DETAILED ACTION
This action is responsive to the communications filed on 02/18/2026. Claim 1-3, 5, 7, 9-13, 15, 17, and 19-20 are pending for examination.
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 02/18/2026 has been entered.
Response to Arguments
Applicant’s arguments with respect to the 35 U.S.C. 103 rejection of claims 1-3, 5, 7, 9-13, 15, 17, and 19-20 have been considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or non-obviousness.
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.
Claims 1-3, 7, 11-13 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Ren et al. ((2024, May). Representation learning with large language models for recommendation. In Proceedings of the ACM on Web Conference 2024 (pp. 3464-3475).), hereafter referred to as Ren in view of Xu et al., (Xu, S., Yoon, H. J., & Tourassi, G. (2014). A user-oriented web crawler for selectively acquiring online content in e-health research. Bioinformatics, 30(1), 104-114.), hereafter referred to as Xu, and in further view of Huang et al. (Huang, Y., Wu, X., Hu, W., Feng, J., & Deng, C. (2022, December). State-Aware Adversarial Training for Utterance-Level Dialogue Generation. In Proceedings of the Towards Semi-Supervised and Reinforced Task-Oriented Dialog Systems (SereTOD) (pp. 62-74).), hereafter referred to as Huang, Liao et al. ((2020). Topic-guided conversational recommender in multiple domains. IEEE Transactions on Knowledge and Data Engineering, 34(5), 2485-2496.), hereafter referred to as Liao, and Yu et al., (Zhang, R., Yu, T., Shen, Y., Jin, H., & Chen, C. (2019). Text-based interactive recommendation via constraint-augmented reinforcement learning. Advances in neural information processing systems, 32.), hereafter referred to as Yu.
Claim 1: Ren teaches the following:
An apparatus for determining an excitation element, wherein the apparatus comprises: at least a processor; and a memory communicatively connected to the at least a processor, wherein the memory contains instructions configuring the at least a processor to: (Ren, page 3471, col. 2, paragraph 1, “In Table 4, we present the epoch time of training on a server with an Intel Xeon Silver 4314 CPU and an NVIDIA RTX 3090 GPU.”, training the model comprises using CPU’s and GPU’s, indicating that a processor and memory with instructions are being used)
using a representation generator, generate a representation data structure based on the system data; (Ren, page 3466, col. 2, paragraph 2, “To mitigate the impact of irrelevant signals on the representation, it is necessary to incorporate auxiliary informative cues. One approach is to introduce textual information, e.g., user and item profiles, which provide insights for user preference learning. These profiles can be encoded using language models to generate representations s ∈ R 𝑑𝑠 that effectively capture the semantic aspects of user preferences. Importantly, both s and e capture shared information that is relevant to the aspects associated with user-item interactions.”, Ren reaches a system for capturing semantic information efficiently through user and item profiles which is supposed to increase the interpretability of recommendation machine learning models. Paragraph 32 of the application states that, “In a non-limiting example, representation data structure 132 may include one or more fields and/or variables which may be modified by a system and/or user device 124, and an evaluation metric may include entries into such fields and/or values assigned to such variables.”, since the user and item profiles represent variables of users and items, it is considered a representation data structure which may include one or more fields and/or variables. When combined with Xu above, representation data structures from Ren are derived from the system data obtained through a user-interface trained web crawler of Xu.)
determine a plurality of evaluation metrics as a function of the representation data structure; (Ren, page 3467, section 3.2, paragraph 1, “User profile: should effectively encapsulate the particular types of items that users are inclined to favor, allowing for a comprehensive representation of their personalized tastes and preferences.”, an example user and item profile is shown in figure 3. As disclosed earlier by referencing paragraph 32, “and an evaluation metric may include entries into such fields and/or values assigned to such variables.”, since the profiles have values set to variables of users and items it is considered an evaluation metric.)
