Information recommendation method and device, electronic equipment and storage medium

Abstract

The invention discloses an information recommendation method and device, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence or other related fields, and the method comprises the steps: constructing user features at least containing a reading completion rate and content features at least containing an information content category based on historical behavior data and information content data of a target user; processing the user features and the content features through a neural collaborative filtering model to obtain first features; processing a historical reading sequence, a corresponding reading completion rate and an information content category in the historical behavior data through a long short-term memory network model to obtain a second feature; processing the historical behavior data and the information content data through a graph convolutional network model to obtain a third feature; and generating a multi-dimensional user portrait based on the first feature, the second feature and the third feature, and generating an information recommendation scheme based on the multi-dimensional user portrait. According to the invention, the technical problem of low information recommendation accuracy in related technologies is solved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to an information recommendation method and apparatus, electronic device, and storage medium. Background Technology

[0002] With the rapid development and iteration of information technology, the financial industry has officially entered the era of big data. The financial market generates massive amounts of information and research reports daily. To enable users and investors across various industries to quickly browse the diverse information released in the financial market, major institutions have launched investment banking channels to centralize information and research reports for user access. However, for investors, the sheer volume of information makes it difficult to efficiently filter valuable information, resulting in information overload. Furthermore, the informational value of specific information varies from person to person.

[0003] Existing technologies typically present information in a uniform push format, failing to fully consider the differentiated needs of various users and making it difficult to achieve targeted pushes. This results in extremely high time costs for users to obtain effective information. Furthermore, traditional methods rely on manual screening and push, which can lead to information delays when financial market information is rapidly updated, causing users to miss crucial investment opportunities. Moreover, traditional service methods lack in-depth analysis of users' personalized behavioral data, failing to capture dynamic changes in user needs and providing continuously adapted information content. This results in low matching between information recommendation services and users, and low push accuracy.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This invention provides an information recommendation method, apparatus, electronic device, and storage medium to at least solve the technical problem of low accuracy in information recommendation in related technologies.

[0006] According to one aspect of the present invention, an information recommendation method is provided, comprising: constructing user features and content features based on historical behavior data and information content data of a target user, wherein the user features include at least a reading completion rate, and the content features include at least an information content category; processing the user features and the content features through a neural collaborative filtering model to obtain a first feature, wherein the first feature is used to characterize the interaction between the target user and the information content; processing the historical reading sequence, the corresponding reading completion rate, and the information content category in the historical behavior data through a long short-term memory network model to obtain a second feature, wherein the second feature is used to reflect the temporal changes in user interests; processing the historical behavior data and the information content data through a graph convolutional network model to obtain a third feature, wherein the third feature is used to characterize the user's potential indirect interests; generating a multi-dimensional user profile based on the first feature, the second feature, and the third feature, and generating an information recommendation scheme based on the multi-dimensional user profile.

[0007] Furthermore, the steps of constructing user features and content features based on the target user's historical behavior data and information content data include: acquiring N pieces of information content data, and classifying each piece of information content data according to a preset classification standard to obtain the information content category, where N is a positive integer; generating classification tags based on the information content categories, and generating the content features based on the information content data and the corresponding classification tags.

[0008] Furthermore, the steps of constructing user features and content features based on the target user's historical behavior data and information content data also include: acquiring M historical behavior data, wherein the historical behavior data includes the target user's historical reading information, reading timestamps, and reading duration, where M is a positive integer; performing anomaly filtering on the historical reading information based on the reading duration, and calculating the reading completion rate corresponding to each historical reading information; sorting the historical reading information based on the reading timestamps to obtain a historical reading sequence; and associating each historical reading information with the target user's user identifier and the reading completion rate corresponding to the historical reading information in the historical reading sequence to obtain the user features.

