Hierarchical representation learning of user interests

Abstract

This disclosure presents methods, apparatus, and computer program products for hierarchical representation learning of user interests. A sequence of historical content items for a user can be obtained. The topic and text of each historical content item in the sequence can be identified to obtain a topic sequence and a text sequence corresponding to the historical content item sequence. A comprehensive topic representation can be generated based on the topic sequence. A comprehensive text representation can be generated based on the text sequence. A user interest representation for the user can be generated based on the comprehensive topic representation and the comprehensive text representation.

Description

Background Technology

[0001] With the development of internet technology and the growth of online information, recommendation systems are playing an increasingly important role in many online services. Different recommendation systems exist based on the content being recommended, such as news recommendation systems, music recommendation systems, movie recommendation systems, and product recommendation systems. These systems typically capture user interests and predict content that the user might be interested in based on those interests, then recommend it to the user. Summary of the Invention

[0002] This invention is provided to introduce a set of concepts, which will be further described in the following detailed description. This invention is not intended to identify key or essential features of the protected subject matter, nor is it intended to limit the scope of the protected subject matter.

[0003] Embodiments of this disclosure provide methods, apparatus, and computer program products for hierarchical representation learning of user interests. A sequence of historical content items for a user can be obtained. The topic and text of each historical content item in the sequence can be identified to obtain a topic sequence and a text sequence corresponding to the historical content item sequence. A comprehensive topic representation can be generated based on the topic sequence. A comprehensive text representation can be generated based on the text sequence. A user interest representation for the user can be generated based on the comprehensive topic representation and the comprehensive text representation.

[0004] It should be noted that one or more of the above aspects include the features specifically pointed out in the following detailed description and claims. Certain illustrative features of the one or more aspects are set forth in detail in the following specification and drawings. These features merely indicate various ways in which the principles of each aspect can be implemented, and this disclosure is intended to include all such aspects and their equivalents. Attached Figure Description

[0005] The following description will take into account several aspects disclosed, which are provided to illustrate rather than limit the aspects disclosed.

[0006] Figure 1 An exemplary process for learning hierarchical representations of user interests according to embodiments of this disclosure is illustrated.

[0007] Figure 2 Exemplary subject sequences and corresponding subject diagrams are shown according to embodiments of this disclosure.

[0008] Figure 3 An exemplary process for constructing a theme graph according to embodiments of this disclosure is shown.

[0009] Figure 4An exemplary process for generating a comprehensive topic representation according to an embodiment of this disclosure is shown.

[0010] Figure 5 An exemplary process for generating a comprehensive text attention representation according to an embodiment of this disclosure is shown.

[0011] Figure 6 An exemplary process for generating a comprehensive text capsule representation according to an embodiment of this disclosure is shown.

[0012] Figure 7 An exemplary process for predicting click probability according to an embodiment of this disclosure is shown.

[0013] Figure 8 An exemplary process for training a click probability prediction model according to embodiments of the present disclosure is shown.

[0014] Figure 9 This is a flowchart of an exemplary method for learning hierarchical representations of user interests according to embodiments of this disclosure.

[0015] Figure 10 An exemplary apparatus for learning hierarchical representations of user interests according to embodiments of this disclosure is shown.

[0016] Figure 11 An exemplary apparatus for learning hierarchical representations of user interests according to embodiments of this disclosure is shown. Detailed Implementation

[0017] This disclosure will now be discussed with reference to several exemplary embodiments. It should be understood that this discussion of embodiments is merely intended to enable those skilled in the art to better understand and thus implement the embodiments of this disclosure, and is not intended to teach any limitation on the scope of this disclosure.

[0018] To enable recommender systems to predict user interests and achieve efficient and personalized recommendations, it is necessary to model user interests and represent them as information in a form that the recommender system can understand and process. Typically, a user's previously clicked, accessed, or viewed historical content items can indicate their interests, thus a representation of their user interests can be generated based on these historical content items. In this paper, a content item can refer to an individual item with specific content. For example, a news article, a piece of music, or a movie can be considered a content item. Existing recommender systems typically use a single embedding vector to represent user interests. However, because user interests are complex—for example, different users often have different interests, the same user may have various interests, and the same content item may evoke different points of interest for different users—a single embedding vector may be insufficient to comprehensively and accurately represent user interests.

[0019] Embodiments of this disclosure propose a hierarchical representation learning method for user interests. For example, a user's historical content item sequence can be obtained. This sequence may include multiple historical content items that the user has previously clicked, accessed, or browsed. These historical content items may include, for example, news, music, movies, videos, books, product information, etc. Subsequently, the topic and text of each historical content item in the sequence can be identified to obtain a topic sequence and text sequence corresponding to the historical content item sequence. The text of a content item may include its title, summary, body, etc. Next, a comprehensive topic representation can be generated based on the topic sequence, and a comprehensive text representation can be generated based on the text sequence. The generated comprehensive topic representation and comprehensive text representation can be used to generate a user interest representation. The comprehensive topic representation and comprehensive text representation can have different levels of information abstraction. For example, the comprehensive topic representation may represent user interests at a coarser granularity, while the comprehensive text representation may represent user interests at a finer granularity. Therefore, the process described above for generating user interest representations can be considered a hierarchical representation learning process for user interests. Furthermore, the above process takes into account multiple aspects of historical content items, such as topic, title, summary, and main content, which can fully reflect the information of historical content items. Therefore, the hierarchical representation learning method for user interests according to embodiments of this disclosure can effectively and comprehensively capture user interests, thereby generating accurate and rich user interest representations. Further, the generated user interest representations can be used by a recommendation system to predict the click probability of a user clicking a target content item. Accurate and rich user interest representations can help the recommendation system predict more accurate click probabilities, thereby achieving efficient and targeted content item recommendations.

[0020] In one aspect, embodiments of this disclosure propose generating a comprehensive topic representation by constructing a topic graph corresponding to a topic sequence. In the topic graph, different nodes can represent different topic categories in the topic sequence, and edges between nodes can represent the order in which a user clicks on different content items with different topic categories. When generating the comprehensive topic representation, the representations of the neighboring nodes of each node can be aggregated using relational information derived from the topic graph representing the relationships between multiple nodes in the topic graph, and the representation of the node can be updated using the aggregated neighboring node representations. The updated representations of the individual nodes can then be combined into a comprehensive topic representation. In this document, a node's neighboring nodes can refer to nodes with which there exists an edge. In this way, the inherent relationships between neighboring nodes can be propagated via structural connections in the topic graph, thereby better capturing information related to the topic sequence.

[0021] In another aspect, embodiments of this disclosure propose multiple methods for generating comprehensive text representations. In one implementation, a comprehensive text representation can be generated based on a text sequence using an attention mechanism. Hereinafter, the representation generated based on a text sequence using an attention mechanism is referred to as a comprehensive text attention representation. In another implementation, a comprehensive text representation can be generated based on a text sequence using a capsule network. Hereinafter, the representation generated based on a text sequence using a capsule network is referred to as a comprehensive text capsule representation. These two implementations can be performed individually or in combination. Comprehensive text attention representations and comprehensive text capsule representations can have different levels of information abstraction. For example, relatively speaking, a comprehensive text attention representation can represent user interests at a coarser granularity, while a comprehensive text capsule representation can represent user interests at a finer granularity.

