the judicial exception is not integrated into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Regarding claim 3
Step 2A Prong 1
Claim 3 recites the following mathematical process, that in each case under the broadest reasonable interpretation, involves mathematical relationships, formulas, calculations, or algorithms implemented using generic computer components (e.g., “processor”, “ conditional generative model”, “graph neural network”) [see MPEP 2106.04(a)(2)(I)].
“drawing the k time-nearest negative feedback interaction vectors drawn from the sequence of negative feedback interaction vectors, and the union operation is performed on these k negative feedback interaction vectors in turn to obtain the joint embedding vector of the k negative feedback interaction vectors
h
U
i
˙
-
, denoted as.
h
U
i
˙
-
=UNION
x
u
i
j
'
-
,
x
u
i
j
'
-
∈
X
u
i
-
(3)
wherein UNION( ) denotes the union operation, and
x
u
i
j
'
-
denotes the first j′ negative feedback interaction vector, and
X
u
i
-
denotes the i-th user ui corresponding sequence of negative feedback interaction vectors;
wherein the equation (3) is understood as follows: for the first j′ negative feedback interaction vector
x
u
i
j
'
-
with the first j′+1 negative feedback interaction vector
x
u
i
j
'
+
1
-
, the result of the union operation
x
u
i
j
'
-
∪
x
u
i
j
'
+
1
-
is obtained, and using the result of the union operation
x
u
i
j
'
-
∪
x
u
i
j
'
+
1
-
with the j′+2 negative feedback interaction vector
x
u
i
j
'
+
2
-
to obtain the result of the concatenation operation
x
u
i
j
'
-
∪
x
u
i
j
'
+
1
-
∪
x
u
i
j
'
+
2
-
and so on, until the k negative feedback interaction vectors are all merged to obtain the joint embedding vector of the k negative feedback interaction vectors
h
U
i
˙
-
” (e.g., vector/tensor concatenation applied iteratively, linear algebra operation )
“then, drawing he k time-nearest positive feedback interaction vectors drawn from the sequence of positive feedback interaction vectors, the joint embedding vector
h
U
i
˙
-
with the k positive feedback interaction vectors in turn to perform the difference-set operation to obtain the negative feedback preference representation
e
U
i
˙
-
, denoted as:
e
U
i
˙
-
=DIFFERENCE(
h
U
i
˙
-
,
x
u
i
j
+
),
x
u
i
j
+
∈
X
u
i
+
(4)
where DIFFERENCE ( ) denotes the difference set operation, and
x
u
i
j
+
denotes the j-th positive feedback interaction vector, and
X
u
i
+
denotes the user ui the corresponding sequence of positive feedback interaction vectors;
wherein the equation (4) is understood as follows: for the joint embedding vector
h
U
i
˙
-
with the j-th positive feedback interaction vector
x
u
i
j
+
, performing a difference operation to obtain the result of the difference operation
h
U
i
˙
-
\
x
u
i
j
+
, and the result of the difference set operation
h
U
i
˙
-
\
x
u
i
j
+
with the j+1-th positive feedback interaction vector
x
u
i
j
'
+
1
+
to obtain the result of the difference operation and so on, until the k positive feedback interaction vectors are used to obtain the negative feedback preference representation
e
U
i
˙
-
” (e.g., vector subtraction applied iteratively, linear algebra operation)
Accordingly, at Step 2A, prong one, the claim recites an abstract idea.
Step 2A Prong 2
In accordance with Step 2A, Prong 2, the claim does not include any additional elements and the judicial exception is not integrated into a practical application.
Step 2B
In accordance with Step 2B, the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
Regarding claim 8
Claim 8 recites a device comprising a memory, a processor, and a computer program. Which corresponds directly to the method steps of claim 1, respectively, with the addition of computer-executable instructions and components, which are insufficient to render the claim subject matter eligible for the same reasons described above.
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.
