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 .
Response to Amendments
The action is responsive to the Applicant’s Amendment filed on 6/16/2026. Claims 1-20 are pending in the application. Claims 1, 3-5, 14, 15, 17, 19, and 20 are amended.
Applicant’s amendments to the claims integrate the processes into a practical application. The 101 rejection of claims 1-20 previously set forth in the Non-Final Office Action mailed 3/16/2026 is hereby withdrawn.
Response to Arguments
3. Applicant’s arguments with respect to the rejections previously made and the amended claims filed on 6/16/2026 have been fully considered but they are not persuasive. In view of the claim amendments, the rejections are being updated accordingly.
Further, regarding the new limitations recited in claims 1, 3-5, 14, 15, 17, 19, and 20, it is submitted that they are properly addressed.
Furthermore, it is also submitted that all limitations in pending claims, including those not specifically argued, are properly addressed. The reason is set forth in the rejections. See claim analysis below for detail.
Claim Rejections - 35 USC § 112
The following is a quotation of the first paragraph of 35 U.S.C. 112(a):
(a) IN GENERAL.—The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor or joint inventor of carrying out the invention.
Claims 3, 10, and 17 are rejected under 35 U.S.C. 112(a) or 35 U.S.C. 112 (pre-AIA ), first paragraph, as failing to comply with the written description requirement. The claim(s) contains subject matter which was not described in the specification in such a way as to reasonably convey to one skilled in the relevant art that the inventor or a joint inventor, or for applications subject to pre-AIA 35 U.S.C. 112, the inventor(s), at the time the application was filed, had possession of the claimed invention.
Claims 4, 14, and 20 recite “encoding a frequency count of transitions from a first topic category to a second topic category into a directed edge between a corresponding pair of nodes based on a chronological order in which the user accessed content items of the first topic category followed by content items of the second topic category in the historical content item sequence; and encoding a frequency count of consecutive accesses to content items of a same topic category into at least one self-loop edge associated with a corresponding node” that is directed to new matter and fails to comply with the written description requirement. The specification does not teach “encoding a frequency count of transitions”. Page 11 of the specification states, “It should be appreciated that the process for constructing the topic graph described above in conjunction with FIG.2 to FIG.3 is merely exemplary. Depending on actual application requirements, the steps in the process for constructing the topic graph may be replaced or modified in any manner, and the process may comprise more or fewer steps.”
Claims 3, 17, and 19 recite “wherein the assigning the plurality of nodes comprises initializing each node with a topic category embedding generated from a pre-trained word embedding model” that is directed to new matter and fails to comply with the written description requirement.
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 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.
Claims 1-20 are rejected under 35 U.S.C. 103 as being unpatentable over Dongho et al. ("News Recommendation with Topic-Enriched Knowledge Graphs", Proceedings of the 7th ACM Conference on Information- Centric Networking, Acmpub27, 19 October 2020 (2020-10-19), pages 695-704) in view of Chatterjee et al. (US 20210097140 A1) and Mao et al. (“Neural News Recommendation with Collaborative News Encoding and Structural User Encoding”, ARXIV.org, (2021-09-02)).