[0022] In another aspect, embodiments of this disclosure propose a machine learning model capable of generating a user interest representation using the hierarchical representation learning method for user interests described above and predicting the click probability of the user clicking a target content item. The target content item can be a content item from a set of candidate content items. The target content item can have the same content as the corresponding historical content item, including, for example, news, music, movies, videos, books, product information, etc. For each content item in the set of candidate content items, the click probability of the user clicking that content item can be predicted, thereby obtaining a set of click probabilities. Based on the predicted set of click probabilities, it can be determined which content items(s) to recommend to the user. The machine learning model used to predict the click probability of the user clicking a target content item can be called a click probability prediction model. Embodiments of this disclosure propose using a negative sampling method to train the click probability prediction model. For example, a training dataset including multiple positive samples and multiple sets of negative samples corresponding to the multiple positive samples can be constructed. Content items previously clicked by the user can be considered positive samples, while the set of content items presented in the same session as the positive sample but not clicked by the user can be considered the set of negative samples corresponding to the positive sample. The posterior click probability corresponding to a positive sample can be generated based on the positive sample and the set of negative samples corresponding to that positive sample. After obtaining multiple posterior click probabilities corresponding to multiple positive samples, a prediction loss can be generated, and the click probability prediction model can be optimized by minimizing this prediction loss.

[0023] It should be understood that although the foregoing and following discussions may involve generating user interest representations for news recommendation systems using news items as an example, the embodiments of this disclosure are not limited to this, but user interest representations for other types of recommendation systems can be generated in a similar manner using music, movies, videos, books, product information, etc.

[0024] Figure 1 An exemplary process 100 for hierarchical representation learning of user interests according to an embodiment of the present disclosure is illustrated. In process 100, a user interest representation 112 may be generated by a user interest representation generation unit 110, based at least on a user's historical content item sequence 102.

[0025] First, the user's historical content item sequence 102 can be obtained. The historical content item sequence 102 can include multiple historical content items that the user has previously clicked, accessed, or browsed, such as historical content item 102-1 to historical content item 102- ,in This refers to the number of historical content items. Historical content items can include, for example, news, music, movies, videos, books, product information, etc. A user's historical content item sequence 102 can indicate the user's interests. Taking news as an example of historical content items, the news items the user has previously clicked on can indicate which news items the user is interested in.

[0026] Subsequently, the topic of each historical content item in the historical content item sequence 102 can be identified. For example, for historical content item 102-1, its topic 114-1 can be identified, and for historical content item 102- It can identify its topic 114- Because even entirely new content items can be directly mapped to a specific topic, considering the topics of historical content items when generating user interest representations can help address data sparsity and cold start problems. The text of each historical content item in the sequence 102 can also be identified. The text of historical content items can include, for example, titles, summaries, and body content. Figure 1 For the sake of brevity, only the title and abstract are shown. For example, for historical content item 102-1, its title 116-1 and abstract 118-1 can be identified, and for historical content item 102- It can identify its title 116- and abstract 118- However, it should be understood that other text, such as the main content, can also be identified for each historical content item. The topic, title, summary, and main content of a historical content item can fully reflect its information. User interest representations generated based on these aspects of historical content items can effectively and comprehensively capture user interests, thereby generating accurate and rich user interest representations.

[0027] The topics and text of historical content items can be identified using known methods. The following example, using news articles as a historical content item, illustrates an exemplary process for identifying topics and text. News articles can be in the form of web pages. The Hypertext Markup Language (HTML) of the web page can be parsed to obtain the title and body content of the news article. The obtained title and / or body content can be input into a trained topic model. The topic model can output the topic corresponding to the news article. Alternatively, the obtained title and / or body content can also be input into a trained summarization model. The summarization model can identify important paragraphs from the body content and use them as a summary of the news article. For other types of historical content items, their topics and text can be identified, for example, using corresponding machine learning models trained in the appropriate way.

[0028] After identifying the topic and text of each historical content item in the historical content item sequence 102, the identified topics and texts can be combined separately to obtain a topic sequence and text sequence corresponding to the historical content item sequence 102. For example, topics 114-1 to 114- can be combined. Combining these into a theme sequence 114, we can arrange headings 116-1 to 116- Combine them into title sequence 116, and list abstracts 118-1 to 118- Combined into summary sequence 118. Title sequence 116 and / or summary sequence 118 can be collectively referred to as text sequence.

[0029] After obtaining topic sequence 114, comprehensive topic representation 122 can be generated using comprehensive topic representation generation unit 120. This will be discussed later in conjunction with... Figure 4 This illustrates an exemplary process for generating a comprehensive topic representation. The comprehensive topic representation 122 can be labeled as... .

[0030] Furthermore, comprehensive text representations can be generated based on text sequences. Comprehensive topic representations and comprehensive text representations can have different levels of information abstraction. For example, relatively speaking, comprehensive topic representations can represent user interests at a coarser granularity, while comprehensive text representations can represent user interests at a finer granularity. Comprehensive text attention representations 152 can be generated based on text sequences using an attention mechanism. Alternatively or additionally, comprehensive text capsule representations 162 can be generated, at least based on text sequences, using capsule networks. Comprehensive text attention representations 152 and 152 can be collectively referred to as comprehensive text representations.

[0031] A title representation sequence 132 can be generated based on the title sequence 116 using a title encoder 130. For each title in the title sequence 116, the title encoder 130 can generate a title representation for that title, thus obtaining a title representation sequence 132 corresponding to the title sequence 116. The title encoder 130 may include a word embedding layer and a convolutional neural network (CNN) layer. The word embedding layer can convert the word sequence in the title into a low-dimensional word embedding vector sequence. The sequence can be generated from the first... ( The title tags identified from the historical content items are as follows: ,in This is the title. The number of words included. This can be determined through a word embedding layer, via a word embedding lookup table. Convert the word sequence into a word embedding vector sequence ,in and These are the vocabulary size and word embedding dimension, respectively. Next, for each word in the headline, a CNN layer can be used to learn the contextual word representation of that word by capturing its local context. For each word, the context is crucial for learning its representation. For example, in the news headline "Xbox One Launches This Week," the context of the word "One," such as "Xbox" and "launch," can help it be understood as belonging to the name of a game console. The context word representation of a word can be tagged as It can be calculated using, for example, the following formula:

[0032] (1)

[0033] in, It is a non-linear activation function. It is the word sequence from position Arrive at the location Concatenation of word embedding vectors of words and These are the kernel and bias parameters of the CNN filter, respectively. It is the number of CNN filters, and It is the window size of the CNN filter.

[0034] After obtaining the contextual word representations of each word in the title, the contextual representations of all words in the title can be combined to generate the title representation. The title representation can be tagged as... The title representations of each title in title sequence 116 can be combined into title representation sequence 132.

[0035] Alternatively or additionally, a summary representation sequence 142 can be generated from the summary sequence 118 via a summary encoder 140. This can be derived from the summary sequence 118. The summary tags identified from each historical content item are as follows ,in This is a summary The number of words included. The summarizing encoder 140 can have the same structure as the title encoder 130. For example, the summarizing encoder 140 may include a word embedding layer and a CNN layer. For each summary, the summarizing encoder 140 can generate a summary representation of that summary. The summary representation can be generated through a process similar to that used to generate the title representation. The summary representation is, for example, labeled as... The summary representations of each summary in summary sequence 118 can be combined into summary representation sequence 142.

[0036] The title representation sequence 132 generated by the title encoder 130 and / or the summary representation sequence 142 generated by the summary encoder 140 can be provided to the comprehensive text attention representation generation unit 150. The comprehensive text attention representation generation unit 150 can generate the comprehensive text attention representation 152 through an attention mechanism. This will be discussed later in conjunction with... Figure 5 This illustrates an exemplary process for generating a comprehensive text attention representation. The comprehensive text attention representation 152 can be labeled as... .