[0009] Further, the step of processing the user features and content features using a neural collaborative filtering model to obtain the first feature includes: inputting the user features and content features into the neural collaborative filtering model, wherein the neural collaborative filtering model comprises: an input layer, an embedding layer, a collaborative filtering layer, and an output layer; encoding the user features and content features respectively through the input layer to obtain corresponding user sparse variables and content sparse variables; mapping the user sparse variables and content sparse variables to vector spaces respectively through the embedding layer to obtain user latent vectors and content latent vectors; concatenating the user latent vectors and content latent vectors and transmitting them to the collaborative filtering layer, where a multilayer perceptron in the collaborative filtering layer calculates the nonlinear interaction pattern between the user and the information content to obtain nonlinear features; and outputting the nonlinear features as the first feature through the output layer.

[0010] Further, the step of processing the historical reading sequence, the corresponding reading completion rate, and the information content category in the historical behavior data through a Long Short-Term Memory (LSTM) network model to obtain the second feature includes: inputting the historical reading sequence, the user feature, and the content feature into the LTM network model, wherein the LTM network model includes at least an embedding layer and an LTM network layer; converting each historical reading information in the historical reading sequence into an embedding vector through the embedding layer, and determining the reading completion rate corresponding to each historical reading information as a vector feature of the embedding vector to obtain an embedding feature sequence; performing time-series analysis on the embedding feature sequence through the LTM network layer to obtain the second feature, wherein the LTM network layer includes at least a forgetting gate, the forgetting gate being used to filter historical reading information with a reading completion rate less than a first preset threshold, and filtering such historical reading information in the embedding feature sequence.

[0011] Further, the step of processing the historical behavior data and the information content data using a graph convolutional network model to obtain the third feature includes: constructing a heterogeneous graph based on the historical behavior data and the information content data, wherein the heterogeneous graph includes user nodes and information nodes; in the heterogeneous graph, for the edges between the user nodes and the information nodes, setting a first type of weight as the user's completion rate of reading the information content; in the heterogeneous graph, for the edges between the information nodes and another information node, setting a second type of weight as information similarity; inputting the heterogeneous graph into the graph convolutional network model, performing link analysis on the heterogeneous graph through convolutional layers to obtain the link probability from the user node to each information node; using the user node and the information node as link objects, and combining the link probabilities to generate the third feature.

[0012] Furthermore, the step of generating an information recommendation scheme based on the multi-dimensional user profile includes: determining the information content category of the information to be recommended; calculating the reading completion rate score of the target user for each information content category based on the multi-dimensional user profile; sorting the information to be recommended corresponding to each information content category according to the reading completion rate to obtain an information recommendation sequence; and generating the information recommendation scheme based on the information recommendation sequence.

[0013] According to another aspect of the present invention, an information recommendation apparatus is also provided, comprising: a construction unit, configured to construct user features and content features based on historical behavior data and information content data of a target user, wherein the user features include at least a reading completion rate, and the content features include at least an information content category; a first processing unit, configured to process the user features and the content features through a neural collaborative filtering model to obtain a first feature, wherein the first feature is used to characterize the interaction between the target user and the information content; a second processing unit, configured to process the historical reading sequence, the corresponding reading completion rate, and the information content category in the historical behavior data through a long short-term memory network model to obtain a second feature, wherein the second feature is used to reflect the temporal changes in user interests; a third processing unit, configured to process the historical behavior data and the information content data through a graph convolutional network model to obtain a third feature, wherein the third feature is used to characterize the user's potential indirect interests; and a generation unit, configured to generate a multi-dimensional user profile based on the first feature, the second feature, and the third feature, and generate an information recommendation scheme based on the multi-dimensional user profile.

[0014] Furthermore, the construction unit includes: a first acquisition module, used to acquire N pieces of information content data, and classify each piece of information content data according to a preset classification standard to obtain the information content category, wherein N is a positive integer; and a first generation module, used to generate classification tags based on the information content categories, and generate the content features based on the information content data and the corresponding classification tags.

[0015] Furthermore, the construction unit further includes: a second acquisition module, used to acquire M historical behavior data, wherein the historical behavior data includes the target user's historical reading information, reading timestamps, and reading duration, and M is a positive integer; a first calculation module, used to perform anomaly filtering on the historical reading information based on the reading duration, and calculate the reading completion rate corresponding to each historical reading information; a first sorting module, used to sort the historical reading information based on the reading timestamps to obtain a historical reading sequence; and an association module, used to associate each historical reading information with the target user's user identifier and the reading completion rate corresponding to the historical reading information in the historical reading sequence to obtain the user characteristics.