Regarding Claim 1, Dongho discloses a method for hierarchical representation learning of user interest ([Page 695, Introduction]: Thus, it is essential for users to automatically receive personalized recommendations based on their interests;[Page 701]: LSTUR[1]is a neural news recommendation method that can learn both long and short-term user representations through a news encoder and user encode), comprising:
obtaining a historical content item sequence of a user comprising a plurality of content items accessed over a time period (Fig. 3; [Page 696]: we propose a Topic-Enriched Knowledge Graph Recommendation System (TEKGR)… which takes a piece of candidate news and a user’s click history as input; [Page 701, section 4.2]: we denote user i's clicked history);
identifying a topic and a text of each historical content item in the historical content item sequence, to obtain a topic sequence and a text sequence corresponding to the historical content item sequence (Fig. 3; [Page 699, sections 4.1.1. - 4.1.3]: The knowledge encoder also has three layers to learn topical information on the news title… The goal of Topic Enhanced KG Construction is to consider contextual knowledge information along with topical information for the recommendation system… Word embedding, the first layer, converts a news title from a sequence of words to a sequence of dense semantic vectors);
automatically constructing, by a computing device, a topic graph from the topic sequence, the automatically constructing comprising: assigning a plurality of nodes to the topic graph, each node representing a distinct topic category from the topic sequence ([Page 696]: Figure 1: Illustration of news title with multiple entities in KG… Figure 2 concisely illustrates the example subgraph of Figure 1.Green nodes (Trump and Madonna) are the recognized entities connected with their 2-hop neighbor nodes);
encoding a frequency count of transitions from a first topic category to a second topic category into a directed edge between a corresponding pair of nodes based on a chronological order in which the user accessed content items of the first topic category followed by content items of the second topic category in the historical content item sequence (Figs. 1-4; [Pages 699-701]: The user features are a concatenation of normalized count features from clicked news’ topics and subtopics and TF-IDF features from the clicked news); and
encoding a frequency count of consecutive accesses to content items of a same topic category into at least one self-loop edge associated with a corresponding node (Figs. 1-4; [Pages 699-701]: As shown in Figure 4, the KG-based news modeling layer has three encoders to extract the news representation vector… For example, a user with a huge interest…. would click…news titles that have financial terms in them… we can obtain the final representation of a news title in word-level news encoder);
generating, by the computing device using a graph neural network comprising multiple graph convolutional layers, a comprehensive topic representation from the topic graph by iteratively aggregating features from neighboring nodes via the directed edges through trainable weight matrices ([Pages 696-699]: LSTUR[1] learns the representations of news from their titles and explicitly given topic information… For a given user, KGNN-LS utilizes a trainable function to highlight valuable relations in KG and transforms the KG into a personalized weighted graph… Through summing contextual representations of words weighted by the attention weights, we can obtain the final representation of a news title in word-level news encoder);
However, Dongho does not explicitly teach “generating, by the computing device, a comprehensive text representation based on the text sequence, the comprehensive text representation being an aggregation of textual contents of historical content items in the historical content item sequence; generating, by the computing device, a user interest representation of the user based on the comprehensive topic representation and the comprehensive text representation; and outputting, by the computing device, the user interest representation to a recommendation engine to generate a personalized content recommendation for the user.”
On the other hand, in the same field of endeavor, Chatterjee teaches
generating, by the computing device, a comprehensive text representation based on the text sequence, the comprehensive text representation being an aggregation of textual contents of historical content items in the historical content item sequence ([0006]: A system and method for generating a conversation graph to present information about a plurality of conversations with similar intents is disclosed. In one aspect, the disclosure provides a method of generating a conversation graph for representation of a task. The method includes receiving a set of conversations and related meta data, where each conversation includes a plurality of word sequences);
Additionally, Mao teaches generating, by the computing device, a user interest representation of the user based on the comprehensive topic representation and the comprehensive text representation ([Abstract]: SUE utilizes graph convolutional networks to extract cluster-structural features of user history, followed by intra-cluster and inter-cluster attention modules to learn hierarchical user interest representations); and
outputting, by the computing device, the user interest representation to a recommendation engine to generate a personalized content recommendation for the user ([Page 1]: Figure 1: (a) An example of user browsing history. (b) An example of news title-content semantic interaction; Figs 3-5; As shown in Figure 4, we visualize our model’s output title (content) attention weights αt(c) over the title (content) words of the news N6 in Figure 1).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the teaching of Dongho to incorporate the teachings of Chatterjee and Mao to include generating a comprehensive text representation based on the text sequence, generating, a user interest representation of the user and outputting the user interest representation to a recommendation engine to generate a personalized content recommendation for the user .
The motivation for doing so would be to represent many conversations in a single graph, as recognized by Chatterjee ([Abstract] of Chetterjee: The resultant classifications are used to represent the many conversations in a single graph by a plurality of nodes), and to provide personalized recommendations, as recognized by Mao ([Introduction] of Mao: With a deluge of news generated every day, an efficient news recommendation system should push relevant news to users to satisfy their diverse personalized interests).
Regarding Claim 2, the combined teachings of Dongho, Chatterjee, and Mao disclose the method of claim 1.
Dongho further teaches wherein the comprehensive topic representation and the comprehensive text representation have different information abstraction levels ([Page 696]: Nevertheless, in addition to the title, our model is also feasible to use full article contents or abstracts of news. Note that our method can be simply utilized in any kind of recommendation scenario with short texts).