Claim(s) 1, 2, 3, and 8 is/are rejected under 35 U.S.C. 103 as being unpatentable over Chenzhong et al. (CN110555112B, referred to as Chenzhong), in view of Ren et al. (“Beta Embeddings for Multi-Hop Logical Reasoning in Knowledge Graphs” referred to as Ren), in view of Sarwar et al. ("Item-based Collaborative Filtering Recommendation Algorithms", referred to as Sarwar), in view of Rendle et al. (BPR: Bayesian Personalized Ranking from Implicit Feedback, referred to as Rendle), in view of He et al.(Neural Collaborative Filtering, referred to as He).
Regarding claim 1, Chenzhong teaches a computer-implemented method for recommending user preferences based on neural set operations, comprising the following steps:
extracting all positive feedback interaction items and negative feedback interaction items from user interaction history information, and constructing positive feedback interaction sequences and negative feedback interaction sequences respectively by arranging all positive feedback interaction items and negative feedback interaction items in the order of interaction time;
obtaining user global preference information(Page 2, Paragraphs 4-6, and 12, and Page 7, Paragraph 2: Describes a positive versus negative extraction and separate list. The lists store an ID sequence, supporting time-ordered sequences. The preference memory (immediate/short-term vs historical/long-term) corresponds to a user global preference information.);
encoding positive and negative feedback interaction sequences into a sequence of positive and negative feedback interaction vectors, respectively, and encoding user global preference information into a user global preference vector(Page 2, Paragraphs 4-8: Describes that the TransE and memory module are vector encodings for items/attributes and user preference vectors.)
calculating similarity between the comprehensive user preference representation and the corresponding vector of multiple candidate interaction items and the similarity is used as the recommendation score to rank the candidate interaction items, and recommending the user preference based on the ranking result (Chenzhong, Page 2,Paragraph 7 “Use the trained positive preference neural network…” describes that the model computes a score for each candidate item and sorts/ranks them, it then uses that ranked list to make recommendations.; Page 9, paragraph 6-8: Describes that the scoring function (sigmoid-normalized probability) is used for ranking candidates.);
wherein said mapping of positive feedback preference representations, negative feedback preference representations, and user global preference vectors using the multi-layer perceptron is calculated to obtain the comprehensive user preference representation, the above specific process comprises:
first connecting the positive feedback preference representation, the negative feedback preference representation, and the global user preference embeddings, and then computing the mapping of the concatenated results using a multi-layer perceptron in order to obtain the comprehensive user preference representation, denoted as: e u i =MLP(concat(e u i + ,e u i − ,m i)) (5)
wherein eu i denotes the comprehensive user preference representation, and ui denotes the i-th user, and eu i + denotes the positive feedback preference representation, and eu i − denotes the negative feedback preference representation, and mi denotes the global user preference embeddings for the ith user, and concat ( ) denotes the tandem operation, and MLP ( ) denotes the mapping computation corresponding to the multi-layer perceptron(Chenzhong, Page 7, paragraph 2, and Page 9, paragraph 2-5; Describes a vector splicing (concatenation) operator to combine learned preference embeddings and then maps the concatenated vector with an MLP to produce a complete preference representation. The system maintains positive and negative preference embeddings (streams) and long-term (historical) user preference embeddings (global vector), corresponding to three inputs being concatenated.)
Although Chenzhong teaches extracting positive and negative feedback interaction items for user interaction history information, constructing positive and negative feedback interaction sequences, obtaining user global preference information, encoding the positive and negative feedback interaction sequences into corresponding positive and negative feedback interaction vectors, encoding the user global preference information into a user global preference vector, calculating similarity between the comprehensive user preference representation and corresponding candidate item vectors for recommendation scoring, and mapping positive preference, negative preference, and global user preference embeddings through a multi-layer perception to obtain a comprehensive user preference representation. It does not fully disclose learning the comprehensive user preference representation with the neural set operations model including the union and difference operations over time-nearest pluralities of positive and negative feedback interaction vectors to obtain the positive and negative feedback preference representations.