[0016] Further, the first processing unit includes: a first input module, used to input the user features and the content features into the neural collaborative filtering model, wherein the neural collaborative filtering model includes: an input layer, an embedding layer, a collaborative filtering layer, and an output layer; an encoding module, used to encode the user features and the content features respectively through the input layer to obtain corresponding user sparse variables and content sparse variables; a mapping module, used to map the user sparse variables and the content sparse variables to a vector space respectively through the embedding layer to obtain user latent vectors and content latent vectors; a concatenation module, used to concatenate the user latent vectors and the content latent vectors and transmit them to the collaborative filtering layer, where a multilayer perceptron in the collaborative filtering layer calculates the nonlinear interaction pattern between the user and the information content to obtain nonlinear features; and an output module, used to output the nonlinear features as the first feature through the output layer.

[0017] Further, the second processing unit includes: a second input module, used to input the historical reading sequence, the user features, and the content features into the long short-term memory network model, wherein the long short-term memory network model includes at least an embedding layer and a long short-term memory network layer; a conversion module, used to convert each historical reading information in the historical reading sequence into an embedding vector through the embedding layer, and determine the reading completion rate corresponding to each historical reading information as a vector feature of the embedding vector to obtain an embedding feature sequence; and an analysis module, used to perform time-series analysis on the embedding feature sequence through the long short-term memory network layer to obtain the second feature, wherein the long short-term memory network layer includes at least a forgetting gate, the forgetting gate is used to filter historical reading information with a reading completion rate less than a first preset threshold, and to filter the historical reading information in the embedding feature sequence.

[0018] Further, the third processing unit includes: a construction module, configured to construct a heterogeneous graph based on the historical behavior data and the information content data, wherein the heterogeneous graph includes user nodes and information nodes; a first setting module, configured to set a first type of weight as the user's reading completion rate of the information content for the edge between the user node and the information node in the heterogeneous graph; a second setting module, configured to set a second type of weight as information similarity for the edge between the information node and another information node in the heterogeneous graph; a third input module, configured to input the heterogeneous graph into the graph convolutional network model, and perform link analysis on the heterogeneous graph through convolutional layers to obtain the link probability from the user node to each information node; and a second generation module, configured to generate the third feature by using the user node and the information node as link objects and combining the link probability.

[0019] Further, the generation unit includes: a determining module, used to determine the information content category of the information to be recommended; a second calculation module, used to calculate the reading completion rate score of the target user for each information content category based on the multi-dimensional user profile; a second sorting module, used to sort the information to be recommended corresponding to each information content category according to the reading completion rate to obtain an information recommendation sequence; and a third generation module, used to generate the information recommendation scheme based on the information recommendation sequence.

[0020] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the information recommendation method described in any one of the above embodiments.

[0021] According to another aspect of the present invention, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the information recommendation method described in any of the preceding embodiments.

[0022] According to another aspect of the present invention, a computer program product is also provided, including computer instructions, wherein when the computer instructions are executed by a processor, they implement the steps of the information recommendation method described in any one of the above embodiments.

[0023] This invention proposes an information recommendation method. First, user features and content features are constructed based on the target user's historical behavior data and information content data. The user features include at least the reading completion rate, and the content features include at least the information content category. Then, a neural collaborative filtering model is used to process the user features and content features to obtain a first feature, which characterizes the interaction between the target user and the information content. Next, a long short-term memory network model is used to process the historical reading sequence, corresponding reading completion rate, and information content category in the historical behavior data to obtain a second feature, which reflects the temporal changes in user interests. Then, a graph convolutional network model is used to process the historical behavior data and information content data to obtain a third feature, which characterizes the user's potential indirect interests. Finally, a multi-dimensional user profile is generated based on the first, second, and third features, and an information recommendation scheme is generated based on this multi-dimensional user profile.