Regarding Claim 3, the combined teachings of Dongho, Chatterjee, and Mao disclose the method of claim 1.
Chatterjee further teaches wherein the assigning the plurality of nodes comprises initializing each node with a topic category embedding generated from a pre-trained word embedding mode1 ([0077]: In some embodiments, conversation meta data is the main input to the clustering algorithm to cluster the conversations… Unseen words in the text may be initialized using a normal distribution between a range (−1, 1)).
Regarding Claim 4, the combined teachings of Dongho, Chatterjee, and Mao disclose the method of claim 1.
Chatterjee further teaches wherein the automatically constructing the topic graph further comprises: computing a degree matrix from the frequency counts of transitions encoded in the directed edges; and normalizing directed edges using the degree matrix to generate normalized edge weights for use in the graph neural network ([0077]: the meta data (e.g., conversation summary) can be in the form of text data, which is converted to numerical representation by using domain word2vec model. In some embodiments, the text may be converted by performing one or more of the following steps: (1) the text is normalized by removing all the punctuation).
Regarding Claim 5, the combined teachings of Dongho, Chatterjee, and Mao disclose the method of claim 1.
Mao further teaches wherein the iteratively aggregating features from neighboring nodes comprises: aggregating features from first-hop neighboring nodes in a first graph convolutional layer; and aggregating features from second-hop neighboring nodes in a second graph convolutional layer to capture higher-order topic transition patterns in the historical content item sequence ([Page 4]: With aspects of user interests encoded within specific clusters, overall user representations can be aggregated by leveraging the correlation among interest clusters.);
Regarding Claim 6, the combined teachings of Dongho, Chatterjee, and Mao disclose the method of claim 1.
Dongho further teaches wherein the generating a comprehensive text representation comprises:
generating a comprehensive text attention representation through an attention mechanism based on the text sequence ([Pages 695- 696]: After obtaining news representation vectors, an attention network compares clicked news to the candidate news in order to get the user’s final embedding… With extracted news representation, we use an attention layer to compare clicked news to the candidate news in order to get a user’s final embedding), and
the generating a user interest representation comprises: generating the user interest representation based on the comprehensive topic representation and the comprehensive text attention representation (Fig. 3; [Page 696, section 4.1.1]: With extracted news representation, we use an attention layer to compare clicked news to the candidate news in order to get a user’s final embedding).
Regarding Claim 7, the combined teachings of Dongho, Chatterjee, and Mao disclose the method of claim 1.
Dongho further teaches wherein the generating a comprehensive text representation comprises: generating a comprehensive text capsule representation using a capsule network based at least on the text sequence, and the generating a user interest representation comprises: generating the user interest representation based on the comprehensive topic representation and the comprehensive text capsule representation ([Page 696, section 4.1.1 and figure 3]: Since recurrent neural networks(RNN) is proved to be effective for sentence modeling[36], we apply Bi-GRU… Through summing contextual representations of words weighted by the attention weights, we can obtain the final representation of a news title in word-level news encoder; [The broad and undefined claimed terms "capsule network" and "text capsule representation" correspond to the network employed and the semantic vector calculated in section 4.1.1]).
Additionally, Mao teaches wherein the comprehensive text capsule representation is a group of neurons that output vectors, and wherein the capsule network models hierarchical relationships in the text sequence to generate the comprehensive text capsule representation ([Pages 47-48]: Based on this sequential formulation, recurrent neural networks…are proposed to encode user history… In recent years, deep neural models have achieved superior performance in news recommendation. Many studies pinpointed that this improvement came from the fine-grained news and user representations, which were extracted by deep neural networks… Concretely, we utilize the semantic memory vector).
Regarding Claim 8, the combined teachings of Dongho, Chatterjee, and Mao disclose the method of claim 1.
Dongho further teaches wherein the generating a comprehensive text representation comprises: generating a comprehensive text attention representation through an attention mechanism based on the text sequence; and generating a comprehensive text capsule representation using a capsule network based at least on the text sequence, and
the generating a user interest representation comprises: generating the user interest representation based on the comprehensive topic representation, the comprehensive text attention representation, and the comprehensive text capsule representation (Fig. 3; [Page 701]: Now that we obtain news representation vector from multiple encoders, we use an attention layer to compare clicked news to the candidate news in order to get the user’s final embedding).