Ren teaches, learning comprehensive preference representation with the neural set operations model including set operation and multi-layer perceptron, comprising performing a union operation on a time-nearest plurality of positive feedback interaction vectors drawn from a sequence of positive feedback interaction vectors (Ren, Page 5, Section 4.2: Describes set operations which are implemented as an MLP, a neural set operations model including set operation and multi-layer perception giving neural operations for intersection and negation with union via De Morgan implemented with MLPs, corresponding to neural probabilistic logical/set operators over embeddings, including intersection and negation, with union realized via De Morgan, and that such operator transformations are implemented with MLPs. ; Chenzhong further teaches separate positive and negative lists/sequences, with immediate (recent) vs. historical preference handling and MLP scoring. Page 2, paragraphs 9-12, page 6, Paragraph 3: Describes a positive feedback sequence and a recency split, corresponding to a selecting a time-nearest plurality from the positive sequence. It also describes a negative feedback sequence with immediate vs historical matrices, where the system then can pick the time nearest negative subset. Page 8 paragraph 6-8: Describes the set-operation pipeline which outputs new embeddings as a representation. ; Accordingly, Chenzhong supplies the time-nearest positive and negative feedback interaction-vector subsets and preference-embedding context, while Ren supplies the neural union/difference set-operator framework applied to those vector embeddings to obtain the claimed positive and negative feedback preference representations.) followed by a difference operation with a time-nearest plurality of negative feedback interaction vectors drawn from a sequence of negative feedback interaction vectors to obtain a positive feedback preference representation; after performing a union operation on a time-nearest plurality of negative feedback interaction vectors extracted from the negative feedback interaction vector sequence, performing the difference set operation on a time-nearest plurality of positive feedback interaction vectors from the positive feedback interaction vector sequence to obtain negative feedback preference representation (Ren teaches, Page 2, Introduction, “Here we propose Beta Embedding” Paragraph: Describes neural operations for intersection and negation, and that a union can be realized from a De Morgan with the intersection and negation. Page 3-4, Section 3, Computation Graph: ;Page 5, section 4.2: Describes neural intersection and negation over embeddings and states that union is realized via intersection with complement (De Morgan) this corresponds to a union and difference computation as neural set operations on vector inputs to output new embeddings. ); Chenzhong teaches separate positive and negative feedback lists which each store an ID sequence of visited items, and maintains recent and historical preference matrices for those feedbacks, updating the historical vector sequentially form the instance (recent) vector. Then concatenates the recent and historical vectors and feeds them toa multi-0layer feed-forward neural network (MLP stage). Ren teaches a neural set-operations model, defining neural probabilistic logical operators including intersection and negation, and states that union is implemented via intersection and complement (De Morgan). Each operator takes embeddings as input and transforms them into a new embedding, the transformations are implemented as MLPs. The union and difference are realized as neural set operations over vector sets. Together Chenzhong supplies the time-nearest positive/negative subsets form sequences and the MLP mapping context, while Ren supplies the neural union/difference operators to compute the claimed positive/negative feedback preference representations.)
the positive feedback preference representation, the negative feedback preference representation and the global user preference embeddings are mapped using a multi-layer perceptron to obtain the comprehensive user preference representation; (Chenzhong Page 7 Paragraph 2, and Page 3, paragraphs 1-10, and Page 9, paragraphs 2-4; Describes taking a user’s preferences embeddings and feeding them through an MLP, first concatenating the users current and historical preference vectors to form a complete preference feature vector, then inputs the vector into a multi-layer feed-forward neural network to learn the users deep preference features. The same memory module maintains historical and current positive/negative preference matrices and sends the fused vectors into the positive/negative preference neural network models for training. Ren Page 5, Section 4.2: Teaches that such set-operation outputs are mapped by the MLPs each logical/set operator :takes one or more embeddings as input and then transforms them into a new embedding,” and the transformation is implemented as a multi-layer perception (MLP) Together, Chenzhong provides the positive-preference, negative-preference and global immediate historical embeddings to be mapped, and Ren confirms the MLP mapping produces the comprehensive user-preference representation from those inputs.)
It would have been obvious to one of ordinary skill in the art at the time of the claimed invention to combine Chenzhong’s positive and negative modeling, recent versus historical preference handling, and global user preference representations with Ren’s neural set operation transformations. Doing so would enable the system to generate unified learned preference representations from recent positive, recent negative, and global preference signals for downstream recommendation scoring and ranking.
Although Chenzhong in view of Ren teaches the above, it does not teach, calculating the cosine similarity between the comprehensive user preference representation and the corresponding vector of multiple candidate interaction items respectively, and using the cosine similarity as the recommendation score.