generate an augmented plurality of evaluation metrics by interpolating, into the plurality of evaluation metrics, an additional evaluation metric; (Ren, page 3467, col. 2, paragraph 1, “Item profile: It should eloquently articulate the specific types of users that the item is apt to attract, providing a clear representation of the item’s characteristics and qualities that align with the preferences and interests of those users”, additional to the user profile which comprise variables to represent the user, an item profile is generated and is interpreted as an additional evaluation metric to create the augmented plurality of evaluation metrics)
Xu, in the same field of web-crawling, teaches the following limitation which the above fails to teach:
receive system data through a system data source obtained using a web crawler trained with information received through a user interface (Xu, page 106, section 4, “Given a web page utility estimator ðwp, Þ trained from a set of human-labeled example web pages f^ ðwp, Þg, we can then develop a user-oriented web crawler that is capable of adaptively acquiring relevant web pages that satisfy the user information requirement.”, human labels and user provided queries are interpreted as the UI-supplied information used to train the web crawler’s utility estimator, which then obtains the system data via crawling.
Xu, page 109, section 4.3, paragraph 1, “Given a specific user web crawling need , the user first conducts a few brief web query efforts wherein he/she would compose a few queries to search the web.”, discloses using user-interfaced webpages to train the crawler.)
Ren teaches generating rich user/item representations from textual profiles (LLM-empowered representation learning) that downstream recommenders can use for accuracy and interpretability. Xu teaches a user-oriented, feedback-trained web crawler that acquires system data (relevant web pages/content) that a representation model like Ren’s would consume. It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to combine Ren’s representation generator with Xu’s UI-trained focused crawling so that the system dynamically harvests user-relevant online content (system data) and then encodes it into representations per Ren. A motivation of which would be thereby improving relevance of inputs adapted to the user. (Xu, page 107, section 4.2, “Hence in practice, given a priority list Δu and a certain amount of downloading time that a user can afford, the crawler will only download the header part of the prioritized results until all available time is used up”, input gained from the web crawler is prioritized based on the user preferences)
Huang, in the same field of machine learning generative input, teaches the following limitations which Ren and Xu fails to teach:
PNG
media_image1.png
250
302
media_image1.png
Greyscale
Figure 1 of Huang
wherein the user interface comprises a chatbot interface configured to prompt a user for information used to collect the system data; (Huang, page 62, col. 1, section 1, paragraph 1, “Task-oriented dialogue systems (Young et al., 2013; Williams et al., 2016; Wu et al., 2020; Su et al., 2021) are designed to assist user in completing daily tasks, which involve reasoning over multiple dialogue turns.”, Huang teaches a multi-turn dialogue interface (i.e., a conversational agent) in which the system prompts the user with a question to narrow a search and the user provides the requested information in response.)
wherein generating the augmented plurality of evaluation metrics comprises determining the additional evaluation metric as a function of an evaluation metric of the plurality of evaluation metrics using a generative machine learning model comprising a generative adversarial network further comprising a discriminator, (Huang, page 64, section 2.2, paragraph 1, “To generate more human-like user utterances, we propose using adversarial training for generation: the generator is guided by the discriminator to produce utterances that are indistinguishable from the original dialogues and consistent with the belief state condition. The discriminator is trained on the dataset consisting of the utterances of original dialogues and the utterances generated by the generator. The learning of generator and discriminator is conducted in an alternate manner”, the system in Huang generates values for missing fields or variables (the evaluation metrics) by using a GAN (generative adversarial network). The generator produces utterances (dialogue/questions) that provide these values, while the discriminator evaluates whether generated utterances align with realistic user utterances. )
wherein the generative machine learning model is iteratively trained by the at least a processor based on received user responses through a feedback loop, (Huang, page 64, section 2.2, paragraph 1, “the generator is guided by the discriminator to produce utterances that are indistinguishable from the original dialogues and consistent with the belief state condition. The discriminator is trained on the dataset consisting of the utterances of original dialogues and the utterances generated by the generator. The learning of generator and discriminator is conducted in an alternate manner, which is detailed in Algorithm 1. ”, this shows how the generator and discriminator work in a feedback loop, wherein the generator produces utterances that guide users to provide value for missing fields, and the discriminator evaluates them to refine future generation. As shown in algorithm 1, this process repeats over multiple training cycles/iterations. The system continually improves its ability to generate dialogue that elicits the correct variable values from users.