Additionally, Mao teaches wherein the comprehensive text capsule representation is a group of neurons that output vectors, and wherein the capsule network models hierarchical relationships in the text sequence to generate the comprehensive text capsule representation ([Pages 47-48]: Based on this sequential formulation, recurrent neural networks…are proposed to encode user history… In recent years, deep neural models have achieved superior performance in news recommendation. Many studies pinpointed that this improvement came from the fine-grained news and user representations, which were extracted by deep neural networks… Concretely, we utilize the semantic memory vector).
Regarding Claim 9, the combined teachings of Dongho, Chatterjee, and Mao disclose the method of claim 8.
Dongho further teaches wherein the comprehensive text attention representation and the comprehensive text capsule representation have different information abstraction levels ([Page 696]: Nevertheless, in addition to the title, our model is also feasible to use full article contents or abstracts of news. Note that our method can be simply utilized in any kind of recommendation scenario with short texts).
Regarding Claim 10, the combined teachings of Dongho, Chatterjee, and Mao disclose the method of claim 7.
Dongho further teaches wherein the generating a comprehensive text capsule representation comprises:
generating an interest capsule representation using the capsule network based on the text sequence; generating a target content item representation of a target content item; and generating the comprehensive text capsule representation through an attention mechanism based on the interest capsule representation and the target content item representation (Fig. 3; [Page 701]: Now that we obtain news representation vector from multiple encoders, we use an attention layer to compare clicked news to the candidate news in order to get the user’s final embedding).
Regarding Claim 11, the combined teachings of Dongho, Chatterjee, and Mao disclose the method of claim 1.
Dongho further teaches further comprising: predicting a click probability of the user clicking a target content item based on the user interest representation and a target content item representation of the target content item ([Page 698, section 4.2 second paragraph]: Using users’ click history along with the link between words in the titles and entities in the knowledge graph, we aim to predict whether user 𝑖 has a potential interest in news title 𝑡 , which has not been clicked before).
Regarding Claim 12, the combined teachings of Dongho, Chatterjee, and Mao disclose the method of claim 11.
Dongho further teaches wherein the click probability is output through a click probability predicting model, and a training of the click probability predicting model comprises:
constructing a training dataset, the training dataset including a plurality of positive samples and a plurality of negative sample sets corresponding to the plurality of positive samples ([Page 701]: For experiments, we applied pre-trained GloVe embedding for the initialization of the word embeddings, which has 200 dimensions. Besides, Microsoft Satori was used to construct the subgraph for a given dataset);
generating a plurality of posterior click probabilities corresponding to the plurality of positive samples ([Page 701]: For the click- through rate (CTR) prediction, we employed the trained model to each line of the dataset in the test dataset and obtained the predicted click probability);
generating a prediction loss based on the plurality of posterior click probabilities; and optimizing the click probability predicting model through minimizing the prediction loss ([Page 698]: we employed multitask learning by utilizing the loss function of short-text modules in both the classification task and recommendation task; [Page 701, section 4.2 second paragraph]: To evaluate the K recommended list, we chose the F1 score metric. We ran each experiment 10 times independently and reported the average and maximum deviation as results).
Regarding Claim 13, the combined teachings of Dongho, Chatterjee, and Mao disclose the method of claim 12.
Dongho further teaches wherein the generating a plurality of posterior click probabilities comprises, for each positive sample: predicting a click probability of the positive sample corresponding to the positive sample; for each negative sample in a negative sample set corresponding to the positive sample, predicting a negative sample click probability corresponding to the negative sample, to obtain a negative sample click probability set corresponding to the negative sample set ([Page 701]: For the click- through rate (CTR) prediction, we employed the trained model to each line of the dataset in the test dataset and obtained the predicted click probability); and
calculating a posterior click probability corresponding to the positive sample based on the positive sample click probability and the negative sample click probability set ([Page 701, section 4.2 second paragraph]: For top-K recommendation, we applied the trained model to choose K items with the highest predicted click probability for each user in the test set. To evaluate the K recommended list, we chose the F1 score metric. We ran each experiment 10 times independently and reported the average and maximum deviation as results).