Sarwar teaches wherein said calculating the similarity of the comprehensive user preference representation to the corresponding vector of the plurality of candidate interaction items, respectively, comprises:
calculating the cosine similarity between the comprehensive user preference representation and the corresponding vector of multiple candidate interaction items respectively, and using the cosine similarity as the recommendation score (Abstract, Section 3.1 and 3.2: Describes using cosine similarity between vectors as the recommendation score that is then used to rank/select items for recommendation. Computing cosine similarity between the representation and each item vector).
It would have been obvious to one of ordinary skill in the art at the time of the claimed invention to combine Chenzhong’s positive and negative modeling and global user preference representations, Ren’s neural set-operation transformations, with Sarwar’s similarity scoring. Doing so would enable the system to score candidate item vectors against the learned comprehensive user preference representation, creating an efficient, scale-invariant similarity metric for recommendation ranking.
Although Chenzhong, in view of Ren, in view of Sarwar teaches the above, it does not teach wherein L denotes the loss function, LSNOM denotes the Bayesian personalized ranking loss corresponding to the recommendation, LSetReg denotes the regularization constraint loss, λr denotes the regularization constraint loss and LSetReg denotes the weights of the recommendations wherein mi denotes the global user preference embedding for the i-th user, M denotes the sequence of global user preference embeddings, xu i j + denotes the j-th positive feedback interaction vector, Xu i + denotes the i-th user ui corresponding sequence of positive feedback interaction vectors, xu i j′ − denotes the j′ negative feedback interaction vector, Xu i − denotes the i-th user ui corresponding negative feedback interaction vector sequence, σ( ) denotes the sigmoid function, ŷu i j denotes the recommendation score calculated from the comprehensive user preference representation and the positive feedback interaction vector in the training set, ŷu i j′ denotes the recommendation score computed from the comprehensive user preference representation and the negative feedback interaction vector in the training set, θ denotes the model parameters, ∥θ∥2 denotes the regularization of the model parameters L2 regularization, and λθ denotes the weights.
Rendle teaches wherein said the neural set operations model is subject to parameter optimization before being applied, the loss function used for parameter optimization being: L=L SNOM+λr L SetReg (6)
wherein L denotes the loss function, LSNOM denotes the Bayesian personalized ranking loss corresponding to the recommendation, LSetReg denotes the regularization constraint loss, λr denotes the regularization constraint loss and LSetReg denotes the weights of the recommendations;
Bayesian personalized ranking loss is denoted as:
LSNOM =
-
∑
m
i
ϵ
M
∑
x
u
i
j
+
ϵ
X
u
i
+
∑
x
u
i
j
'
-
ϵ
X
u
i
-
ln
σ
y
^
u
i
j
-
y
^
u
i
j
'
)
+
λ
θ
∥θ∥2 (7)
wherein mi denotes the global user preference embedding for the i-th user, M denotes the sequence of global user preference embeddings, xu i j + denotes the j-th positive feedback interaction vector, Xu i + denotes the i-th user ui corresponding sequence of positive feedback interaction vectors, xu i j′ − denotes the j′ negative feedback interaction vector, Xu i − denotes the i-th user ui corresponding negative feedback interaction vector sequence, σ( ) denotes the sigmoid function, ŷu i j denotes the recommendation score calculated from the comprehensive user preference representation and the positive feedback interaction vector in the training set, ŷu i j′ denotes the recommendation score computed from the comprehensive user preference representation and the negative feedback interaction vector in the training set, θ denotes the model parameters, ∥θ∥2 denotes the regularization of the model parameters L2 regularization, and λθ denotes the weights (Section 4.1-4.3 : Teaches a Bayesian personalized ranking optimization criterion for personalized ranking from implicit feedback, including a log-sigmoid pairwise ranking loss with model-parameter regularization and further teaches optimizing that objective using stochastic gradient descent. (Chenzhong Page 8-9, Step5: Teaches training the recommender model parameters, including weight matrices, bias vectors, and corresponding user and scenic-spot feature vectors, using a loss-based optimization procedure during training, specifically by back-propagation with a logarithmic cross-entropy loss function and stochastic gradient descent.)).