[0024] This invention employs a method of deeply mining multi-dimensional features of user reading behavior and information content. By integrating innovative approaches such as neural collaborative filtering models, long short-term memory network models, and graph convolutional network models, it achieves a comprehensive understanding of user interests and preferences. This results in a significant improvement in the personalization level and technical accuracy of information recommendation systems, thereby solving the technical problem of low accuracy in information recommendation in related technologies.

[0025] Specifically, this invention first constructs user and content features using historical behavioral data and information content data of target users. User features focus on reading completion rate, while content features focus on information classification systems. Then, a neural collaborative filtering model processes user and content features to uncover non-linear interaction relationships and capture the deep association between users and information content, forming the first feature. Next, a long short-term memory network model is used to perform temporal analysis on user reading sequences, combining reading completion rate and information category to extract the second feature reflecting dynamic changes in user interests, enhancing the timeliness and adaptability of the recommendation scheme. Finally, a graph convolutional network model is used to analyze links on the heterogeneous graph of user-information, revealing potential indirect interest points of users to obtain the third feature, expanding the scope of the recommendation system. Ultimately, the multi-dimensional user profile constructed based on these three features becomes the core of generating customized, high-precision information recommendation schemes, ensuring that information recommendations not only meet users' immediate needs but also reveal their potential interests. This fundamentally solves the problem of low accuracy in information recommendations caused by single user profiles and insufficient capture of interest changes in related technologies. Attached Figure Description

[0026] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0027] Figure 1 This is a flowchart of an optional information recommendation method according to an embodiment of the present invention;

[0028] Figure 2 This is a model diagram of an optional neural collaborative filtering model according to an embodiment of the present invention;

[0029] Figure 3 This is a model diagram of an optional long short-term memory network model according to an embodiment of the present invention;

[0030] Figure 4 This is a model diagram of an optional graph convolutional network model according to an embodiment of the present invention;

[0031] Figure 5 This is a schematic diagram of an optional information recommendation device according to an embodiment of the present invention;

[0032] Figure 6 This is a structural block diagram of an electronic device for performing an information recommendation method according to an embodiment of the present invention. Detailed Implementation

[0033] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0034] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0035] To facilitate understanding of the present invention by those skilled in the art, some terms or nouns involved in the various embodiments of the present invention are explained below:

[0036] NCF, or Neural Collaborative Filtering, is a deep learning model that captures the complex relationship between user preferences and item features by modeling data from user-item interactions. The NCF model maps user and item features to a low-dimensional space, then learns the non-linear interactions between these features using a Multi-Layer Perceptron (MLP), ultimately generating a predicted rating or level of preference for the user towards the item.

[0037] LSTM, or Long Short-Term Memory, is a special type of recurrent neural network (RNN) specifically designed for processing and predicting events in time series, especially those with long intervals and delays. LSTM selectively remembers or forgets information in a sequence by introducing three gating mechanisms—a forget gate, an input gate, and an output gate—effectively handling long-term dependencies in the sequence. In this invention, LSTM is used to analyze the temporal characteristics of user reading behavior and capture the changing trajectory of user interests.

[0038] GCN, or Graph Convolutional Networks, is a deep learning model applicable to graph-structured data, primarily used for node classification, link prediction, and graph classification tasks. Unlike traditional CNNs (Convolutional Neural Networks), GCNs can perform convolution operations on both node features and the graph structure simultaneously, capturing information between adjacent nodes. In this invention, GCNs are used to analyze potential relationships between information and, combined with user historical behavior, to uncover users' potential indirect interests in related content.

[0039] The following embodiments of the present invention can be applied to various systems / applications / devices that require precise personalized information recommendations and optimized distribution of investment banking business information, enabling a personalized information recommendation platform based on deep user behavioral feature analysis and intelligent prediction algorithms. The present invention uses a neural collaborative filtering model to model the interaction relationship between users and information content, capturing nonlinear features to reflect user preferences. Then, it combines a long short-term memory network model to handle the temporal dependence of user reading behavior, identifying dynamic changes in user interests and better adapting to the characteristics of user interests evolving over time. Simultaneously, a graph convolutional network model is used to analyze potential indirect user interests and mine the correlations between content, ensuring the comprehensiveness and depth of the recommendations.

