A context-aware based document recommendation method
By employing a context-aware literature recommendation method, and utilizing deep learning recurrent neural networks and improved graph neural networks, a dynamic interest model is constructed. This solves the problem of inaccurate recommendations caused by neglecting user context in existing technologies, and enables real-time identification and personalized recommendations of user interests.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- INST OF SCI & TECHN INFORMATION OF CHINA
- Filing Date
- 2023-12-07
- Publication Date
- 2026-07-21
AI Technical Summary
Existing literature recommendation technologies ignore user scenarios, resulting in an inability to capture users' real-time information and true needs, leading to problems such as biased understanding of interests and information cocoons.
We adopt a context-aware literature recommendation method, which constructs a dynamic interest model by analyzing user search behavior, and uses deep learning recurrent neural networks and improved graph neural networks to integrate user context information for real-time recommendation.
It enables real-time identification and accurate recommendation of user search intent, avoiding the information cocoon phenomenon and improving the personalization and accuracy of recommendations.
Smart Images
Figure CN117931986B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a real-time document recommendation method, specifically a context-aware document recommendation method, belonging to the field of big data mining and application technology. Background Technology
[0002] The rapid development of internet technology has led to severe information overload. Recommendation technology is a common and effective means of solving this problem, and it has been widely and successfully applied in fields such as news and social media.
[0003] In recent years, with the surge in digital literature resources, users' demand for personalized recommendations of literature has become increasingly strong, prompting numerous studies by researchers both domestically and internationally. In applied research, related studies mainly revolve around content-based recommendation, collaborative filtering recommendation, and deep learning recommendation. Content-based literature recommendation methods focus on the characteristics of literature content, discovering similar literature by calculating the similarity between users and literature, and between different literatures, and recommending literature that aligns with user interests. Whether analyzing citation relationships between papers, collaboration relationships between authors, or recommendations based on article topics, all rely on article content for feature extraction and paper scoring. Collaborative filtering algorithms primarily focus on the similarity of feature relationships between users and are widely used in literature recommendation, calculating the relevance between users and documents they haven't interacted with before and then filtering recommendations. Deep learning methods can be used to analyze the content of literature, including titles, abstracts, and full texts, to extract semantic and thematic features. By constructing deep learning models, relevant literature can be recommended to users based on their interests and historical behavior.
[0004] Most existing document recommendation technologies rely on the binary relationship between users and documents for recommendation, that is, by calculating feature similarity, including feature relationships between users and documents, between documents, and between users. This provides a coarse-grained analysis of existing static user behavior data, lacks attention to user scenarios, and ignores the impact of contextual factors on user behavior. As a result, it fails to capture users' real-time information and true needs, leading to various problems such as misunderstanding of user interests and information cocoons. Summary of the Invention
[0005] The purpose of this invention is to address the problems and shortcomings of existing technologies. In order to solve the technical problems in the field of literature that rely solely on static data for analysis, neglecting the attention to user scenarios, failing to capture users' real-time information and true needs, and resulting in inaccurate literature recommendations, this invention creatively proposes a context-aware literature recommendation method.
[0006] Because scenarios are complex and multifaceted, with different definitions across different fields, they need to be understood in conjunction with application scenarios to fully explore the connection between scenarios and applications. Context awareness refers to using information sensing, collection, and processing technologies to enable computers and other devices to understand their own and their users' current situation as accurately as possible, in order to better serve users' changing needs in different environments. Its basic idea is to describe and model the acquired user contextual information, and understand the user's behavioral motivations through modeling. Context-aware recommendation technology mainly studies how to integrate contextual information into the recommendation process. Context-aware recommendation systems that incorporate contextual information can integrate data representing user personalities, which can improve the accuracy of recommendation results and enhance the personalization of recommendations.
[0007] Therefore, based on fine-grained analysis of user behavior, this invention proposes a dynamic real-time document recommendation method that integrates user context awareness. This method begins with a user's document retrieval and, through analysis of user retrieval behavior, forms a temporally related sequence of retrieval behaviors, achieving real-time identification of user retrieval interests and intentions. Simultaneously, considering the complex relationship between the current context and user interests, an improved graph neural network method is employed to achieve document recommendation, exploring a novel approach to document recommendation.