[0037] Alternatively or additionally, the title representation sequence 132 and / or the summary representation sequence 142 may be provided to the comprehensive text capsule representation generation unit 160. The comprehensive text capsule representation generation unit 160 may employ a capsule network to generate the comprehensive text capsule representation 162. Preferably, when generating the comprehensive text capsule representation 162, in order to more accurately measure the user's interest in the target content item 164, a target content item representation 172 of the target content item 164 may also be considered. The target content item representation 172 may be generated by the target content item representation generation unit 170. This will be discussed later in conjunction with... Figure 6 This illustrates an exemplary process for generating a comprehensive text capsule representation. The comprehensive text capsule representation 162 can be labeled as... .

[0038] The integrated text attention representation 152 and the integrated text capsule representation 162 can have different levels of information abstraction. For example, relatively speaking, the integrated text attention representation 152 can represent user interests at a coarser granularity, while the integrated text capsule representation 162 can represent user interests at a finer granularity. The integrated text attention representation 152 and / or the integrated text capsule representation 162 can be collectively referred to as integrated text representation. A user interest representation 112 can be generated based on the integrated topic representation 122 and the integrated text representation. The user interest representation 112 can be labeled as... For example, the comprehensive topic representation 122, the comprehensive text attention representation 152, and the comprehensive text capsule representation 162 can be combined into a user interest representation 112 through the combination unit 180, as shown in the following formula:

[0039] (2)

[0040] In process 100, user interests can be represented using, for example, a comprehensive topic representation 122, a comprehensive text attention representation 152, and a comprehensive text capsule representation 162. These representations have different levels of information abstraction and model user interests at different granularities. Furthermore, process 100 considers multiple aspects of historical content items, such as topic, title, summary, and main content, which can fully reflect the information of historical content items. Therefore, the hierarchical representation learning method for user interests according to embodiments of this disclosure can effectively and comprehensively capture user interests, thereby generating accurate and rich user interest representations. Further, the generated user interest representations can be used by a recommendation system to predict the click probability of a user clicking a target content item. Accurate and rich user interest representations can help the recommendation system predict more accurate click probabilities, thereby achieving efficient and targeted content item recommendations.

[0041] It should be understood that, despite Figure 1The diagram shows that the text identified from each historical content item includes both titles and summaries; however, it is also possible to identify only one of the titles and summaries from each historical content item. Accordingly, when generating a comprehensive text representation, it is possible to base it on only one of the title sequence and summary sequence. Additionally, besides titles and / or summaries, other text from each historical content item, such as body content, can also be identified. Accordingly, when generating a comprehensive text representation, it is also possible to base it on the identified other text sequences. Furthermore, it should be understood that although... Figure 1 The diagram illustrates the generation of a user interest representation 112 based on three elements: a comprehensive topic representation 122, a comprehensive text attention representation 152, and a comprehensive text capsule representation 162. However, when generating the user interest representation 112, it is also possible to consider only one or two of the three elements: the comprehensive topic representation 122, the comprehensive text attention representation 152, and the comprehensive text capsule representation 162.

[0042] According to embodiments of this disclosure, a comprehensive topic representation can be generated by constructing a topic graph corresponding to a topic sequence. Figure 2 An exemplary topic sequence 200a and a corresponding topic diagram 200b according to an embodiment of this disclosure are shown. Topic sequence 200a may be, for example, news-related, indicating a series of topics corresponding to a series of news articles clicked by a user in succession. Topics 201 to 211 may be, in sequence, “Entertainment,” “Sports,” “Automobiles,” “Sports,” “Entertainment,” “Sports,” “Technology,” “Entertainment,” “Technology,” “Technology,” and “Technology.” A topic diagram corresponding to topic sequence 200a, such as topic diagram 200b, can be constructed.

[0043] Figure 3 An exemplary process 300 for constructing a theme graph according to an embodiment of the present disclosure is shown. Through process 300, a theme graph corresponding to a sequence of topics can be constructed.

[0044] At position 310, multiple topic categories included in the topic sequence can be identified. For example, for Figure 2 The theme sequence 200a can include four theme categories: “Entertainment”, “Sports”, “Automotive”, and “Technology”.

[0045] At point 320, multiple identified topic categories can be set as multiple nodes. For example, as... Figure 2 As shown in the theme diagram 200b, the four theme categories "Entertainment", "Sports", "Automobiles" and "Technology" can be set as nodes 250, 252, 254 and 256 respectively.

[0046] Subsequently, a set of edges between multiple nodes can be determined. For example, for every two nodes in the multiple nodes, at point 330, it can be determined whether there is a transition between the two topic categories corresponding to these two nodes, based on the topic sequence. The topic sequence corresponds to the user's click order. It can be determined whether there is a transition between two topic categories based on whether the user clicked two content items corresponding to two topic categories consecutively. For example, for nodes 250 and 252 shown in topic graph 200b, it can be determined from topic sequence 200a that there is a transition between the two topic categories corresponding to these two nodes, namely "Entertainment" and "Sports"; while for nodes 254 and 250 shown in topic graph 200b, it can be determined from topic sequence 200a that there is no transition between the two topic categories corresponding to these two nodes, namely "Automotive" and "Entertainment". In addition, the user may click two or more content items with the same topic category consecutively. For example, in topic sequence 200a, topics 209 to 211 are all "Technology", which means that the user clicked three content items with the same topic category, "Technology".

[0047] At point 340, in response to determining that a transition exists between two topic categories, the direction of the transition and the number of transitions corresponding to that direction can be determined. The direction and number of transitions can be determined based on the topic sequence. For example, for the two topic categories "Entertainment" and "Sports," from topic sequence 200a, it can be seen that the transition directions include from topic category "Entertainment" to topic category "Sports," such as from topic 201 to topic 202, and from topic 205 to topic 206. Accordingly, the number of transitions corresponding to this transition direction can be "2". Alternatively, the transition direction also includes from topic category "Sports" to topic category "Entertainment," such as from topic 204 to topic 205. Accordingly, the number of transitions corresponding to this transition direction can be "1".

[0048] At point 350, the direction and number of edges between two nodes can be determined based on the determined transformation direction and number of transformations. The direction of the edges between two nodes can be consistent with the determined transformation direction, as indicated by the arrows in topic graph 200b. The number of edges between two nodes can be consistent with the determined number of transformations. The number of each edge can be marked near that edge, as shown by the numbers near each edge in topic graph 200b. Node 256 has an edge with itself, and the number of edges is "2" because the user clicked three content items in succession, all with the topic category "Technology".

[0049] Steps 330 to 350 can be performed for every two nodes among these multiple nodes. At position 360, a set of edges between these multiple nodes can be obtained.

[0050] At position 370, these multiple nodes and the resulting set of edges can be combined to form a topic graph.

[0051] It should be understood that the above text, in combination with... Figures 2 to 3 The described process for constructing a theme graph is merely exemplary. Depending on the specific application requirements, the steps in the process for constructing a theme graph can be replaced or modified in any way, and the process may include more or fewer steps. Furthermore, the specific order or hierarchy of the steps in process 300 is merely exemplary, and the process for constructing a theme graph may be performed in an order different from that described.

[0052] Figure 4 An exemplary process 400 for generating a comprehensive topic representation according to an embodiment of the present disclosure is shown. In process 400, a comprehensive topic representation 412 can be generated by a comprehensive topic representation generation unit 410 based on a topic sequence 402. The topic sequence 402, the comprehensive topic representation generation unit 410, and the comprehensive topic representation 412 can respectively correspond to Figure 1 The process 400 includes a topic sequence 114, a comprehensive topic representation generation unit 120, and a comprehensive topic representation 122. In process 400, a topic representation sequence 422 corresponding to the topic sequence 402 can be generated, a topic graph 432 corresponding to the topic sequence 402 can be constructed, and a comprehensive topic representation 412 can be generated based on the generated topic representation sequence 422 and the constructed topic graph 432.