[0040] This invention integrates advanced deep learning technologies such as neural collaborative filtering, long short-term memory networks, and graph convolutional networks, avoiding the limitations of traditional recommendation algorithms based on single metrics (such as click volume or search history). It also overcomes the accuracy challenges caused by data sparsity and the dynamic nature of user interests, significantly improving the personalization level and technical accuracy of information recommendations in the financial field. It is not only suitable for rapidly changing financial information scenarios, but also provides more targeted information services for various financial institutions and investors, promoting the application and deepening of fintech and artificial intelligence in the financial industry.

[0041] The present invention will now be described in detail with reference to various embodiments.

[0042] Example 1

[0043] According to an embodiment of the present invention, an information recommendation method embodiment is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0044] The implementation subject of this invention can be an intelligent information recommendation system for investment banks, which combines deep learning technology and user behavior data analysis technology for personalized information service scenarios in the financial market. In particular, it addresses the problems of information overload and insufficient personalized recommendations. Through the construction of multi-dimensional user profiles and the adjustment of dynamic recommendation strategies, specifically through the steps of capturing nonlinear features using a neural collaborative filtering model, capturing reading behavior sequence dependencies using a long short-term memory network model, and capturing potential indirect interest associations using a graph convolutional network model, the aim is to provide investment banking clients with accurate, personalized, and dynamically adjusted information recommendations.

[0045] In existing technologies, some commonly used recommendation systems make recommendations based on users' query history, historical click history, favorites history, and click volume of trending news, making it difficult to fully capture user behavior data. One reason is that these viewing histories do not fully reflect a user's interest in the information. For users, opening news for extensive reading is basic; they may close it immediately if they are not interested, but in such cases, clicking on the news can be mistakenly interpreted as an indication of interest. Another reason is that a significant proportion of users are "silent users," who are not accustomed to saving or liking news.

[0046] Therefore, reading time data is needed to refine user profiles and achieve more accurate information recommendations. The essence of reading time is the deep interaction characteristics between users and content, and its core value lies in two aspects: First, it distinguishes between users' superficial interests (such as accidental clicks, headlines that attract attention, etc.) and deep interests (serious reading); second, it is more user-friendly for content scenarios without explicit ratings (such as news, information, etc.), without requiring users to actively evaluate or rate it.

[0047] The aforementioned implementing entities can be integrated into various financial information service platforms, mobile clients, online investment advisory systems, etc. Through deep learning neural collaborative filtering models, long short-term memory network models, and graph convolutional network models, as well as data processing techniques such as reading completion rate calculation and outlier filtering in user behavior data analysis technology, they can achieve personalized and intelligent recommendations for investment banking information. This is particularly suitable for the efficient distribution of financial market information and personalized services in the context of big data, helping to avoid user churn due to content mismatch and investment opportunity loss due to information lag, thereby enhancing the professionalism and market competitiveness of financial services.

[0048] The embodiments of the present invention will now be described in detail with reference to the specific implementation steps.

[0049] Figure 1 This is a flowchart of an optional information recommendation method according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps:

[0050] Step S101: Construct user features and content features based on the target user's historical behavior data and information content data. The user features include at least the reading completion rate, and the content features include at least the information content category.

[0051] Specifically, target users refer to the recipients of information recommendation system services from a specific financial institution. These can be registered users or anonymous visitors with known behavioral patterns. In this invention, target users are various types of investors in the financial market, including but not limited to individual investors, institutional investors, and corporate decision-makers. Historical behavioral data covers the target user's past interactions within the information recommendation system, including but not limited to search history, reading history, click records, collection behavior, and sharing actions. Information content data includes all available financial information and its specific attributes, such as: information length, information type (e.g., macroeconomic analysis, corporate financial report interpretation, industry news), industry, difficulty level, number of clicks, and number of collections. Refined analysis of information content data helps improve the accuracy of recommendations, thereby effectively matching them with user characteristics.