It would have been obvious to one of ordinary skill in the art at the time of the claimed invention to combine Chenzhong’s neural network training and parameter optimization, with Rendle’s Bayesian personalized ranking loss and regularization framework. Doing so would have enabled the system to optimize its learned preference representations directly for personalized ranking from implicit feedback using a known regularized training objective.
Although Chenzhong in view of Ren, in view of Sarwar, in view of Rendle teaches the above, they do not teach, wherein said regularization constraint loss LSetReg is expressed as:
LSetReg =
Σ
f
2
∈
R
1
z
Σ
z
ϵ
Z
f
z
(8)
wherein z is a set of vectors including a sequence of positive feedback interaction vectors corresponding to a single input sample, a sequence of negative feedback interaction vectors, and a vector of global user preferences, the positive feedback preference representation and the negative feedback preference representation obtained by the set operation, and the comprehensive user preference representation computed by the multi-layer perceptron mapping, Z denotes the set of z corresponding to multiple input samples, ƒ(z) denotes the single regularization constraint function corresponding to the ensemble operation, and R is the set of regularization constraint functions including the ƒ(z) as:
ƒ(z)=1−DEC(UNION(z,EMPTY),EMPTY)ƒ(z)=1−DEC(UNION(z,z),z)ƒ(z)=1−DEC(INTERSECTION(z,EMPTY),EMPTY)ƒ(z)=1−DEC(INTERSECTION(z,z),z)ƒ(z)=1−DEC(DIFFERENCE(z,EMPTY),z)ƒ(z)=1−DEC(DIFFERENCE(EMPTY,z),EMPTY)
wherein EMPTY is a randomly initialized vector, UNION( ) denotes the union operation, DEC( ) denotes the similarity function, DIFFERENCE( ) denotes the difference operation, and DIFFERENCE( ) denotes the intersection operation.
He teaches, neural scoring and MLP-based mapping of embeddings to downstream representations/scores, corresponding to the claimed decoder/similarity-function context, while Chenzhong and Ren provide the claimed positive/negative interaction-vector and set-operation context, and Rendle provides regularized ranking optimization.
wherein said regularization constraint loss LSetReg is expressed as:
LSetReg =
Σ
f
2
∈
R
1
z
Σ
z
ϵ
Z
f
z
(8)
wherein z is a set of vectors including a sequence of positive feedback interaction vectors corresponding to a single input sample, a sequence of negative feedback interaction vectors, and a vector of global user preferences, the positive feedback preference representation and the negative feedback preference representation obtained by the set operation, and the comprehensive user preference representation computed by the multi-layer perceptron mapping, Z denotes the set of z corresponding to multiple input samples, ƒ(z) denotes the single regularization constraint function corresponding to the ensemble operation, and R is the set of regularization constraint functions including the ƒ(z) as: (He et al. Neural Collaborative Filtering, Page 3, section 3.1 describes user/item embeddings fed through an MLP to produce prediction scores, corresponding to standard user embedding plus an MLP that maps concatenated inputs to a downstream representation/score.; Chenzhong Page 5, paragraph 3: Describes per-user positive and negative feedback lists corresponding to positive/negative feedback interaction vectors. ;Ren Section 4.2: Describes neural set-operations toolkit (projection, intersection negation and union) over embeddings corresponding to set
ƒ(z)=1−DEC(UNION(z,EMPTY),EMPTY)ƒ(z)=1−DEC(UNION(z,z),z)ƒ(z)=1−DEC(INTERSECTION(z,EMPTY),EMPTY)ƒ(z)=1−DEC(INTERSECTION(z,z),z)ƒ(z)=1−DEC(DIFFERENCE(z,EMPTY),z)ƒ(z)=1−DEC(DIFFERENCE(EMPTY,z),EMPTY)
wherein EMPTY is a randomly initialized vector, UNION( ) denotes the union operation, DEC( ) denotes the similarity function, DIFFERENCE( ) denotes the difference operation, and DIFFERENCE( ) denotes the intersection operation (He, Section 3.1:Describes neural scoring, a similarity/compatibility function between learned representations, and where a recommender takes two learned embeddings and maps them to a prediction score via a neural decoder, similarity/compatibility function. Corresponding to a decoder/decision function instantiation for DEC.; Ren, Section 4.2: Describes logical/set operators over embeddings covering named set operations for regularizes.).