Huang, page 64, col. 1 paragraph 1, “The task we focus on is to generate a user utterance Ut conditioned on the turn-level dialogue state St and corresponding system response Rt .”, the generation of input is trained on turn-level dialogue. A turn-level dialogue state represents the current state of the conversation at a specific turn, including all previously-known information provided by the user. The examiner interprets this as received user responses through a feedback since it generates a user utterance after each user response.)
wherein the discriminator is configured to evaluate generated content of the generative machine learning model; (Huang, page 64, col. 1, section 2.2, paragraph 3, “The discriminator D is a binary classifier that aims to determine whether the user utterance is generated or from the original dataset. In order to make sense of belief state condition, the concatenation of turn-level belief state and user utterance is used as input to the discriminator.”, the discriminator is used to evaluate generated utterances from the generative model.)
Ren teaches generating representations from textual information (e.g., user and item profiles) for recommendation models. Xu teaches a user-oriented crawling pipeline in which the method trains a model and launches an adaptive web crawler that executes queries to obtain a collection of search result web pages. Huang teaches a multi-turn dialogue system interaction in which the system prompts the user with a question to narrow a search and the user responds with requested information. It would have been obvious to a person of ordinary skill in the art before the effective filing date of the invention to have incorporated the teachings disclosed by Ren (i.e. creating enriched semantics for recommendation systems) and Xu with the teachings disclosed by Huang (i.e. state-based GAN input generation). In particular, Huang’s dialogue-system prompting provides the chatbot-style mechanism for soliciting the user information, while Xu uses the user-provided information in the crawling/training workflow to obtain the collection of web pages, thereby meeting teaching a “chatbot interface … configured to prompt a user for information used to collect the system data” in combination. A motivation for the combination is to enable a system to dynamically generate targeted utterances (questions) that guide users in providing necessary input variables, ensuring a more context-driven focus, (Huang, page 63, col. 1 paragraph 1, “The experimental results show that our proposed method has higher controllability for state-aware dialogue even though it has higher or comparable naturalness to existing methods, and improves the discriminability of generation. Furthermore, we investigate the effectiveness of our approach via downstream dialogue state tracking (DST) tasks. Experimental results demonstrate the high-quality data generated by our proposed framework improves the performance over state-of-the-art models.”, the state-aware adversarial training of Huang can improve controllability, dialogue state tracking, and response quality compared to other known methods.).
Liao, in the same field of system machine learning for optimizing recommendations, teaches the following limitations which Ren, Xu, and Huang fails to teach:
determine an excitation element by: training an excitation element machine learning model on a training dataset including a plurality of example evaluation metrics as inputs correlated to a plurality of example excitation elements as outputs; (Liao, abstract, “In this paper, we thus present a Topic-guided Conversational Recommender (TCR) which is specifically designed for the multi-domain setting. It augments the sequence-to-sequence (seq2seq) models with a neural latent topic component to better guide the response generation. To better leverage the dialogue history and the back-end data structure, we adopt a graph convolutional network (GCN) to model the relationships between different recommendation candidates while also capture the match between candidates and the dialogue history”, the semantic representation output of Ren can be used with the recommendation system of Liao to determine a recommendation or optimization to a system. Specification 111 of this application states that, “As used herein, an excitation element is a data structure describing a change to be implemented to improve a system, a process performed by a system, or both”, it is interpreted by the examiner that the recommendation machine learning model of Liao is synonymous to the excitation element machine learning model of this application.