[0052] Furthermore, in this invention, user characteristics are not limited to traditional personal information or preference settings. More importantly, comprehensive characteristics that reflect users' reading preferences and investment inclinations are extracted. One core characteristic is the reading completion rate, which quantifies the user's level of engagement with the information, eliminating interference from accidental clicks or brief browsing, and providing an important basis for accurately grasping user interests. In addition to basic information, content characteristics may also be included, reflecting the intrinsic value and attractiveness of the information, such as the richness of the topic, authority, and timeliness.

[0053] User characteristics include at least the reading completion rate, defined as the ratio of the actual time a user spends reading a piece of information to the estimated reading time. This reflects the proportion of content actually read in its entirety or the degree to which a single piece of content is read completely, thus measuring the user's genuine interest in the information. Through standardization, the reading completion rate can more objectively reflect a user's willingness to read in depth, thereby improving the accuracy of personalized recommendations after removing outliers and data from "silent users." Content characteristics include at least the information content category, referring to a detailed classification of information, such as macroeconomics, stock market dynamics, bond analysis, insurance regulations, and the real estate industry. Accurately classifying information content helps build a richer information tagging system, improves the recommendation algorithm's ability to handle information diversity, and ensures the breadth and depth of recommended content.

[0054] In one optional approach, the steps of constructing user and content features based on the target user's historical behavior data and information content data include: acquiring N pieces of information content data and classifying each piece of information content data according to a preset classification standard to obtain information content categories, where N is a positive integer; generating classification tags based on the information content categories; and generating content features based on the information content data and the corresponding classification tags. The purpose of the above steps is to perform in-depth analysis and tagging of information content in order to more accurately match it with user preferences.

[0055] Specifically, the preset classification criteria may include: subject area (macroeconomics, stock market, bond analysis, industry news, regulation interpretation, etc.); article length (short, medium, long); difficulty level (beginner, intermediate, professional); popularity (click rate, collection rate, number of comments, etc.).

[0056] An optional approach is to classify news content data using the following steps: extract key information and keywords from the news content using text analysis tools; select classification features based on preset classification criteria (e.g., article length determined by word count, topic domain identified by topic model); build a classification model (e.g., K-means clustering, Naive Bayes classifier, or deep learning model), train the model to identify and classify news content; and apply the trained model to new news content data, automatically assigning classification labels. These labels can be represented by a series of numerical values ​​or characters, such as 1 for macroeconomic analysis, 2 for stock market information, and 3 for bond analysis reports, or in character form, such as "MACRO" for macroeconomics and "STOCK" for the stock market.

[0057] Furthermore, the process of generating content features based on information content data and classification tags involves converting textual information into a computable numerical representation. For example, text representation techniques such as the bag-of-words model, TF-IDF (Term Frequency-Inverse Document Frequency), and word2vec can be used to convert information text into feature vectors. Embedding classification tags into feature vectors, or adding them separately as binary or numerical features, can enhance the model's sensitivity to classification information. The content feature is ultimately represented as a high-dimensional feature vector, where each element represents a score for a certain attribute of the information, such as topic score, length score, and difficulty score. For example, a long article about macroeconomic analysis with a technical level of difficulty might have a content feature vector represented as (0.85, 1.0, 0.9, 0.8) (representing topic matching degree, length score, difficulty score, and popularity score, respectively).