[0053] Topic sequence 402 may include multiple topics. Topic sequence 402 may be provided to topic encoder 420 to generate topic representation sequence 422. For each topic in topic sequence 402, topic encoder 420 may generate a topic representation for that topic, thereby obtaining topic representation sequence 422 corresponding to topic sequence 402. Topic encoder 420 may have a structure similar to title encoder 130 or summary encoder 140. The difference is that topic encoder 420 includes multilayer perceptron (MLP) layers instead of CNN layers. Specifically, since the number of words included in a topic is usually less than the number of words included in a title or summary, topic encoder 420 may use MLP layers to generate topic representations. That is, topic encoder 420 may include word embedding layers and MLP layers. The topic tags identified in each historical content item are as follows ,in The topic The number of words included. This can be determined through a word embedding layer, via a word embedding lookup table. Convert the word sequence into a word embedding vector sequence Next, the topic can be obtained through the MLP layer. The subject is represented as shown in the following formula:

[0054] (3)

[0055] in, and These are the matrix and bias parameters of the MLP layer, respectively.

[0056] After obtaining the topic representations of each topic, the topic representations of each topic in topic sequence 402 can be combined into topic representation sequence 422.

[0057] A topic graph 432 corresponding to the topic sequence 402 can be constructed using graph construction unit 430. The topic graph 432 can be composed of multiple nodes corresponding to multiple topic categories included in the topic sequence 402, and a set of edges between these nodes. This can be achieved, for example, by combining the previously mentioned methods... Figures 2 to 3 The process described is used to construct the theme graph 432.

[0058] After constructing the theme graph 432, relational information 434 can be derived from it. Relational information 434 can be used to represent the relationships between multiple nodes in the theme graph 432. Edge information between every two nodes in the theme graph can be obtained, including edge information between a node and itself. Subsequently, the number of edges associated with each node in the theme graph can be calculated, and relational information can be derived based on the edge information and the number of edges.

[0059] First, the edge information between every two nodes can be obtained from the topic graph, including the edge information between a node and itself. Edge information can be related to the direction of the edge. Nodes To the node The graph edge information can be used to represent nodes. To the node Does an edge exist between them, and if so, how many edges are there? Node To the node The graph edge information can be, for example, labeled as When node To the node When there are no edges, And when node To the node When an edge exists, ,in This is the number of edges. Return to reference. Figure 2 The graph edge information from node 254 to node 252 can be "1", and the graph edge information from node 256 to itself can be "2".

[0060] The above steps can be performed on every two nodes in the topic graph to obtain a set of graph edge information. Then, this set of graph edge information can be combined into a matrix. ,in It represents the number of nodes in the matrix. It can be called a graph adjacency matrix.

[0061] Next, the number of edges associated with each node in the topic graph can be calculated. For example, the number of edges associated with a node can be calculated by summing the graph edge information associated with each node. For example, with node... The number of associated edges can be represented as .

[0062] The above steps can be performed on each node in the topic graph to obtain the number of multiple edges. Then, the number of these multiple edges can be combined into a matrix. .matrix This can be called the graph degree matrix. For example, it can be a diagonal matrix. In this diagonal matrix, all elements except those on the main diagonal are 0.

[0063] At least the graph adjacency matrix can be And graph degree matrix This serves as relational information 434 derived from the topic graph 432. Relational information 434 can be further used to generate a comprehensive topic representation 412. In one embodiment, the comprehensive topic representation 412 can be generated based on the topic representation sequence 422 and the relational information 434.

[0064] The topic representation sequence 422 may include topic representations of each topic in the topic sequence 402. The topic graph 432 may include multiple nodes corresponding to multiple topic categories in the topic sequence 402. An initial topic graph representation 424 may be generated based on the topic representation sequence 422 and the topic graph 432. The initial topic graph representation 424 may include initial node representations of each node in the topic graph 432. The initial node representation of each node may be consistent with the topic representation of the topic category corresponding to that node. The initial topic graph representation 424 may be labeled as... ,in It is the first The node representation of a node.

[0065] Subsequently, the initial topic graph representation 424 can be updated to topic graph representation 442 based on relation information 434 using a graph attention network 440. The graph attention network 440 can have a two-layer structure. In each layer, the graph attention network 440 can aggregate the representations of each node's neighboring nodes and update the node's representation using the aggregated neighboring node representations. The process of updating the initial topic graph representation 424 to topic graph representation 442 based on relation information 434 can be represented by the following formula:

[0066] (4)

[0067] in, This represents the 440th iteration of the graph attention network. Layer output ( ), It is a non-linear activation function, and It is the first learnable The weight matrix of layer -1. In formula (4), the symbol "~" indicates a renormalization operation, such as in constructing the graph adjacency matrix. or graph degree matrix Add a self-connection to each node in the graph. This operation allows the use of each node's first self-connection. The -1 level representation is used to update the node's -1 level. Layer representation. In cases where nodes themselves are self-connected, for example... Figure 2 For node 256, it is not necessary to perform a self-join operation on this node. This represents the updated representation obtained after two rounds of convolution, which can be used as a topic graph representation 442.

[0068] After obtaining the theme image representation 442 Then, it can be represented based on the theme graph. To generate a comprehensive topic representation 412 As shown in the following formula:

[0069] (5)

[0070] It should be understood that the above text, in combination with... Figure 4The described process for generating a comprehensive topic representation is merely exemplary. Depending on the specific application requirements, the steps in the process for generating the comprehensive topic representation can be replaced or modified in any way, and the process may include more or fewer steps. Furthermore, the specific order or hierarchy of the steps in process 400 is merely exemplary, and the process for generating the comprehensive topic representation may be performed in an order different from the described order.

[0071] Figure 5 An exemplary process 500 for generating a comprehensive text attention representation according to an embodiment of the present disclosure is illustrated. In process 500, a comprehensive text attention representation 512 may be generated by a comprehensive text attention representation generation unit 510 based on a text representation sequence, such as a title representation sequence 502 and a summary representation sequence 504. The title representation sequence 502, the summary representation sequence 504, the comprehensive text attention representation generation unit 510, and the comprehensive text attention representation 512 may respectively correspond to Figure 1 The sequence includes a title representation sequence 132, a summary representation sequence 142, a comprehensive text attention representation generation unit 150, and a comprehensive text attention representation 152. In process 500, a comprehensive text attention representation 512 can be generated using an attention mechanism.