Page 2490, section 3.4, paragraph 1, “As the generation of responses is controlled via the sentinel token $ as a hard gate, the generation procedure actually works in a two-step way. The substitution of $ with venue recommendation result is separate from the token generation process. In order to achieve good results, we train the whole model in a sequential way. At the beginning, we train the global topic control component separately on the altered dataset where all venue names are replaced with $. The training objective of this component is LTopic detailed as Eq. (8). Then we change back the dataset and train the GCN component for venue ranking on it. The dialogue context is embedded via the trained global topic control model. The training objective is LGCN as detailed in Eq. (10). Finally, we initialize the whole model with the components trained and fine-tune them altogether. The final training objective is as follows:”, this disclosure explains the training process of a conversational recommendation machine learning model. Natural language input is used and the examiner interprets this as being compatible with the user/item profile generation of Ren.)
and generating an excitation element as a function of the augmented plurality of evaluation metrics using the trained excitation element machine learning model; (Liao, page 2485, foot note 1, “The recommendation candidates are like venue names or train numbers in dataset. For ease of illustration, we all describe as venues throughout the paper”, for ease of illustration the paper uses “venue” as a word to convey multiple data types.
Page 2490, col. 1, paragraph 2, “In branch 2, the h is fed to the GCN-based recommender. Following the process introduced in Section 3.2,
the recommender ranks the venues and outputs the top ranked venue name”, the system ranks the recommendations outputted from the model and returns the best one)
and display the excitation element to a user through a user interface at a display device. (Liao, page 2486, figure 1, “A sample dialogue between a user (U) and an agent (A) from the dataset. We observe the need for global topic control and accurate recommendation”, as shown in the example conversation in figure 1, the system displays the excitation element (the recommendation) to the user.)
Liao teaches a recommendation system machine learning model in the form of a natural language conversation. It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have incorporated the teachings disclosed by Ren, Xu, and Huang with the teachings disclosed by Liao (i.e. a natural language recommendation system). A motivation for the combination is to have a recommendation system that can provide more information about users, items, or intentions in an intuitive way through natural language multi-round conversation, (Liao, page 2487, col. 1, section 2.2, paragraph 1, “For recommender systems, conversational systems can provide more information about user intentions, such as user preferred type of food or the location of a hotel, by interactively soliciting and identifying user intentions based on multi-round natural language conversation.”).
Yu, in the same field of system machine learning for optimizing recommendations, teaches the following limitations which Ren, Liao, Huang and Xu fails to teach:
And determine an error signal describing a quality of a state of a system after display of the excitation element, (Yu, page 2, section 2.2, “We employ an RL-based formulation for sequential recommendation of items to users, utilizing user feedback in natural language. Denote st ∈ S as the state of the recommendation environment at time t and at ∈ Aasthe recommender-defined items from the candidate items set A. In the context of a recommendation system, as discussed further below, the state st corresponds to the state of sequential recommender, implemented via a LSTM [23] state tracker. At time t, the system recommends item at based on the current state st at time t. After viewing item at, a user may comment on the recommendation in natural language (a sequence of natural-language text) xt, as feedback. The recommender then receives a reward rt and perceives the new state st+1. Accordingly, we can model the recommendation-feedback loop as an MDP M = S,A,P,R, where P : S × A ×S → R is the environment dynamic of recommendation and R : S × A → R is the reward function used to evaluate recommended items. The recommender seeks to learn a policy parameterized by θ, i.e., πθ(a|s), that corresponds to the distribution of items conditioned on the current state of the recommender.”, Yu’s recommended item/action teaches an output selected by a recommender policy and presented to the user. Yu further teaches that the user’s reaction to the displayed recommendation is used to determine a reward and next recommender state. The reward/penalty information in Yu is an “error signal” or quality signal because it indicates whether the displayed recommendation satisfied, violated, or improved the user’s stated preferences and the recommender state.)
wherein the error signal is used as a cost function (Yu, page 2, section 2.1, paragraph 1, “The goal of an agent is to learn an optimal policy that maximizes JR(π). A constrained Markov decision process (CMDP) [3] extends the MDP framework by introducing the constraint C(s,a) (mapping astate-action pair to costs, similar to the usual reward) 2 and a threshold α ∈ [0,1].”, Yu’s constraint function/discriminator evaluates whether a recommended action violates user preference. Yu uses this constraint as a cost/penalty term in the reinforcement-learning objective. Thus, Yu’s feedback-derived violation signal corresponds to the claimed “error signal,” and Yu’s constraint/cost function corresponds to the claimed use of that signal “as a cost function.”)