[0058] In a specific optional implementation scenario, User A has a habit of frequently reading company financial reports and prefers in-depth analysis, while User B prefers macroeconomic analysis and shorter news articles. To achieve personalized recommendations, the system of this invention needs to perform detailed classification and feature extraction of the information. First, all information content data is retrieved from the database, where N represents the total number of articles and can be any positive integer. Preset classification criteria include, but are not limited to, the topic range of the information (e.g., macroeconomics, stock market, bond market, foreign exchange, gold, real estate, technology, healthcare, etc.), article length (short, medium, long), difficulty level (beginner, intermediate, advanced), click count, and number of favorites. Click count and number of favorites can be used to measure the popularity of the information, while length and difficulty can be used to standardize the reading time. Natural language processing techniques, such as topic models, keyword extraction, and classification algorithms (decision trees, support vector machines, etc.), can be used to classify the information content data, automatically assigning preset classification labels based on keywords and topic sentences in the information text. Category tags are codes or text identifiers used to indicate attributes such as topic, length, and difficulty of information. They are typically in machine-readable formats, such as numeric codes (1 for macroeconomic analysis, 2 for corporate financial reports), letter codes (L for long articles, S for short articles), or mixed codes (L1 for long macroeconomic analysis). Using category tags facilitates rapid retrieval and categorization. Further generating content features involves combining information content data with category tags. For example, for a long, in-depth analysis of the technology industry (category tag L3), the system combines text content, keyword density, and the amount of cited data to generate a multi-dimensional feature vector including topic, length, difficulty, and keyword frequency. This is represented as a high-dimensional numerical vector, with each dimension corresponding to an attribute of the information, such as topic proportion, text length, or difficulty score. The more refined the information categorization, the better it can recommend information within the corresponding category based on user behavior data and reading habits. Finally, the content features can be represented as: , where n=N.

[0059] In one optional scheme, the steps of constructing user features and content features based on the target user's historical behavior data and information content data further include: obtaining M historical behavior data, wherein the historical behavior data includes the target user's historical reading information, reading timestamps, and reading duration, where M is a positive integer; performing anomaly filtering on the historical reading information based on reading duration and calculating the reading completion rate corresponding to each historical reading information; sorting the historical reading information based on reading timestamps to obtain a historical reading sequence; and associating each historical reading information with the target user's user identifier and the reading completion rate corresponding to the historical reading information in the historical reading sequence to obtain user features.

[0060] Specifically, historical behavioral data can include records of various interactions a user has had with information content over a past period, such as browsing, clicking, searching, saving, and sharing. The key focus is the user's historical reading history, i.e., a list of information the user has previously read, along with corresponding behavioral details, such as reading timestamps (recording the precise time each reading occurred) and reading duration (the actual time the user remained on the information page). When acquiring this data, the system retrieves all reading records associated with the user's identifier from the database. M represents the total number of historical information reads by the user, which can be any positive integer, large enough to cover the user's recent and even long-term reading behavior.

[0061] Furthermore, the system implementation needs to perform outlier detection on the reading duration of historical information to remove noise that may be caused by user errors or other abnormal behaviors. For example, it filters out information records with a reading duration of less than 3 seconds (considered as accidental clicks) and records with a reading duration of more than 1 hour (possibly because the page was opened but not closed). In addition, it checks whether the reading duration deviates significantly from the reading time estimated based on the information length; if it exceeds three times the estimated time, it will also be considered an outlier and excluded.

[0062] Furthermore, the system can also calculate the reading completion rate. The reading completion rate quantifies a user's level of interest in the information by comparing the actual reading time with the expected reading time. The expected reading time can be calculated by considering the number of words or paragraphs in the information, combined with the user's average reading speed. The formula for calculating the reading completion rate is: Reading Completion Rate = Actual Reading Time / Expected Reading Time. This metric allows for a more accurate assessment of the user's level of engagement with the information, thereby improving the accuracy of recommendations.

[0063] In addition, it is necessary to sort users' historical reading information based on reading timestamps to create a chronological reading sequence. This is used to capture the evolution of users' interests, as reading behavior at different points in time may reflect dynamic changes in users' interests. Each piece of historical reading information in the historical reading sequence is associated with user identifiers, reading completion rates, etc., to form a comprehensive user feature vector. This vector not only includes users' preferences for different types of information but also reflects the strength of these preferences through reading completion rates, and captures changes in interests over time through sequence dependencies.