[0072] The title representation sequence 502 can include context word representations of each word in the title sequence. A comprehensive title attention representation 522 can be generated based on the title representation sequence 502 through the attention layer 520. For each word in the title sequence, an additive attention weight for that word can be calculated based on the interaction between context word representations. The additive attention weights of each word are labeled as follows: It can be calculated using, for example, the following formula:

[0073] (6)

[0074] in, and These are the first ones obtained through formula (1) above. The word and the first The context words of each word, and This is the number of words included in the title sequence. Subsequently, a comprehensive title attention representation can be generated using, for example, the following formula: 522 :

[0075] (7)

[0076] The summary representation sequence 504 may include context word representations of each word in the summary sequence. The comprehensive summary attention representation 532 can be generated in a similar manner to generating the comprehensive title attention representation 522. For example, the comprehensive summary attention representation 532 can be generated based on the summary representation sequence 504 using attention layer 530. For each word in the summary sequence, an additive attention weight for that word can be calculated based on the interaction between context word representations. The [missing information - likely a typo or incomplete sentence] can be used to... The additive attention weights of each word are labeled as follows: It can be calculated using, for example, the following formula:

[0077] (8)

[0078] in, and These are the first ones obtained through formula (1) above. The word and the first The context words of each word, and This is the number of words included in the summary sequence. Subsequently, a comprehensive summary attention representation can be generated using, for example, the following formula: 532 :

[0079] (9)

[0080] Next, the combined unit 540 can be used to base the comprehensive title attention representation 522. And comprehensive summary attention representation 532 To generate a comprehensive text attention representation 512 This can be achieved, for example, by representing comprehensive headline attention. and comprehensive summary attention representation Summation is performed to generate a comprehensive text attention representation. As shown in the following formula:

[0081] (10)

[0082] It should be understood that the above text, in combination with... Figure 5The described process for generating a comprehensive text attention representation is merely exemplary. Depending on the specific application requirements, the steps in the process for generating the comprehensive text attention representation can be replaced or modified in any way, and the process may include more or fewer steps. For example, although in process 500, a comprehensive text attention representation 512 is generated based on both the title representation sequence 502 and the summary representation sequence 504, in some embodiments, the comprehensive text attention representation 512 may be generated based on only one of the title representation sequence 502 and the summary representation sequence 504. Furthermore, in addition to the title representation sequence 502 and the summary representation sequence 504, the comprehensive text attention representation 512 may also be generated based on other text representation sequences corresponding to other text, such as the body content. Moreover, the specific order or hierarchy of the steps in process 500 is merely exemplary, and the process for generating the comprehensive text attention representation may be performed in an order different from the described order.

[0083] Figure 6 An exemplary process 600 for generating a comprehensive text capsule representation according to an embodiment of the present disclosure is illustrated. In process 600, a capsule network may be used to generate the comprehensive text capsule representation, at least based on a text representation sequence. The text representation sequence may include a title representation sequence 602 and a summary representation sequence 604. The title representation sequence 602 and the summary representation sequence 604 may respectively correspond to Figure 1 The title representation sequence 132 and the abstract representation sequence 142 are shown in the figure. A comprehensive text capsule representation 612 can be generated at least through a comprehensive text capsule representation generation unit 610. The comprehensive text capsule representation generation unit 610 and the comprehensive text capsule representation 612 can respectively correspond to... Figure 1 The integrated text capsule representation generation unit 160 and integrated text capsule representation 162 are included.

[0084] The comprehensive text capsule representation generation unit 610 may include a capsule layer 620 with dynamic routing and a label-aware attention layer 630. Through the capsule layer 620 and the label-aware attention layer 630, fine-grained user interest representations can be learned.

[0085] The title representation sequence 602 and the summary representation sequence 604 can be provided to the capsule layer 620. Within the capsule layer 620, two levels of capsules can be used: low-level capsules and high-level capsules. The goal of dynamic routing is to iteratively compute the representations of high-level capsules based on the representations of low-level capsules. This operation can be viewed as a further encapsulation of the low-level features to obtain more abstract features. The [missing information - likely a typo, should be removed]... The lower-level capsules are marked as follows: And will the first The marking of a high-grade capsule is as follows: The process of dynamic routing can be represented, for example, by the following formula:

[0086] (11)

[0087] (12)

[0088] in, This refers to the representation of lower-level capsules that need to be learned. Representation of high-grade capsules The matching matrix between them This indicates a lower grade of capsule. Representation of high-grade capsules Routing parameters between It refers to the quantity of high-grade capsules, and This is a representation of the calculated lower-level capsules. Representation of high-grade capsules The weights between them.

[0089] After dynamic routing, the updated representations of higher-level capsules can be computed. First, candidate vectors are computed, and then a squash function is applied to utilize the mold length to characterize the probabilities of higher-level features. This process can be represented by the following formula:

[0090] (13)

[0091] (14)

[0092] in, It refers to the number of low-grade capsules. This indicates the calculation of the modulus, and Indication of high-grade capsules Candidate vectors.

[0093] The above operations can be used to obtain... The representation of a high-grade capsule The interest capsule representation 622 is output by capsule layer 620. The above dynamic routing process can be regarded as a soft-clustering process, which can aggregate the user's historical interactions into several clusters. Therefore, it can be guaranteed that the representation learned by each interest capsule is as different as possible, so that different interest capsules can represent the user's different interests from different aspects.

[0094] To more accurately measure user interest in target content item 632, preferably, when generating the comprehensive text capsule representation 612, a target content item representation 642 of target content item 632 can be considered. Target content item 632 can be a content item from a set of candidate content items. Target content item 632 can have the same content as the corresponding historical content item, including, for example, news, music, movies, videos, books, product information, etc. The target content item representation 642 of target content item 632 can be generated by the target content item representation generation unit 640. Target content item 632, target content item representation generation unit 640, and target content item representation 642 can respectively correspond to... Figure 1 The target content item 164, the target content item representation generation unit 170, and the target content item representation 172 are included.

[0095] First, the text of the target content item 632 can be extracted, such as title 644, summary 646, etc. Then, a text representation of the extracted text can be generated. For example, a title representation 652 of title 644 can be generated using a title encoder 650. Alternatively or additionally, a summary representation 662 of summary 646 can be generated using a summary encoder 660. The title encoder 650 and the summary encoder 660 can have similar characteristics to... Figure 1 The structure is the same as that of the title encoder 130. Next, a target content item representation 642 can be generated based on the text representation using an attention mechanism. For example, a title attention representation 672 can be generated based on the title representation 652 using an attention layer 670. The title attention representation 672 can be labeled as... Alternatively or additionally, a summary attention representation 682 can be generated from the summary representation 662 via an attention layer 680. The summary attention representation 682 can be labeled as... The processing at attention layers 670 and 680 can be combined with... Figure 5 The processing at attention layers 520 and 530 is similar. Accordingly, the title attention representation 672 can be generated using a process similar to that used to generate the comprehensive title attention representation 522, and the summary attention representation 682 can be generated using a process similar to that used to generate the comprehensive summary attention representation 532. The title attention representation can be combined using the combination unit 690. and summary attention representation Combine them to generate the target content item representation 642. As shown in the following formula:

[0096] (15)

[0097] After generating the interest capsule representation 622 and the target content item representation 642, a comprehensive text capsule representation 612 can be generated through an annotation-aware attention layer 630. At the annotation-aware attention layer 630, the relevance between the interest capsule representation 622 and the target content item representation 642 can be obtained by calculating the likelihood value between them. In one implementation, the attention query can be the target content item representation 642, and both the key and value can be the interest capsule representation 622. The first... The attention weight of each interest capsule is labeled as It can be calculated using, for example, the following formula:

[0098] (16)

[0099] Subsequently, a comprehensive text capsule representation 612 can be generated using, for example, the following formula. :

[0100] (17)

[0101] It should be understood that the above text, in combination with... Figure 6 The described process for generating comprehensive text capsule representations is merely exemplary. Depending on the specific application requirements, the steps in the process for generating comprehensive text capsule representations can be replaced or modified in any way, and the process may include more or fewer steps. For example, although in process 600, the interest capsule representation 622 is generated based on both the title representation sequence 602 and the summary representation sequence 604, in some embodiments, the interest capsule representation 622 may be generated based on only one of the title representation sequence 602 and the summary representation sequence 604. Furthermore, in addition to the title representation sequence 602 and the summary representation sequence 604, the interest capsule representation 622 may also be generated based on other text representation sequences corresponding to other text, such as the body content. Additionally, when generating the target content item representation 642, in addition to the title 644 and the summary 646, the target content item representation 642 may also be generated based on other text, such as the body content. Furthermore, although in process 600, the comprehensive text capsule representation 612 is generated based on both the interest capsule representation 622 and the target content item representation 642, in some embodiments, the comprehensive text capsule representation 612 may be generated based solely on the interest capsule representation 622.