PNG
media_image2.png
260
305
media_image2.png
Greyscale
Algorithm 1 of Yu
to retrain at least one of the representation generator and the excitation element machine learning model using a reinforcement learning algorithm, (Yu, page 4, last paragraph, “With the visual feature cvis t and textual feature ctxt t ,the recommender perceives the state in an auto-regressive manner. At time t, the state is st = f(g([cvis t ,ctxt t ]),st−1), where g is an MLPfor textual and visual matching, and f is the LSTM unit [23]. Since our goal in each user session is to find items with a set of desired attribute values, we use the policy πθ with multi-discrete action spaces [22, 12]. For each attribute, the desired attribute value by the user is sampled from a categorical distribution. Given the state st, the probability of choosing a particular attribute value is output by a three-layer fully connected neural network with a softmax activation function. The recommender samples the values of different attributes from πθ. If K items are recommended at each time, we select the items that are top K closest to the sampled attribute values under Euclidean distance in the visual attribute space.”, Yu’s recommender policy corresponds to the claimed excitation element machine learning model because it selects the recommendation/action that is provided to the user. Yu retrains or updates the recommender policy parameters using reinforcement-learning updates based on the observed reward and penalty signals. The claim only requires retraining “at least one of” the representation generator or excitation element ML model; therefore, Yu’s retraining of the recommender/excitation-element model is sufficient.)
such that future excitation elements are modified to improve the state of the system relative to a prior state. (Yu, page 3, figure 1, “Overview of the reward constrained recommender model. When receiving the recommended images, the user gives natural-language feedback, and this feedback will be used for the next item recommendation, as well as preventing future violations.”,
Yu, page 5, paragraph 1, “Following Algorithm 1, we update the discriminator after each user session, where a user interacts with the system for several time steps, or quits. To further enhance the results, when making recommendations, we reject some items based on this discriminator. If an item at sampled by the recommender has high probability of violating the previous comments {xi}t−1 i=1, we ignore this item and sample another item to recommend.”, Yu’s reinforcement-learning policy update changes future recommendation actions based on prior feedback, rewards, and penalties. Therefore, future recommended items—corresponding to future excitation elements—are modified to avoid prior preference violations and improve the recommender state relative to earlier interactions. This teaches or at least suggests the claimed modification of future excitation elements to improve the system state relative to a prior state.)
It would have been obvious to a person of ordinary skill in the art before the effective filing date to modify the combined Ren/Liao/Huang/Xu system to include Yu’s reinforcement-learning feedback update. Ren teaches generating semantic representation data for recommender systems; Liao teaches generating and displaying conversational recommendations; Xu teaches collecting user-oriented system data using a trained web crawler; and Huang teaches adversarial/generative machine-learning techniques including discriminator-based evaluation. Yu further teaches using post-recommendation user feedback, rewards, costs, and penalties to update the recommender policy so future recommendations are improved. A person of ordinary skill would have been motivated to incorporate Yu’s reinforcement-learning update into the Ren/Liao/Huang/Xu recommendation pipeline to adapt the recommendation model based on user feedback after display, reduce future preference violations, and improve subsequent system outputs.
Claim 2: Ren, Xu, Huang, Liao, and Yu teaches the limitations of claim 1. Liao further teaches:
wherein determining the plurality of evaluation metrics as a function of the representation data structure comprises: using the user interface, displaying the representation data structure to the user and using the user interface, receiving from a user the plurality of evaluation metrics (Liao, page 2486, figure 1,
PNG
media_image3.png
389
442
media_image3.png
Greyscale
This is an example user interface (UI) as shown in figure 1 of Liao. As shown, the UI displays the representation data structure like “travel day” and then receives evaluation metrics such as the value “Wednesday” from the user.)
The rationale to combine Ren with Liao is similar to that applied for claim 1 above.