Regarding Claim 14, Dongho discloses an apparatus for hierarchical representation learning of user interest, comprising: at least one processor; and a memory storing computer-executable instructions that ([Page 695]: Online news websites, such as MSN News and Google News, collect news contents from various sources to provide them to users), when executed, cause the at least one processor to perform operations comprising:
obtain a historical content item sequence of a user comprising a plurality of content items accessed over a time period (Fig. 3; [Page 696]: we propose a Topic-Enriched Knowledge Graph Recommendation System (TEKGR)… which takes a piece of candidate news and a user’s click history as input; [Page 701, section 4.2]: we denote user i's clicked history);
identifying a topic and a text of each historical content item in the historical content item sequence, to obtain a topic sequence and a text sequence corresponding to the historical content item sequence (Fig. 3; [Page 699, sections 4.1.1. - 4.1.3]: The knowledge encoder also has three layers to learn topical information on the news title… The goal of Topic Enhanced KG Construction is to consider contextual knowledge information along with topical information for the recommendation system… Word embedding, the first layer, converts a news title from a sequence of words to a sequence of dense semantic vectors);
automatically constructing, by a computing device, a topic graph from the topic sequence, the automatically constructing comprising: assigning a plurality of nodes to the topic graph, each node representing a distinct topic category from the topic sequence ([Page 696]: Figure 1: Illustration of news title with multiple entities in KG… Figure 2 concisely illustrates the example subgraph of Figure 1.Green nodes (Trump and Madonna) are the recognized entities connected with their 2-hop neighbor nodes);
encoding a frequency count of transitions from a first topic category to a second topic category into a directed edge between a corresponding pair of nodes based on a chronological order in which the user accessed content items of the first topic category followed by content items of the second topic category in the historical content item sequence (Figs. 1-4; [Pages 699-701]: The user features are a concatenation of normalized count features from clicked news’ topics and subtopics and TF-IDF features from the clicked news); and
encoding a frequency count of consecutive accesses to content items of a same topic category into at least one self-loop edge associated with a corresponding node (Figs. 1-4; [Pages 699-701]: As shown in Figure 4, the KG-based news modeling layer has three encoders to extract the news representation vector… For example, a user with a huge interest…. would click…news titles that have financial terms in them… we can obtain the final representation of a news title in word-level news encoder);
generating, by the computing device using a graph neural network comprising multiple graph convolutional layers, a comprehensive topic representation from the topic graph by iteratively aggregating features from neighboring nodes via the directed edges through trainable weight matrices ([Pages 696-699]: LSTUR[1] learns the representations of news from their titles and explicitly given topic information… For a given user, KGNN-LS utilizes a trainable function to highlight valuable relations in KG and transforms the KG into a personalized weighted graph… Through summing contextual representations of words weighted by the attention weights, we can obtain the final representation of a news title in word-level news encoder);
However, Dongho does not explicitly teach “generating a comprehensive text representation based on the text sequence, the comprehensive text representation being an aggregation of textual contents of historical content items in the historical content item sequence; generating a user interest representation of the user based on the comprehensive topic representation and the comprehensive text representation; and outputting the user interest representation to a recommendation engine to generate a personalized content recommendation for the user.”
On the other hand, in the same field of endeavor, Chatterjee teaches
generating a comprehensive text representation based on the text sequence, the comprehensive text representation being an aggregation of textual contents of historical content items in the historical content item sequence ([0006]: A system and method for generating a conversation graph to present information about a plurality of conversations with similar intents is disclosed. In one aspect, the disclosure provides a method of generating a conversation graph for representation of a task. The method includes receiving a set of conversations and related meta data, where each conversation includes a plurality of word sequences);
Additionally, Mao teaches generating a user interest representation of the user based on the comprehensive topic representation and the comprehensive text representation ([Abstract]: SUE utilizes graph convolutional networks to extract cluster-structural features of user history, followed by intra-cluster and inter-cluster attention modules to learn hierarchical user interest representations); and
outputting the user interest representation to a recommendation engine to generate a personalized content recommendation for the user ([Page 1]: Figure 1: (a) An example of user browsing history. (b) An example of news title-content semantic interaction; Figs 3-5; As shown in Figure 4, we visualize our model’s output title (content) attention weights αt(c) over the title (content) words of the news N6 in Figure 1).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the teaching of Dongho to incorporate the teachings of Chatterjee and Mao to include generating a comprehensive text representation based on the text sequence, generating, a user interest representation of the user and outputting the user interest representation to a recommendation engine to generate a personalized content recommendation for the user .