It would have been obvious to one of ordinary skill in the art at the time of the claimed invention to combine Chenzhong’s positive/negative feedback interaction vector modeling, Ren’s neural set-operation framework over embeddings, Rendle’s regularized ranking optimization, with He’s neural decoder/scoring architecture. Doing so would have enabled the system to apply regularized neural scoring to learned set-operation outputs and embedding representations used in personalized recommendation, thereby improving training stability and supporting consistent optimization of the recommender model.
Regarding claim 2, Chenzhong in view of Ren teaches, the computer-implemented method for recommending user preferences based on neural set operations according to claim 1, characterized in that after performing a union operation on the time-nearest multiple positive feedback interaction vectors drawn from the sequence of positive feedback interaction vectors, performing a difference operation with the time-nearest multiple negative feedback interaction vectors drawn from the sequence of negative feedback interaction vectors to obtain a positive feedback preference representation; the specific process of performing the difference operation comprises:
first, drawing the k time-nearest positive feedback interaction vectors from the sequence of positive feedback interaction vectors, and the union operation is performed on these k positive feedback interaction vectors in turn to obtain the joint embedding vector of the k positive feedback interaction vectors hu i +, denoted as, h u i +=UNION(x u i j +),x u i j + ∈X u i + (1)
where UNION( ) denotes the concatenation operation, and xu i j + denotes the jth positive feedback interaction vector, and Xu i + denotes the ith user ui corresponding sequence of positive feedback interaction vectors;
wherein the equation (1) is understood as follows: for the j-th positive feedback interaction vector xu i j + with the j+1-th positive feedback interaction vector xu i (j+1) +, the result of the union operation xu i j +∪xu i (j+1) + is obtained, then using the result of the union operation xu i j +∪xu i (j+1) + with the j+2th positive feedback interaction vector xu i (j+2) + to obtain the result of the concatenation operation xu i j +∪xu i (j+1) +∪xu i (j+2) +, until the k positive feedback interaction vectors are all concatenated to obtain the joint embedding vector of the k positive feedback interaction vectors hu i +; (Chenzhong, Page 2, paragraph 12, Describes positive and negative lists with IQ sequences.; Page 8, paragraph 4-8: Describes positive sequences, recency (current vs historical) and sequential updates. Then concatenates (splicing) those together to form a complete positive vector. Together these disclose positive sequences, recency construct (current vs historical with sequential updates) and vector concatenation. This produces a union over the time-nearest positive vectors to produce joint embeddings.)
then, drawing the k time-nearest negative feedback interaction vectors drawn from the sequence of negative feedback interaction vectors, the joint embedding vector hu i + with the k negative feedback interaction vectors in order to perform the difference set operation to obtain the positive feedback preference representation eu i +, denoted as: e u i +=DIFFERENCE(h u i + ,x u i j′ −),x u i j′ − ∈X u i − (2)
where DIFFERENCE ( ) denotes the difference set operation, and xu i j′ − denotes the first j′ negative feedback interaction vector, and Xu i − denotes the i-th user's corresponding sequence of negative feedback interaction vectors;
wherein the equation (2) is understood as follows: for the joint embedding vector hu i + with the first j′ negative feedback interaction vector xu i j′ −, and the difference set operation is performed to obtain hu i +\xu i j′ −, and the difference operation is performed to obtain the result of the difference operation hu i +\xu i j′ −\xu i (j′+1) − with hu i +\xu i j′ − and the j′+1 negative feedback interaction vector xu i (j′+1) −, until the k negative feedback interaction vectors are differenced to obtain a positive feedback preference representation eu i +.(Chenzhong , Page 2, paragraph 12, Page 8, paragraph 4-8: teaches which vectors to use (the k time-nearest negatives from the negative sequence). Ren Page 5, section 4.2: teaches how to compute difference between embeddings and how each operation outputs a new embedding to iteratively apply it across the k negatives.)