Claim 3: Ren, Xu, Huang, Liao, and Yu teaches the limitations of claim 1. Ren further teaches:
wherein generating the augmented plurality of evaluation metrics comprises determining the additional evaluation metric as a function of an evaluation metric of the plurality of evaluation metrics using a large language model (LLM). (Ren, page 3466, col.2, paragraph 1, “Text-enhanced User Preference Learning. To mitigate the impact of irrelevant signals on the representation, it is necessary to incorporate auxiliary informative cues. One approach is to introduce textual information, e.g., user and item profiles, which provide insights for user preference learning. These profiles can be encoded using language models to generate representations s ∈ R 𝑑𝑠 that effectively capture the semantic aspects of user preferences.”)
Claim 7: Ren, Xu, Huang, Liao, and Yu teaches the limitations of claim 1. Ren further teaches:
training a representation machine learning model on a training dataset including a plurality of example elements of system data as inputs correlated to a plurality of example representation data structures as outputs; (Ren, page 3473, section A.1, “The focus is on the training procedure of these implementations. Prior to training, user and item profiles are preprocessed, and their semantic embeddings s are generated using text models. Algorithm 1 presents the training procedure for RLMRecCon, while Algorithm 2 outlines the process for RLMRec-Gen.”, Ren utilizes a machine learning model to generate representation data structures from users or items.)
and generating the representation data structure as a function of the system data using the trained representation machine learning model. (Ren page 3469, section 4.1.4 paragraph 1, “We use the text-embedding-ada-002 [22] to generate semantic representations s. During training, all methods are trained with a fixed batch size of 4096 and a learning rate of 1e-3 using the Adam optimizer.”)
Claims 11-13 recite limitations substantially similar to claims 1-3 respectively, and as such a similar analysis applies.
Claim 17 recites limitations substantially similar to claim 7 as such a similar analysis applies.
Claims 5 and 15 are rejected under 35 U.S.C. 103 as being unpatentable over Ren, Xu, Huang, Liao, and Yu as applied to claims 1-3, 7, and 11-13 above, and further in view of Dahan et al. (US 20140304106 A1), referred to as Dahan.
Claim 5: Ren, Xu, Huang, Liao, and Yu teaches the limitations of claim 1. Dahan, in the same field of recommendation machine learning model implementation, teaches the following limitation which Ren, Xu, Huang, Liao, and Yu fail to teach:
wherein generating the augmented plurality of evaluation metrics comprises: selecting an aggregate evaluation metric using a classifier and determining the additional evaluation metric as a function of the selected aggregate evaluation metric. (Dahan, paragraph 81, “At state 202, population data in input by the recommendation system from one or more sources. The population data may include unit sales, average prices, and choice calibration data used to calibrate population level purchase data. The population data may derive from observable marketplace data, and the calibration data may be collected through surveys of individuals making choices.”, the recommendation machine learning model of Dahan utilizes aggregate evaluation metrics of average price, as well as other additional metrics as input. Specification paragraph 106 of this application states, “As used herein, an aggregate evaluation metric is a data structure describing a feature typical of a system of a group of systems, a data structure describing a feature derived from multiple systems of a group of systems, or both. In a non-limiting example, an aggregate evaluation metric may include an average height of a population”, showing that an aggregate evaluation includes an average.)
Dahan teaches using aggregate evaluation metrics which Ren and Liao fail to teach. It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have incorporated the teachings disclosed by Ren, Xu, Huang, Liao, and Yu with the teachings disclosed by Dahan (i.e. incorporate aggregate evaluation metrics). A motivation for the combination is to use aggregate information with mathematical algorithms in order to further determine useful information for a recommendation model, (Dahan, paragraph 81, “For example, if the item type is televisions, the population data for the population of television buyers (e.g., of many or all different models of televisions) may include the number of unit sales over a specific period of time, for a specific geographical region, and the average or median sales price. Optionally, the process updates the population data periodically (e.g., monthly) and/or in response to a specific type of event to identify new trends in the desirability of certain features. A population algorithm may then analyze the population data (e.g., using statistical analysis to determine the mean and variance, of weightings on each attribute and for each attribute level, and the covariances between attributes). For example, the population data may be processed using an accelerated heterogeneity algorithm, which performs a statistical analysis that estimates covarying bell curves for respective product characteristics.”).
Claim 15 recites limitations substantially similar to claim 5, and as such a similar analysis applies.