[0008] First, we obtain the behavior logs of scientific research users on the literature retrieval platform and analyze the log data to obtain the basic retrieval steps of the users.
[0009] Then, the conversation flow of a single user is segmented according to the basic steps of retrieval, the scenarios that need to be activated for recommendation are selected, and the scenarios are defined and characterized.
[0010] Finally, deep learning recurrent neural networks, which are adept at processing temporal information, are used to process the user's temporal information, obtain the user's interests, and make recommendations based on the user's interests.
[0011] Beneficial effects
[0012] Compared with existing technologies, this method has the following advantages:
[0013] 1. This method incorporates user search context information, which can fully obtain the user's true search intent, making it easier to make accurate recommendations to the user.
[0014] 2. This method utilizes users' real-time search behavior information to perceive users' search intent in real time and helps to avoid the occurrence of information cocoon phenomenon.
[0015] 3. This method analyzes a large number of user search logs to determine and characterize the context of user search behavior.
[0016] 4. This method utilizes an improved graph neural network approach, enabling more real-time and accurate recommendations. Attached Figure Description
[0017] Figure 1 This is a schematic diagram of a context-based user dynamic interest model.
[0018] Figure 2 This is a schematic diagram of the general user retrieval process based on the NSTL platform;
[0019] Figure 3 A schematic diagram of the user retrieval session flow framework;
[0020] Figure 4 This is a schematic diagram of the complete behavior sequence of a portion of NSTL users;
[0021] Figure 5 Here is pseudocode for cell session segmentation based on Distilled BERT;
[0022] Figure 6 This is a schematic diagram of a user interest model based on a recurrent neural network.
[0023] Figure 7 This is a schematic diagram of a literature recommendation model based on an improved graph neural network. Detailed Implementation
[0024] The present invention will now be described in further detail with reference to the accompanying drawings.
[0025] A context-aware literature recommendation method is proposed, which constructs a dynamic user interest model based on context characterization.
[0026] The basic idea of context awareness is to describe and model the acquired contextual information, understand the user's behavioral motivation, and proactively push relevant content. In the context of literature retrieval applications, this invention focuses on the user's preferences for literature content shown during the clicking process, based on the user's basic behavioral patterns, to identify the user's context from the user's click behavior characteristics and make recommendations.
[0027] like Figure 1 As shown, this invention presents a user dynamic interest model based on user retrieval behavior sequences.
[0028] First, we obtain the behavioral logs of research users on the literature retrieval platform and analyze the log data to determine the basic retrieval steps taken by the users, such as... Figure 2 As shown.
[0029] Then, the conversation flow of a single user is segmented according to the basic steps of retrieval, the scenarios that need to be activated for recommendation are selected, and the scenarios are defined and characterized.
[0030] Finally, deep learning recurrent neural networks, which are adept at processing temporal information, are used to process the user's temporal information, obtain the user's interests, and make recommendations based on the user's interests.
[0031] Specifically, this method includes the following steps:
[0032] Step 1: User behavior log analysis.
[0033] During the document retrieval process, users typically interact sequentially with multiple modules across different interfaces on a single platform, generating a large amount of behavioral information. Therefore, analyzing user retrieval behavior essentially involves analyzing the patterns and personalized habits exhibited by users when using a document retrieval platform, thereby discovering and uncovering users' potential needs and interests.
[0034] Through extensive user behavior analysis, user behaviors during the information retrieval process are summarized into three key categories: querying, navigation, and location. Querying refers to users entering search terms or clicking on the platform after entering it, with the platform generating a result set based on the user's input or clicks. Navigation refers to users performing a search on the literature retrieval platform, with the platform returning a set of results. The first page of the search results displays only the most relevant documents and provides a pagination function. Location refers to users clicking on a specific document on the platform to view its details in the returned list of documents, or clicking on a specific reference on the document details page.
[0035] Furthermore, the "purchase and add," forwarding, favorite, and related recommendations" functions provided by literature retrieval platforms also fall under the category of location-based behaviors. In these contexts, location information represents the user's interests and ultimate goals.
[0036] By analyzing user search behavior, specific user search steps can be derived. For example... Figure 2 As shown.