The motivation for doing so would be to represent many conversations in a single graph, as recognized by Chatterjee ([Abstract] of Chetterjee: The resultant classifications are used to represent the many conversations in a single graph by a plurality of nodes), and to provide personalized recommendations, as recognized by Mao ([Introduction] of Mao: With a deluge of news generated every day, an efficient news recommendation system should push relevant news to users to satisfy their diverse personalized interests).
Regarding Claim 15, Dongho discloses at least one non-transitory machine-readable medium comprising instructions for hierarchical representation learning of user interest ([Page 695]: Online news websites, such as MSN News and Google News, collect news contents from various sources to provide them to users), that, when executed by at least one processor, causes the at least one processor to perform operations comprising:
obtaining a historical content item sequence of a user comprising a plurality of content items accessed over a time period (Fig. 3; [Page 696]: we propose a Topic-Enriched Knowledge Graph Recommendation System (TEKGR)… which takes a piece of candidate news and a user’s click history as input; [Page 701, section 4.2]: we denote user i's clicked history);
identify a topic and a text of each historical content item in the historical content item sequence, to obtain a topic sequence and a text sequence corresponding to the historical content item sequence (Fig. 3; [Page 699, sections 4.1.1. - 4.1.3]: The knowledge encoder also has three layers to learn topical information on the news title… The goal of Topic Enhanced KG Construction is to consider contextual knowledge information along with topical information for the recommendation system… Word embedding, the first layer, converts a news title from a sequence of words to a sequence of dense semantic vectors);
automatically constructing, by a computing device, a topic graph from the topic sequence, the automatically constructing comprising: assigning a plurality of nodes to the topic graph, each node representing a distinct topic category from the topic sequence ([Page 696]: Figure 1: Illustration of news title with multiple entities in KG… Figure 2 concisely illustrates the example subgraph of Figure 1.Green nodes (Trump and Madonna) are the recognized entities connected with their 2-hop neighbor nodes);
encoding a frequency count of transitions from a first topic category to a second topic category into a directed edge between a corresponding pair of nodes based on a chronological order in which the user accessed content items of the first topic category followed by content items of the second topic category in the historical content item sequence (Figs. 1-4; [Pages 699-701]: The user features are a concatenation of normalized count features from clicked news’ topics and subtopics and TF-IDF features from the clicked news); and
encoding a frequency count of consecutive accesses to content items of a same topic category into at least one self-loop edge associated with a corresponding node (Figs. 1-4; [Pages 699-701]: As shown in Figure 4, the KG-based news modeling layer has three encoders to extract the news representation vector… For example, a user with a huge interest…. would click…news titles that have financial terms in them… we can obtain the final representation of a news title in word-level news encoder);
generating, by the computing device using a graph neural network comprising multiple graph convolutional layers, a comprehensive topic representation from the topic graph by iteratively aggregating features from neighboring nodes via the directed edges through trainable weight matrices ([Pages 696-699]: LSTUR[1] learns the representations of news from their titles and explicitly given topic information… For a given user, KGNN-LS utilizes a trainable function to highlight valuable relations in KG and transforms the KG into a personalized weighted graph… Through summing contextual representations of words weighted by the attention weights, we can obtain the final representation of a news title in word-level news encoder);
However, Dongho does not explicitly teach “generating a comprehensive text representation based on the text sequence, the comprehensive text representation being an aggregation of textual contents of historical content items in the historical content item sequence; generating a user interest representation of the user based on the comprehensive topic representation and the comprehensive text representation; and outputting the user interest representation to a recommendation engine to generate a personalized content recommendation for the user.”