Claims 9 and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Ren, Xu, Huang, Liao, and Yu as applied to claims 1-3, 7, and 11-13 above, and further in view of Li et al. ("Deep Learning-Based Recommendation System: Systematic Review and Classification." IEEE Access (2023).) and Meiseles et al. ((2020). Source model selection for deep learning in the time series domain. IEEE Access, 8, 6190-6200.), referred to as Li and Meiseles respectively.
Claim 9: Ren, Xu, Huang, Liao, and Yu teaches the limitations of claim 1. Li, in the same field of recommendation machine learning model implementation, teaches the following limitation which Ren, Xu, Huang, Liao, and Yu fail to teach:
classify an evaluation metric of the augmented plurality of evaluation metrics to a domain of a plurality of domains using a domain classifier; (Li, page 113806, section 2, paragraph 1, “Our approach to domain classification involves extracting and categorizing domain-specific information using rigorous strategies. We extract relevant terms from various sources, such as article titles, authors, and index keywords, to classify 58 domains into eight categories based on industries and data types”, Li teaches a domain classifier using natural language input. Li trains a model to recognize key terms from various sources and determines a domain.)
Ren, Liao, and Li utilize machine learning models and algorithms to create systems for the purpose of improving upon recommendation systems. Ren teaches a system that encapsulates semantic information of users and items so that it can be used as enriched input to recommendation systems. Liao teaches a recommendation system machine learning model in the form of a natural language conversation. Li teaches using a domain classifier to determine a domain from which key words are directing to. It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have incorporated the teachings disclosed by Ren, Xu, Huang, Liao, and Yu with the teachings disclosed by Li (i.e. using a domain classifier to determine domain of the recommendation system). A motivation for the combination is to develop machine learning models specific to a domain, allowing for increased accuracy in domain-specific information, (Li, col. 2, paragraph 1, “Each domain category represents a context where deep learning techniques are harnessed to provide relevant, personalized, and effective recommendations, catering to users’diverse needs and preferences.”).
Meiseles, in the same field of domain modeling, teaches the following limitation which Ren, Xu, Huang, Liao, and Yu and Li fail to teach:
and select the excitation element machine learning model from a plurality of excitation element machine learning models as a function of the domain. (Meiseles, page 6195, figure 2, “A visual depiction of source model selection for a single target dataset, from left to right. (a) First, the training target sequences are encoded using each of the truncated pre-trained source models. (b) Next, the MSC is calculated for each of the source model encodings in encoding space. (c) Finally, the source models are sorted by their MSC, in non-increasing order, producing a ranking of estimated performance of transfer learning from each source model.”, Meiseles outlines a system that determines a best machine learning model to use for a given domain.)
Ren, Liao, Li, and Meiseles utilize machine learning models and algorithms to create systems for the purpose of improving upon recommendation systems. Ren teaches a system that encapsulates semantic information of users and items so that it can be used as enriched input to recommendation systems. Liao teaches a recommendation system machine learning model in the form of a natural language conversation. Li teaches using a domain classifier to determine a domain from which key words are directing to. Meiseles teaches a system for choosing an optimal model based on a given domain. It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have incorporated the teachings disclosed by Ren, Xu, Huang, Liao, Yu and Li, with the teachings disclosed by Meiseles (i.e. determining a best model based on domain). A motivation for the combination is to leverage a most optimal model with a given target domain, (Meiseles, page 6198, col. 2, section A.1, “As we mentioned earlier the expected top-1 accuracy for random selection is 1:01%. Our method ranks the optimal source model first 6 times, yielding a top-1 accuracy of 7.1%.IDS ranks the optimal source model first 4 times, yielding a top-1 accuracy of 4.7%.”).
Claim 19 recites limitations substantially similar to claim 9, and as such a similar analysis applies.
Claims 10 and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Ren, Xu, Huang, Liao, Yu, Li, and Meiseles as applied to claims 9 and 19 above, and further in view of Eitel et al. (US 20240054510 A1), referred to as Eitel.