[0037] Step 2: Construct the user session data stream.
[0038] User search information refers to the information from when a user enters the platform until they leave. Since the data in the original search logs is a series of clicks, it is necessary to analyze the user's sequential clickstream to obtain the click relationships between the user and specific information.
[0039] Statistical user sequential clickstream such as Figure 3 As shown, the user clickstream generally includes the following steps:
[0040] First, the unified identifier sessionid of the logs is used to match each specific behavior of the same user, and the user's search clicks at that time are concatenated according to the time sequence. For all the click information of the concatenated user, it is sorted by time, and the query behavior, navigation behavior and the location behavior determined by navigation are distinguished. The order of each query behavior, navigation behavior and location behavior is determined according to the time sequence, and finally the user's click stream is generated completely.
[0041] Through the above operations, the user retrieval process was reconstructed, completing the first step of organizing user log data. The extracted partial complete user behavior sequences, such as... Figure 4 As shown.
[0042] Step 3: After extracting the complete user behavior sequence, obtain the conversation stream, and determine the context based on the conversation stream to obtain user interests.
[0043] User interests are revealed by the topics a user discusses throughout the entire conversation flow. However, topics can vary within the same conversation, leading to topic shifts. Therefore, it's necessary to perform pairwise calculations on topic features in chronological order to obtain the similarity between consecutive content feature records.
[0044] Specifically, a similarity threshold is set to determine whether two consecutive content feature records in the same session belong to the same topic. If the similarity is higher than the set threshold, the consecutive actions are determined to belong to the same topic; conversely, if the similarity is lower than the set threshold, the consecutive actions are determined not to belong to the same topic, and the session is split between the two content feature records to form corresponding unit sessions.
[0045] After calculating and splitting all conversations, a dataset of unit conversation streams with similar topics is ultimately formed. The conversation streams are segmented according to different contexts, and the similarity of different search terms is compared. For example, the DistilledBERT model can be used for similarity comparison. This model performs well on some natural language processing tasks; it maps each word to a fixed-length vector, capturing the semantic and contextual relationships between words. The pseudocode for unit conversation segmentation based on the DistilledBERT model is as follows: Figure 5 As shown.
[0046] Step 4: Determine and depict the scenario.
[0047] After determining the session flow, the scenario during the retrieval process is identified by analyzing the acquired user log data, triggering subsequent recommendations.
[0048] During the user's search phase, information such as the selected search method, search terms, document type, author, institution, keywords, subject terms, publication time, and source (journal terminology, year, volume, and issue) is included. It also includes information on the specific documents that the user has clicked to view. This information can fully reflect the dynamic process of a user's search at that time.
[0049] Therefore, the data dimensions in user logs can be summarized into the following scenarios:
[0050] The unique identifier of a unit session: The original session is constructed using the session ID from the user information. A unit session is a research unit after the original session has been split into topics, and a unique identifier is assigned to this unit session.
[0051] Unique identifier for behavior: A behavior refers to a behavior record within a unit session. Each behavior record has a corresponding identifier, which is sorted by access time and then simply assigned a value.
[0052] Behavior types: including three types of behavior: query, navigation, and location;
[0053] Content features include document type, search terms, logical operators, matching method (exact or fuzzy), and search fields;
[0054] Behavioral characteristics: Features beyond content preferences exhibited in user behavior are meaningful for understanding user intent; these characteristics are collectively referred to as behavioral characteristics. For example, a user's choice of sorting order—relevance sorting, time descending order, citation count descending order, and NSTL metric descending order—represents four different areas of interest: topic, novelty, classic, and popularity.
[0055] Professionalism: This refers to whether the user is clear about their search intent. User query behavior includes quick search and advanced search. In the context characterization process, different weight parameters should be set according to the differences in search methods. If two search methods exist simultaneously, the one with the highest intensity should be selected to represent the user's behavioral intent.
[0056] Attention level: The degree of user attention to the current page can be reflected by the duration of page dwell. The duration of dwell on the current page can be used to determine the validity of actions taken from the source page.
[0057] Step 5: Construct a multi-layer recurrent graph neural network (RGNN) interest extraction model.