[0102] The above text combined Figures 1 to 6An exemplary process for generating a user interest representation according to embodiments of this disclosure is described. After generating the user interest representation, it can be further used to predict the click probability of the user clicking a target content item. The target content item can be a content item from a set of candidate content items. The target content item can have the same content as a corresponding historical content item, including, for example, news, music, movies, videos, books, product information, etc. For each content item in the set of candidate content items, the click probability of the user clicking that content item can be predicted, thereby obtaining a set of click probabilities. Based on the predicted set of click probabilities, it can be determined which content items(s) to recommend to the user. Figure 7 An exemplary process 700 for predicting click probability according to an embodiment of the present disclosure is illustrated. In process 700, a user's historical content item sequence 702 and a target content item 704 may be provided to a click probability prediction model 710. The click probability prediction model 710 may predict and output the click probability of the user clicking the target content item 704 based on the historical content item sequence 702 and the target content item 704.

[0103] Historical content item sequence 702 can correspond to Figure 1 The historical content item sequence 102 can include a set of historical content items previously clicked by the user. The historical content item sequence 702 can be provided to the user interest representation generation unit 720 in the click probability prediction model 710. The user interest representation generation unit 720 can correspond to... Figure 1 User interest representation generation unit 720 can generate user interest representation 722. .

[0104] Target content item 704 can correspond to Figure 1 The target content item 164. Target content item 704 can be provided to the target content item representation generation unit 730 in the click probability prediction model 710. The target content item representation generation unit 730 can correspond to Figure 1 The target content item in the representation of generation unit 170 and Figure 6 The target content item representation generation unit 640 in the middle. The target content item representation generation unit 730 can generate the target content item representation 732 of the target content item 704. Preferably, the user interest representation generation unit 720 may consider the target content item representation 732 when generating the user interest representation 722. To more accurately measure user interest in target content item 704.

[0105] User interest representation 722 and target content item representation 732 can be provided to prediction layer 740 in click probability prediction model 710. Prediction layer 740 can predict the click probability of a user clicking on target content item 704 based on user interest representation 722 and target content item representation 732. Click probability can be labeled as... In one implementation, the click probability can be predicted by applying the dot product method to calculate the distance between the user interest representation 722 and the target content item representation 732. As shown in the following formula:

[0106] (18)

[0107] It should be understood that the above text, in combination with... Figure 7 The described process for predicting click probabilities is merely exemplary. Depending on the specific application requirements, the steps in the process for predicting click probabilities can be replaced or modified in any way, and the process may include more or fewer steps. Furthermore, the specific order or hierarchy of the steps in process 700 is merely exemplary, and the process for predicting click probabilities can be performed in an order different from the described order.

[0108] Embodiments of this disclosure propose using a negative sampling method to train a click probability prediction model, for example... Figure 7 Click probability prediction model 710. Figure 8 An exemplary process 800 for training a click probability prediction model according to an embodiment of this disclosure is illustrated. The click probability prediction model trained by process 800, when actually deployed, can predict the click probability of a user clicking a target content item.

[0109] At point 810, a training dataset for training the click probability prediction model can be constructed. In one implementation, a list-wise strategy can be used to construct the training dataset. Taking a click probability prediction model that predicts the probability of a user clicking on a target news item as an example, the training dataset for training this model can be constructed using news items previously clicked by the user and news items not previously clicked by the user. For example, multiple news items previously clicked by the user can be considered as multiple positive samples. For each positive sample, a set of news items presented in the same session as that positive sample but not clicked by the user can be considered as a set of negative samples corresponding to that positive sample. Accordingly, the constructed training dataset can include multiple positive samples and multiple sets of negative samples corresponding to the multiple positive samples.

[0110] Subsequently, multiple posterior click probabilities corresponding to multiple positive samples can be generated. For example, at 820, the positive sample click probability corresponding to each positive sample can be predicted. The positive sample click probability corresponding to the th positive sample can be labeled as .

[0111] At 830, for each negative sample in the set of negative samples corresponding to the positive sample, the negative sample click probability corresponding to the negative sample can be predicted to obtain a set of negative sample click probabilities corresponding to the set of negative samples. The set of negative sample click probabilities corresponding to the set of negative samples of the th positive sample can be labeled as , where is the number of negative samples included in the set of negative sample click probabilities. In this way, the click probability prediction problem can be expressed as a pseudo +1-way classification task.

[0112] At 840, the posterior click probability corresponding to the positive sample can be calculated based on the positive sample click probability and the set of negative sample click probabilities. The posterior click probability corresponding to the th positive sample can be labeled as . In one implementation, the posterior click probability corresponding to the positive sample can be calculated by normalizing the positive sample click probability and the set of negative sample click probabilities using the softmax function, as shown in the following formula:

[0113] (19)

[0114] The operations at step 820 to step 840 can be performed for each positive sample in the set of multiple positive samples in the training data set. Thus, at 850, multiple posterior click probabilities corresponding to multiple positive samples can be obtained.

[0115] At 860, a prediction loss can be generated based on the multiple posterior click probabilities. In one implementation, the prediction loss can be generated by calculating the negative log-likelihood of the multiple posterior click probabilities, as shown in the following formula:

[0116] (20)

[0117] where, is the set of positive samples composed of multiple positive samples.

[0118] At 870, the click probability prediction model can be optimized by minimizing the prediction loss.

[0119] It should be understood that the above text, in combination with... Figure 8 The described process for training a click probability prediction model is merely exemplary. Depending on the specific application requirements, the steps in the process for training the click probability prediction model can be replaced or modified in any way, and the process may include more or fewer steps. Furthermore, the specific order or hierarchy of the steps in process 800 is merely exemplary, and the process for training the click probability prediction model may be performed in an order different from the described order.

[0120] Figure 9 This is a flowchart of an exemplary method 900 for learning hierarchical representations of user interests according to embodiments of the present disclosure.

[0121] At position 910, the user's historical content item sequence can be obtained.

[0122] At position 920, the topic and text of each historical content item in the historical content item sequence can be identified to obtain a topic sequence and text sequence corresponding to the historical content item sequence.

[0123] At position 930, a comprehensive topic representation can be generated based on the topic sequence.

[0124] At position 940, a comprehensive text representation can be generated based on the text sequence.

[0125] At point 950, the user's interest representation can be generated based on the comprehensive topic representation and the comprehensive text representation.

[0126] In one implementation, the integrated topic representation and the integrated text representation may have different information abstraction levels.

[0127] In one implementation, generating a comprehensive topic representation may include: generating a topic representation sequence corresponding to the topic sequence; constructing a topic graph corresponding to the topic sequence; and generating the comprehensive topic representation based on the topic representation sequence and the topic graph.

[0128] The construction of the theme graph may include: determining multiple theme categories included in the theme sequence; setting the multiple theme categories as multiple nodes; determining a set of edges between the multiple nodes; and combining the multiple nodes and the set of edges to form the theme graph.

[0129] The determination of a set of edges may include, for every two nodes among the plurality of nodes: determining, based on the topic sequence, whether there is a transition between the two topic categories corresponding to the two nodes; in response to determining that there is a transition between the two topic categories, determining the transition direction and the number of transitions corresponding to the transition direction; and based on the determined transition direction and the number of transitions, determining the direction and number of edges between the two nodes.