On the other hand, in the same field of endeavor, Chatterjee teaches
Generating a comprehensive text representation based on the text sequence, the comprehensive text representation being an aggregation of textual contents of historical content items in the historical content item sequence ([0006]: A system and method for generating a conversation graph to present information about a plurality of conversations with similar intents is disclosed. In one aspect, the disclosure provides a method of generating a conversation graph for representation of a task. The method includes receiving a set of conversations and related meta data, where each conversation includes a plurality of word sequences);
Additionally, Mao teaches generating a user interest representation of the user based on the comprehensive topic representation and the comprehensive text representation ([Abstract]: SUE utilizes graph convolutional networks to extract cluster-structural features of user history, followed by intra-cluster and inter-cluster attention modules to learn hierarchical user interest representations); and
outputting the user interest representation to a recommendation engine to generate a personalized content recommendation for the user ([Page 1]: Figure 1: (a) An example of user browsing history. (b) An example of news title-content semantic interaction; Figs 3-5; As shown in Figure 4, we visualize our model’s output title (content) attention weights αt(c) over the title (content) words of the news N6 in Figure 1).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the teaching of Dongho to incorporate the teachings of Chatterjee and Mao to include generating a comprehensive text representation based on the text sequence, generating, a user interest representation of the user and outputting the user interest representation to a recommendation engine to generate a personalized content recommendation for the user .
The motivation for doing so would be to represent many conversations in a single graph, as recognized by Chatterjee ([Abstract] of Chetterjee: The resultant classifications are used to represent the many conversations in a single graph by a plurality of nodes), and to provide personalized recommendations, as recognized by Mao ([Introduction] of Mao: With a deluge of news generated every day, an efficient news recommendation system should push relevant news to users to satisfy their diverse personalized interests).
Regarding Claim 16, the combined teachings of Dongho, Chatterjee, and Mao disclose the apparatus of claim 14.
Mao further teaches wherein the comprehensive topic representation and the comprehensive text representation have different information abstraction levels ([Pages 49-52]: Besides intracluster refinement of user interests, modeling inter cluster correlation is also essential to leverage the overall information of user history… The GCN extracts structural features on graph G, refining specific user interest representations within clusters and aggregating overall user history information among clusters… deep neural models can learn refined representations adaptively, which are more effective than general feature engineering with fixed hancrafted features).
Regarding Claim 17, the combined teachings of Dongho, Chatterjee, and Mao disclose the apparatus of claim 14.
Chatterjee further teaches wherein the assigning the plurality of nodes comprises initializing each node with a topic category embedding generated from a pre-trained word embedding mode1 ([0077]: In some embodiments, conversation meta data is the main input to the clustering algorithm to cluster the conversations… Unseen words in the text may be initialized using a normal distribution between a range (−1, 1)).
Regarding Claim 18, the combined teachings of Dongho, Chatterjee, and Mao disclose the at least one non-transitory machine-readable medium of claim 15.
Mao further teaches wherein the comprehensive topic representation and the comprehensive text representation have different information abstraction levels ([Pages 49-52]: Besides intracluster refinement of user interests, modeling inter cluster correlation is also essential to leverage the overall information of user history… The GCN extracts structural features on graph G, refining specific user interest representations within clusters and aggregating overall user history information among clusters… deep neural models can learn refined representations adaptively, which are more effective than general feature engineering with fixed hancrafted features).
Regarding Claim 19, the combined teachings of Dongho, Chatterjee, and Mao disclose the at least one non-transitory machine-readable medium of claim 15.
Chatterjee further teaches wherein the assigning the plurality of nodes comprises initializing each node with a topic category embedding generated from a pre-trained word embedding mode1 ([0077]: In some embodiments, conversation meta data is the main input to the clustering algorithm to cluster the conversations… Unseen words in the text may be initialized using a normal distribution between a range (−1, 1)).
Regarding Claim 20, the combined teachings of Dongho, Chatterjee, and Mao disclose the at least one non-transitory machine-readable medium of claim 19.
wherein the automatically constructing the topic graph further comprises: computing a degree matrix from the frequency counts of transitions encoded in the directed edges; and normalizing directed edges using the degree matrix to generate normalized edge weights for use in the graph neural network ([0077]: the meta data (e.g., conversation summary) can be in the form of text data, which is converted to numerical representation by using domain word2vec model. In some embodiments, the text may be converted by performing one or more of the following steps: (1) the text is normalized by removing all the punctuation).
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 extension fee 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 SHIRLEY D. HICKS whose telephone number is (571)272-3304. The examiner can normally be reached Mon - Fri 7:30 - 4:00.
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, Charles Rones can be reached on (571) 272-4085. 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.
/S D H/Examiner, Art Unit 2168
/CHARLES RONES/Supervisory Patent Examiner, Art Unit 2168
Prosecution Insights