[0058] After determining, defining, and characterizing the search scenario, it is necessary to determine the user's search interests. Since users exhibit dynamic interests during the browsing process, their real-time interests change, and their interest levels change accordingly. Therefore, in the document recommendation model, this invention establishes a dynamic interest model for users based on their dynamic interests. This model can reflect user preferences in real time and effectively avoid the information cocoon problem.
[0059] Specifically, a recurrent graph neural network, which can effectively process temporal information, is used to obtain user interests and preferences. From Figure 6 As can be seen, the processed user behavior information and conversation flow are arranged in a temporal topology to form a graph. In this graph, nodes represent different context nodes, and the connections between nodes represent the contexts involved in the current user interaction. The graph's topology matrix is then input into a recurrent graph neural network (RGNN) to infer the user's interests and preferences. Specifically, word embedding is performed on the original conversation flow data, and the resulting topology vector is input into the RGNN layer to obtain the input information x. t The matrix is used for gated cyclic unit calculations. During the calculation process, the matrix first passes through the reset gate r. t and update gate z t Perform linear processing on the original data and reset W in the gate. r and U r With the W in the update gate z and U z This is a randomly generated initial weight matrix. Using the above weight matrix, the input information x is weighted respectively. t and the hidden state h from the previous moment t-1 The calculated matrix undergoes a linear transformation. Finally, the user's search behavior and the context in which they are located are input into the RGNN model to obtain the user's interests in different contexts.
[0060] Step 6: Construct a context-triggered graph neural network recommendation model.
[0061] In the process of literature recommendation, it is necessary to consider both the interests and preferences of research users and the context in which they are situated. How to extract and analyze the complex features of this information and then recommend literature is a crucial problem to be solved in this field. Currently, graph neural networks have shown great potential in the field of personalized recommendation, as they can express the similarity of nodes through the information flow on a graph. However, their drawbacks include high algorithmic complexity.
[0062] This method requires extracting complex features such as research users' preferences and current context during the literature recommendation process. Therefore, a graph neural network (Graph Neural Network) is used to construct the recommendation model, and improvements are made to address its high complexity. The improved Graph Neural Network-based recommendation model is as follows: Figure 7 As shown.
[0063] from Figure 7 As can be seen, the recommendation model proposed in this method includes a topology graph, a topology relation matrix, Markov chain sampling, multi-level feature dimensionality reduction and convolution, fully connected components, and recommendation output.
[0064] The topology graph includes three scenarios: retrieval, location, and access. Different nodes are linked to relationships. After inputting the topology graph into the topology matrix, the Gibbs sampling algorithm is used to sample nodes in each convolutional layer. Linear discriminant analysis (LDA) is used for multi-level dimensionality reduction, feature reduction, and convolution. The resulting features are then input into a fully connected layer, and finally, the recommendation results are output.
[0065] To avoid the recommendation principle of direct reasoning, this invention improves the graph neural network through inductive reasoning. The inductive evolution is shown in Equation 1:
[0066]
[0067] l = 0, 1, ..., M-1
[0068] Where L represents the output of the convolutional layer, m represents the layer number, t represents time, v represents the vertex set, H represents the node set of each layer, A represents the adjacency matrix, σ represents the non-linear transformation, i.e., the activation function, q represents intermediate layer nodes, l represents the current layer, and M is the layer number. This indicates the last level where the current node j is located. This indicates the layer l where node j is located.
[0069] The obtained loss function L b As shown in Equation 2:
[0070]
[0071] Where L is the output of the convolutional layer, and M is the number of vertices. represents the layer number of node i, and g represents the activation function.
[0072] By deriving Equations 1 and 2 from the input data, the final recommendation result is obtained.
[0073] The steps of this recommendation model include:
[0074] Step 1: Construct the topology graph.
[0075] Step 2: Construct the topological relation matrix.
[0076] Step 3: Gibbs-based sampling.
[0077] Step 4: Multi-stage feature dimensionality reduction and convolution.
[0078] Step 5: Obtain user intent features and output recommendation results.