[0130] The process of generating the comprehensive topic representation may include: deriving relational information representing the relationships between multiple nodes in the topic graph from the topic graph; and generating the comprehensive topic representation based on the topic representation sequence and the relational information.

[0131] The derived relationship information may include: obtaining graph edge information between every two nodes in the theme graph; calculating the number of edges associated with each node in the theme graph; and deriving the relationship information based on the graph edge information and the number of edges.

[0132] In one implementation, the text may include at least one of a title, a summary, and body content. The text sequence may include at least one of a title sequence, a summary sequence, and a body content sequence.

[0133] In one implementation, generating a comprehensive text representation may include: generating a comprehensive text attention representation based on the text sequence using an attention mechanism. Generating a user interest representation may include: generating the user interest representation based on the comprehensive topic representation and the comprehensive text attention representation.

[0134] In one implementation, generating a comprehensive text representation may include: generating a comprehensive text capsule representation using a capsule network, at least based on the text sequence. Generating a user interest representation may include: generating the user interest representation based on the comprehensive topic representation and the comprehensive text capsule representation.

[0135] In one implementation, generating a comprehensive text representation may include: generating a comprehensive text attention representation based on the text sequence using an attention mechanism; and generating a comprehensive text capsule representation using a capsule network, at least based on the text sequence. Generating a user interest representation may include: generating the user interest representation based on the comprehensive topic representation, the comprehensive text attention representation, and the comprehensive text capsule representation.

[0136] The integrated text attention representation and the integrated text capsule representation can have different levels of information abstraction.

[0137] The generation of the comprehensive text capsule representation may include: generating interest capsule representations based on the text sequence using the capsule network; generating target content item representations for target content items; and generating the comprehensive text capsule representations based on the interest capsule representations and the target content item representations through an attention mechanism.

[0138] The generation of the target content item representation may include: extracting the text of the target content item; generating a text representation of the text; and generating the target content item representation based on the text representation using an attention mechanism.

[0139] In one implementation, method 900 may further include: predicting the click probability of the user clicking the target content item based on the user interest representation and the target content item representation of the target content item.

[0140] The click probability can be output by a click probability prediction model. Training the click probability prediction model may include: constructing a training dataset, which includes multiple positive samples and multiple negative sample sets corresponding to the multiple positive samples; generating multiple posterior click probabilities corresponding to the multiple positive samples; generating a prediction loss based on the multiple posterior click probabilities; and optimizing the click probability prediction model by minimizing the prediction loss.

[0141] The generation of multiple posterior click probabilities may include, for each positive sample: predicting the positive sample click probability corresponding to the positive sample; for each negative sample in the set of negative samples corresponding to the positive sample, predicting the negative sample click probability corresponding to the negative sample, to obtain a set of negative sample click probabilities corresponding to the set of negative samples; and calculating the posterior click probability corresponding to the positive sample based on the positive sample click probability and the set of negative sample click probabilities.

[0142] In one implementation, the historical content item or the target content item may include at least one of news, music, movies, videos, books, and product information.

[0143] It should be understood that method 900 may also include any step / process for learning hierarchical representations of user interests according to the embodiments of this disclosure described above.

[0144] Figure 10 An exemplary apparatus 1000 for hierarchical representation learning of user interests according to an embodiment of the present disclosure is shown.

[0145] The apparatus 1000 may include: a historical content item sequence acquisition module 1010, for acquiring a user's historical content item sequence; a topic sequence and text sequence acquisition module 1020, for identifying the topic and text of each historical content item in the historical content item sequence to obtain a topic sequence and text sequence corresponding to the historical content item sequence; a comprehensive topic representation generation module 1030, for generating a comprehensive topic representation based on the topic sequence; a comprehensive text representation generation module 1040, for generating a comprehensive text representation based on the text sequence; and an interest representation generation module 1050, for generating the user's user interest representation based on the comprehensive topic representation and the comprehensive text representation. Furthermore, the apparatus 1000 may also include any other modules configured for hierarchical representation learning of user interests according to the embodiments of the present disclosure described above.

[0146] Figure 11 An exemplary apparatus 1100 for hierarchical representation learning of user interests according to an embodiment of the present disclosure is shown.

[0147] The apparatus 1100 may include: at least one processor 1110; and a memory 1120 storing computer-executable instructions. When executed, the computer-executable instructions cause the at least one processor 1110 to: obtain a sequence of historical content items for a user; identify the topic and text of each historical content item in the sequence to obtain a topic sequence and a text sequence corresponding to the historical content item sequence; generate a comprehensive topic representation based on the topic sequence; generate a comprehensive text representation based on the text sequence; and generate a user interest representation for the user based on the comprehensive topic representation and the comprehensive text representation.

[0148] It should be understood that the processor 1110 may also perform any other steps / processes of the method for hierarchical representation learning of user interests according to the above embodiments of this disclosure.

[0149] Embodiments of this disclosure propose a computer program product for hierarchical representation learning of user interests, comprising a computer program executed by at least one processor to: obtain a sequence of historical content items for a user; identify the topic and text of each historical content item in the sequence of historical content items to obtain a topic sequence and a text sequence corresponding to the sequence of historical content items; generate a comprehensive topic representation based on the topic sequence; generate a comprehensive text representation based on the text sequence; and generate a user interest representation for the user based on the comprehensive topic representation and the comprehensive text representation. Furthermore, the computer program may also be executed to implement any other steps / processes for implementing the method for hierarchical representation learning of user interests according to the embodiments of this disclosure described above.

[0150] Embodiments of this disclosure may be embodied in a non-transitory computer-readable medium. The non-transitory computer-readable medium may include instructions that, when executed, cause one or more processors to perform any operation of the method for hierarchical representation learning of user interests according to embodiments of this disclosure as described above.

[0151] It should be understood that all operations in the methods described above are merely exemplary, and this disclosure is not limited to any operation in the methods or the order of such operations, but should cover all other equivalent variations under the same or similar concept. Furthermore, unless otherwise specified or clearly indicated from the context as relating to the singular form, the articles “a” and “an” as used in this specification and the appended claims should generally be interpreted as meaning “one” or “one or more”.

[0152] It should also be understood that all modules in the apparatus described above can be implemented in various ways. These modules can be implemented as hardware, software, or a combination thereof. Furthermore, any of these modules can be further functionally divided into sub-modules or combined together.

[0153] Processors have been described in conjunction with various devices and methods. These processors can be implemented using electronic hardware, computer software, or any combination thereof. Whether these processors are implemented as hardware or software will depend on the specific application and the overall design constraints imposed on the system. As an example, the processors, any portions of processors, or any combination of processors given in this disclosure can be implemented using microprocessors, microcontrollers, digital signal processors (DSPs), field-programmable gate arrays (FPGAs), programmable logic devices (PLDs), state machines, gated logic units, discrete hardware circuits, and other suitable processing components configured to perform the various functions described in this disclosure. The functionality of the processors, any portions of processors, or any combination of processors given in this disclosure can be implemented using software executed by a microprocessor, microcontroller, DSP, or other suitable platform.

[0154] Software should be broadly considered to refer to instructions, instruction sets, code, code segments, program code, programs, subroutines, software modules, applications, software applications, software packages, routines, subroutines, objects, running threads, procedures, functions, etc. Software may reside on a computer-readable medium. Computer-readable media may include, for example, memory, which may be, for example, magnetic storage devices (e.g., hard disks, floppy disks, magnetic stripes), optical disks, smart cards, flash memory devices, random access memory (RAM), read-only memory (ROM), programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), registers, or removable disks. Although memory is shown as separate from the processor in several aspects set forth in this disclosure, memory may also reside within the processor, such as in caches or registers.