Claim 10: Ren, Xu, Huang, Liao, Yu, Li and Meiseles teach the limitations of claim 9. Eitel, in the same field of recommendation machine learning model implementation, teaches the following limitation which Ren, Xu, Huang, Liao, Yu, Li and Meiseles fail to teach:
wherein the plurality of domains comprises a spiritual domain, a marriage domain, a family domain, a health domain, a virtue domain, an emotional domain, a financial domain, a vocational domain, an intellectual domain, a lifestyle domain, an interest domain, and a social domain. (Eitel, paragraph 4, “A user's intentions can take various forms. For example, “intentions” can be used to describe a person's ideal envisioning of themselves. This envisioning generally includes several large aspirations or goals a person wishes to achieve throughout their life. Example ‘domains’ of these life aspirations and goals can include career development, financial security, achieving a certain social status or standing, improvement in physical health and fitness, mental and emotional wellness, family, friends and romantic relationships, hobbies, and many other aspirations. In some examples, “intentions,” can (and often do) include a multitude of aspirations or goals spread across different life domains.”, Eitel teaches a recommendation system that categorizes a user’s intention’s through multiple domains of life. It is interpreted by the examiner these domains are substantially similar to the domains listed in this application.)
Ren, Liao, and Eitel utilize machine learning models and algorithms to create systems for the purpose of improving upon recommendation systems. Ren teaches a system that encapsulates semantic information of users and items so that it can be used as enriched input to recommendation systems. Liao teaches a recommendation system machine learning model in the form of a natural language conversation. Eitel teaches categorizing a user’s intentions into certain domains of life as to further optimize the learning model based on domain. It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have incorporated the teachings disclosed by Ren, Xu, Huang, Liao, Yu, Li, and Meiseles with the teachings disclosed by Eitel (i.e. categorizing domains of life to a user’s intentions). A motivation for the combination is to develop machine learning models specific to a domain, allowing for increased accuracy in domain-specific information, (Eitel, paragraph 6, “To identify an intention gap between one's intentions and their actions (or inaction), the meKey system first ascertains what life domain aspirations the user possesses. In some embodiments, this solicitation can take the form of a series of questions or a survey to be completed by the user, or it can be questions interspersed and integrated into their interaction with their media.”).
Claim 20 recites limitations substantially similar to claim 10, and as such a similar analysis applies.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Yu, Y., Chen, X., Ai, Q., Yang, L., & Croft, W. B. (2018, October). Towards conversational search and recommendation: System ask, user respond. In Proceedings of the 27th acm international conference on information and knowledge management (pp. 177-186).
Zamani, H., Mitra, B., Chen, E., Lueck, G., Diaz, F., Bennett, P. N., ... & Dumais, S. T. (2020, July). Analyzing and learning from user interactions for search clarification. In Proceedings of the 43rd international acm sigir conference on research and development in information retrieval (pp. 1181-1190).
Jiang, J., Yu, N., & Lin, C. Y. (2012, April). Focus: learning to crawl web forums. In Proceedings of the 21st International Conference on World Wide Web (pp. 33-42).
Hernandez, J., Marin-Castro, H. M., & Morales-Sandoval, M. (2020). A semantic focused web crawler based on a knowledge representation schema. Applied Sciences, 10(11), 3837.
Sun, Y., & Zhang, Y. (2018, June). Conversational recommender system. In The 41st international acm sigir conference on research & development in information retrieval (pp. 235-244).
Lei, W., He, X., Miao, Y., Wu, Q., Hong, R., Kan, M. Y., & Chua, T. S. (2020, January). Estimation-action-reflection: Towards deep interaction between conversational and recommender systems. In Proceedings of the 13th international conference on web search and data mining (pp. 304-312).
Any inquiry concerning this communication or earlier communications from the examiner should be directed to HYUNGJUN B YI whose telephone number is (703)756-4799. The examiner can normally be reached M-F 9-5.
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, Usmaan Saeed can be reached on (571) 272-4046. 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.
/H.B.Y./Examiner, Art Unit 2124
/USMAAN SAEED/Supervisory Patent Examiner, Art Unit 2146
Prosecution Insights