Regarding claim 3, Chenzhong in view of Ren teaches, the computer-implemented method for recommending user preferences based on neural set operations according to claim 1, characterized in that, the time-nearest multiple negative feedback interaction vectors drawn from the sequence of negative feedback interaction vectors are merged and then differenced with the time-nearest multiple positive feedback interaction vectors drawn from the sequence of positive feedback interaction vectors to obtain the negative feedback preference representation; the specific process comprises:
drawing the k time-nearest negative feedback interaction vectors drawn from the sequence of negative feedback interaction vectors, and the union operation is performed on these k negative feedback interaction vectors in turn to obtain the joint embedding vector of the k negative feedback interaction vectors hu i −, denoted as. h u i −=UNION(x u i j′ −)x u i j′ − ∈x u i − (3)
wherein UNION( ) denotes the union operation, and xu i j′ − denotes the first j′ negative feedback interaction vector, and Xu i − denotes the i-th user ui corresponding sequence of negative feedback interaction vectors; (Chenzhong, Page 2, paragraph 10-12, and page 8, paragraph 6-8: Describes positive and negative lists with IQ sequences.; Page 8, paragraph 4-8: Describes negative sequences, recency split(current vs historical) basis to select k time-nearest negatives and a disclosed vector splicing operator to merge multiple vectors into a joint embedding.)
wherein the equation (3) is understood as follows: for the first j′ negative feedback interaction vector xu i j′ − with the first j′+1 negative feedback interaction vector xu i (j′+1) −, the result of the union operation xu i j′ −∪xu i (j′+1) − is obtained, and using the result of the union operation xu i j′ −∪xu i (j′+1) − with the j′+2 negative feedback interaction vector xu i (j′+2) − to obtain the result of the concatenation operation xu i j′ −∪xu i (j′+1) −∪xu i (j′+2) − and so on, until the k negative feedback interaction vectors are all merged to obtain the joint embedding vector of the k negative feedback interaction vectors hu i −;
then, drawing the k time-nearest positive feedback interaction vectors drawn from the sequence of positive feedback interaction vectors, the joint embedding vector hu i − with the k positive feedback interaction vectors in turn to perform a difference-set operation to obtain a negative feedback preference representation eu i −, denoted as: e u i −=DIFFERENCE(h u i − ,x u i j +),x u i j + ∈X u i + (4)
wherein DIFFERENCE ( ) denotes the difference set operation, and xu i j + denotes the j-th positive feedback interaction vector, and Xu i + denotes the user ui the corresponding sequence of positive feedback interaction vectors;
wherein the equation (4) is understood as follows: for the joint embedding vector hu i − with the j-th positive feedback interaction vector xu i j +, performing a difference operation to obtain the result of the difference operation hu i −\xu i j +, and the result of the difference set operation hu i −\xu i j + with the j+1-th positive feedback interaction vector xu i (j+1) + to obtain the result of the difference operation and so on, until the k positive feedback interaction vectors are used to obtain the negative feedback preference representation eu i −.( Chenzhong , Page 2, paragraph 9-12, Page 8, paragraph 4-8: teaches which vectors to use (the k time-nearest negatives from the positive sequence). Ren page 5, section 4.2: teaches how to compute difference between embeddings and how each operation outputs a new embedding to iteratively apply it across the k positives.)
Regarding claim 8, which recites substantially the same limitations as claim 1 and further recites a device… comprising a non-transitory memory, a processor and a computer program stored in said memory and executable on said processor (Chenzhong describes use of their inventive concept to be used in computer technology, and Ren’s methods are implemented via computers, it is understood that recitation of computer components to implement method steps is common and necessary in the art). to perform the steps of claim 1. It is therefore rejected on the same premise.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). 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 DONALD T RODEN whose telephone number is (571)272-6441. The examiner can normally be reached Mon-Thur 8:00-5:00 EST.
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, Omar Fernandez Rivas can be reached at (571) 272-2589. 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.
/D.T.R./Examiner, Art Unit 2128
/OMAR F FERNANDEZ RIVAS/Supervisory Patent Examiner, Art Unit 2128
Prosecution Insights