[0155] The above description is provided to enable any person skilled in the art to practice the various aspects described herein. Various modifications to these aspects will be apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects. Therefore, the claims are not intended to be limited to the aspects shown herein. All structural and functional equivalents of the elements of the various aspects described herein, known or likely to be known by one of ordinary skill in the art, are expressly incorporated herein and covered by the claims.

Claims

1. A method for learning hierarchical representations of user interests, comprising: Obtain the user's historical content item sequence; Identify the topic and text of each historical content item in the historical content item sequence to obtain a topic sequence and text sequence corresponding to the historical content item sequence; Generate a comprehensive topic representation based on the topic sequence, wherein generating the comprehensive topic representation includes: Generate a topic representation sequence corresponding to the topic sequence; Construct a theme graph corresponding to the theme sequence, where nodes in the theme graph represent theme categories in the theme sequence, and edges in the theme graph represent the order in which different content items with different theme categories are clicked; Relationship information representing the relationships between multiple nodes in the theme graph is derived from the theme graph. This relationship information includes: information indicating whether an edge exists between any two nodes and, if so, the number of edges; and information indicating the number of edges associated with each node. The comprehensive topic representation is generated based on the topic representation sequence and the relation information; Generate a comprehensive text representation based on the text sequence; and The user's interest representation is generated based on the comprehensive topic representation and the comprehensive text representation.

2. The method according to claim 1, wherein, The comprehensive topic representation and the comprehensive text representation have different levels of information abstraction.

3. The method according to claim 1, wherein, Building a theme graph includes: Identify the multiple topic categories included in the topic sequence; Set the multiple topic categories into multiple nodes; Determine a set of edges between the plurality of nodes; and The plurality of nodes and the set of edges are combined to form the theme graph.

4. The method according to claim 3, wherein, Determining a set of edges includes, for every two nodes among the plurality of nodes: Based on the topic sequence, determine whether there is a transition between the two topic categories corresponding to the two nodes; In response to determining that a transition exists between the two topic categories, the direction of the transition and the number of transitions corresponding to the direction of the transition are determined; as well as Based on the determined transformation direction and number of transformations, the direction and number of edges between the two nodes are determined.

5. The method according to claim 1, wherein, The exported relationship information includes: Obtain the edge information between every two nodes in the topic graph; Calculate the number of edges associated with each node in the topic graph; and The relationship information is derived based on the graph edge information and the quantity.

6. The method according to claim 1, wherein, The text includes at least one of a title, a summary, and a body content, and the text sequence includes at least one of a title sequence, a summary sequence, and a body content sequence.

7. The method according to claim 1, wherein, The generated comprehensive text representation includes: Based on the text sequence, a comprehensive text attention representation is generated using an attention mechanism, and Generating user interest representations includes: The user interest representation is generated based on the comprehensive topic representation and the comprehensive text attention representation.

8. The method according to claim 1, wherein, The generated comprehensive text representation includes: Based at least on the text sequence, a capsule network is used to generate a comprehensive text capsule representation, and Generating user interest representations includes: The user interest representation is generated based on the comprehensive topic representation and the comprehensive text capsule representation.

9. The method according to claim 1, wherein, The generated comprehensive text representation includes: Based on the text sequence, a comprehensive text attention representation is generated using an attention mechanism; and Based at least on the text sequence, a capsule network is used to generate a comprehensive text capsule representation, and Generating user interest representations includes: The user interest representation is generated based on the comprehensive topic representation, the comprehensive text attention representation, and the comprehensive text capsule representation.

10. The method according to claim 9, wherein, The integrated text attention representation and the integrated text capsule representation have different information abstraction levels.

11. The method according to claim 8 or 9, wherein, The generated integrated text capsule representation includes: Based on the text sequence, the capsule network is used to generate interest capsule representations; The target content item representation of the generated target content item; and The comprehensive text capsule representation is generated based on the interest capsule representation and the target content item representation through an attention mechanism.

12. The method according to claim 11, wherein, The generated target content item includes: Extract the text of the target content item; The text representation of the generated text; and Based on the text representation, the target content item representation is generated through an attention mechanism.

13. The method according to claim 1, further comprising: The probability of a user clicking on a target content item is predicted based on the user interest representation and the target content item representation.

14. The method according to claim 13, wherein, The click probability is output through a click probability prediction model, and the training of the click probability prediction model includes: Construct a training dataset, which includes multiple positive samples and multiple sets of negative samples corresponding to the multiple positive samples; Generate multiple posterior click probabilities corresponding to the multiple positive samples; A prediction loss is generated based on the multiple posterior click probabilities; and The click probability prediction model is optimized by minimizing the prediction loss.

15. The method according to claim 14, wherein, The generation of multiple posterior click probabilities includes, for each positive sample: Predict the positive sample click probability corresponding to the positive sample; For each negative sample in the set of negative samples corresponding to the positive sample, predict the negative sample click probability corresponding to the negative sample to obtain the set of negative sample click probabilities corresponding to the set of negative samples; as well as The posterior click probability corresponding to the positive sample is calculated based on the set of positive sample click probabilities and negative sample click probabilities.

16. The method according to claim 11, wherein, The historical content item or the target content item includes at least one of news, music, movies, videos, books, and product information.

17. An apparatus for learning hierarchical representations of user interests, comprising: At least one processor; as well as A memory storing computer-executable instructions, which, when executed, cause the at least one processor to: Obtain the user's historical content item sequence. Identify the topic and text of each historical content item in the historical content item sequence to obtain a topic sequence and text sequence corresponding to the historical content item sequence. Generate a comprehensive topic representation based on the topic sequence, wherein generating the comprehensive topic representation includes: Generate a topic representation sequence corresponding to the topic sequence. Construct a theme graph corresponding to the theme sequence, where nodes in the theme graph represent theme categories in the theme sequence, and edges in the theme graph represent the order in which different content items with different theme categories are clicked. Relationship information representing the relationships between multiple nodes in the theme graph is derived from the theme graph. This relationship information includes: information indicating whether an edge exists between any two nodes and, if so, the number of edges; and information indicating the number of edges associated with each node. The comprehensive topic representation is generated based on the topic representation sequence and the relation information. Generate a comprehensive text representation based on the text sequence, and The user's interest representation is generated based on the comprehensive topic representation and the comprehensive text representation.

18. A computer program product for learning hierarchical representations of user interests, comprising a computer program executed by at least one processor for: Obtain the user's historical content item sequence; Identify the topic and text of each historical content item in the historical content item sequence to obtain a topic sequence and text sequence corresponding to the historical content item sequence; A comprehensive topic representation is generated based on the aforementioned topic sequence, wherein... The generated comprehensive topic representation includes: Generate a topic representation sequence corresponding to the topic sequence; Construct a theme graph corresponding to the theme sequence, where nodes in the theme graph represent theme categories in the theme sequence, and edges in the theme graph represent the order in which different content items with different theme categories are clicked; Relationship information representing the relationships between multiple nodes in the theme graph is derived from the theme graph. This relationship information includes: information indicating whether an edge exists between any two nodes and, if so, the number of edges; and information indicating the number of edges associated with each node. The comprehensive topic representation is generated based on the topic representation sequence and the relation information; Generate a comprehensive text representation based on the text sequence; and The user's interest representation is generated based on the comprehensive topic representation and the comprehensive text representation.

Images