[0079] Various deep learning methods, such as graph neural networks, construct multi-layered networks to represent recommendation goals at multiple levels and represent word order features of the data through multiple high-level features. Therefore, in the process of multi-layer convolution, to avoid high dependence on the data, it is necessary to increase the data dimensionality. Excessive data dimensionality can lead to the curse of dimensionality, affecting recommendation performance. How to reduce feature dimensionality is one of the important problems in deep learning research. Dimensionality reduction methods can transform multiple different features into several comprehensive criteria. This method uses multi-level dimensionality reduction and output, specifically employing a method based on linear discriminant analysis (LDA) for dimensionality reduction.
[0080] First, the input topological matrix undergoes a first-level dimensionality reduction, which involves sampling the matrix, calculating the mean, and multiplying it with its transpose to obtain the dimensionality-reduced features, which are then fed into a convolutional layer for convolution. The convolutional features are then subjected to a second-level dimensionality reduction, involving sampling, calculating the mean, and multiplying with the transpose. The features obtained after this second dimensionality reduction are then fed into a second convolutional layer for further convolution. This process is repeated until a predetermined threshold is met. Finally, the resulting dimensionality-reduced features are input into a fully connected matrix.
Claims
1. A context-aware document recommendation method, characterized in that, This includes building a dynamic user interest model based on contextual characterization; First, we obtained the behavior logs of scientific research users on the literature retrieval platform and analyzed the log data to obtain the basic retrieval steps of the users. Then, construct the user session data stream, segment the session flow of a single user according to the basic steps of retrieval, select the scenarios that need to be activated for recommendation, and define and characterize the scenarios; Finally, deep learning recurrent neural networks are used to process users' time-series information to obtain users' interests, and recommendations are made based on users' interests. The interest extraction model, constructed using a multi-layer recurrent graph neural network (RGNN), is as follows: The processed user behavior information and session flow are arranged in a time-series graph topology, where the nodes of the graph represent different scenario nodes, and the connections between nodes represent the scenarios involved in this user. The topology matrix of the graph is input into a recurrent graph neural network (RGNN) to infer user interests and preferences. Specifically, word embeddings are performed on the original conversation stream data, and the resulting topology vector is input into the RGNN layer to obtain the input information. The matrix is used for gated loop unit calculations. During the calculation process, it first passes through the reset gate. and Update Gate Perform linear processing on the original data and reset the gate. and With the update gate and This is a randomly generated initial weight matrix; using the above weight matrix, the input information is respectively... Hidden state from the previous moment The calculated matrix undergoes a linear transformation; finally, the user's search behavior and the context are input into the RGNN model to obtain the user's interests in different contexts. Construct a context-triggered graph neural network recommendation model, including topology graph, topology relation matrix, Markov chain sampling, multi-level feature dimensionality reduction and convolution, fully connected layers, and recommendation output; The topology graph includes three scenarios: retrieval, location, and access. Different nodes are linked together with relationships. After the topology graph is input into the topology matrix, the Gibbs sampling algorithm is used to sample nodes in each convolutional layer. Linear discriminant analysis is used for multi-level dimensionality reduction to perform feature reduction and convolution. The obtained features are input into the fully connected layer, and finally the recommendation results are output. The graph neural network is improved, and its evolution is summarized as shown in Equation 1: （1） Where L is the output of the convolutional layer, m represents the layer number, and t represents time. H represents the set of vertices, H represents the set of nodes at each level, and A represents the adjacency matrix. This represents a nonlinear transformation, i.e., an activation function; Indicates an intermediate layer node; l represents the current layer, and M is the layer number. This indicates the last level where the current node j is located. Indicates the layer l where node j is located; The obtained loss function As shown in Equation 2: （2） Where L is the output of the convolutional layer, and M is the number of vertices. Indicates the layer number of node i. Indicates the activation function; The final recommendation result is obtained by deriving Equations 1 and 2 from the input data; The steps of this recommendation model include: Step 1: Construct the topology graph; Step 2: Construct the topological relation matrix; Step 3: Gibbs-based sampling; Step 4: Multi-stage feature dimensionality reduction and convolution; Step 5: Obtain user intent features and output recommendation results.
2. The context-aware document recommendation method as described in claim 1, characterized in that, User behavior log analysis includes the following steps: Analyze user search behavior to identify patterns and personalized habits among users when using the literature retrieval platform, thereby discovering and uncovering users' potential needs and interests. By analyzing user behavior, user behaviors during the information retrieval process can be categorized into three key behaviors: query, navigation, and location. Among them, query refers to the process where users enter the search platform, input search terms according to their own wishes, or make corresponding clicks, and the platform generates a corresponding set of results based on the user's input or clicks; navigation refers to the process where users perform a set of searches on the literature search platform, and the platform returns a corresponding set of results. The first page of the search results only displays a few documents that the system considers most relevant, and provides a page-turning function; positioning refers to the process where users click on a specific document on the platform, view the document details in the list of documents returned by the search, or click on a specific reference on the document details page; Based on the analysis of user search behavior, the specific user search steps are obtained.
3. The context-aware document recommendation method as described in claim 1, characterized in that, This includes building user session data streams; User search information refers to the information from when a user enters the platform until they leave; by analyzing the user's sequential clickstream, the click relationship between the user and specific information can be obtained. User sequential click flow includes the following steps: First, the unified identifier sessionid of the logs is used to match each specific behavior of the same user, and the user's search and click situation at that time is concatenated according to the time sequence; For all the click information of the connected users, sort them by time, distinguish the query behavior, navigation behavior and the location behavior determined by navigation, and determine the order of each query behavior, navigation behavior and location behavior according to the time sequence, and finally generate the complete click stream of the user. After extracting the complete user behavior sequence, the conversation stream is obtained, and the user's interests are obtained by determining the context based on the conversation stream; User interests are reflected in the topics of the user throughout the conversation flow. The similarity between two content feature records is obtained by calculating the topic features in pairs in chronological order. Set a similarity threshold to determine whether two content feature records in the same session belong to the same topic; if the similarity is higher than the set threshold, the actions are determined to belong to the same topic; conversely, if the similarity is lower than the set threshold, the actions are determined not to belong to the same topic, and the session is split between the two content feature records to form the corresponding unit session. After calculating and splitting all the sessions, a dataset of unit session streams with similar topics is formed; among them, the session streams are segmented according to different scenarios, and the similarity of different search terms is compared.
4. The context-aware document recommendation method as described in claim 1, characterized in that, The definition and description of the scenario include: After determining the session flow, the scenario in the retrieval process is identified by analyzing the acquired user log data, which triggers subsequent recommendations. During the user's search phase, information such as the selected search method, search terms, document type, author, institution, keywords, subject terms, publication time, source, journal terminology, year, volume, and issue is included. It also includes information on the specific documents that the user has clicked to view. This information can fully reflect the dynamic process of a user's search at that time. The data dimensions in user logs can be summarized into the following scenarios: Unique identifier for a unit session: The original session is constructed using the session ID in the user information; a unit session is the research unit after the original session has been split into topics, and at this time, a unique identifier is assigned to this unit session. Unique identifier for behavior: A behavior refers to a behavior record within a unit session. Each behavior record has a corresponding identifier, which is sorted by access time and then simply assigned a value. Behavior types: including three types of behavior: query, navigation, and location; Content features include document type, search terms, logical operators, matching methods, and search fields; Behavioral characteristics: Characteristics of user behavior beyond content preferences are meaningful for understanding user intent; these characteristics are collectively referred to as behavioral characteristics. Professionalism: refers to whether the user is clear about their search intent; user query behavior includes quick search and advanced search; in the context characterization process, different weight parameters are set according to the differences in search methods. If two search methods exist at the same time, the maximum strength is selected to represent the user's behavioral intent. Attention level: The degree of attention users pay to the current page.
5. The context-aware document recommendation method as described in claim 1, characterized in that, Dimensionality reduction is performed using a linear discriminant analysis-based method. First, the input topological matrix is reduced to a single dimension. This involves sampling the matrix, calculating the mean, and multiplying it with its transpose to obtain the reduced features, which are then input into the convolutional layer for convolution. The features obtained by convolution are then subjected to a second-level dimensionality reduction, which involves sampling, averaging, and transposing the samples for multiplication. The features after the second dimensionality reduction are then input into the second convolutional layer for convolution. This process is repeated until the set threshold is met. Finally, the dimensionality-reduced features are input into the fully connected matrix.