[0064] In one possible implementation scenario, it's first necessary to collect user search history, browsing history, liked articles, saved articles, click counts, shared articles, and the user's reading time for each article (represented by the time the user spends on the article page). Each data feature is then associated with a unique timestamp, representing the time the user browsed or saved the article. Furthermore, a dataset (i.e., user features) needs to be established by linking the user's unique identifier with the above data, serving as a crucial basis for user behavior analysis. Collecting reading time facilitates analyzing the topics of deep interest to users from the perspective of browsing time. User features can be denoted as: , where m=M.

[0065] Step S102: Process user features and content features through a neural collaborative filtering model to obtain a first feature, wherein the first feature is used to characterize the interaction between the target user and the information content.

[0066] Specifically, Neural Collaborative Filtering (NCF) learns user-information interaction functions through a multilayer perceptron, capturing more complex nonlinear relationships between users and information. It takes the feature vectors corresponding to information and users as input, and uses deep learning to determine the user's preference for information based on nonlinear interaction patterns. NCF maps user and information content features to a low-dimensional vector space through an embedding layer, and then uses a neural network to capture their interaction patterns, ultimately obtaining deep interaction features between users and content.

[0067] Figure 2 This is a model diagram of an optional neural collaborative filtering model according to an embodiment of the present invention, such as... Figure 2 As shown, the neural collaborative filtering model includes an input layer, an embedding layer, a collaborative filtering layer, and an output layer. After obtaining sparse variables from user and content features through one-hot encoding, these are mapped to latent vectors p and q through the embedding layer. The Hadamard product is used to interact the user and information feature vectors, i.e.: ;in, This represents the processing function of the first layer of the neural network. u and i are the two core objects that the model processes. u represents the user, and i represents the content, such as a piece of technology news or a piece of sports news. This refers to a user latent vector, which encodes the implicit interests and preferences of user u. For example, this vector... High values ​​on certain dimensions indicate that user u has a strong interest in "technology" and "long articles". This refers to the content latent vector, which encodes the implicit attributes of content i. For example, this vector... The high values ​​on certain dimensions indicate that this news article has strong "technology" and "long article" attributes.

[0068] The latent vectors are concatenated and fed into a multilayer perceptron for training. The multilayer perceptron utilizes activation functions and a multilayer structure; by adjusting the number of hidden layers, neurons, and activation functions, it is suitable for various scenarios such as regression, classification, prediction, and feature learning. Here, a multilayer perceptron is used for feature learning, learning the non-linear interaction function between the user and the information content, ultimately obtaining non-linear features. The specific training process is as follows:

[0069] ,in, This is the input to the first layer of the neural network, containing all the raw, potential information about the user and the content, ready to be fed into subsequent deeper layers for further processing; in this embodiment... The corresponding specific operation is splicing; That is, the stacked matrices are combined into a long vector.

[0070] ,in, This describes the performance of any layer (the Lth layer) in the neural network, where L represents the layer number index. For example, L=2 represents the second layer, and L=3 represents the third layer. It is the input of the Lth layer and also the output of the previous layer. For the first layer, ; This is the weight matrix for the Lth layer, where each value (weight) determines the input to the previous layer. The importance of each feature in the current layer to each neuron can also be understood as the "connection strength" between different features; It is the transpose of the weight matrix; It is the bias vector of the Lth layer, used to add a basic activation threshold to each neuron, making the model more flexible, so that the neuron will be activated even when all inputs are 0; It is the activation function of the Lth layer (such as ReLU, Sigmoid), which is a non-linear function.

[0071] Through the above process, the neural collaborative filtering model captures users' potential interests and content attributes. In the collaborative filtering interaction layer, it learns the non-linear interactions between users and information, concatenating or fusing features to obtain deeper non-linear features. For example, user A has a higher completion rate for long articles in the "technology" category, while user B has fewer clicks on short articles in the "agriculture" category. This process mainly utilizes the precise matching capability of NCF. After deeply capturing these features, it can more accurately match users' deep interests, thereby achieving better recommendation results.

[0072] In one optional approach, the step of processing user features and content features using a neural collaborative filtering model to obtain a first feature includes: inputting user features and content features into the neural collaborative filtering model, wherein the neural collaborative filtering model comprises: an input layer, an embedding layer, a collaborative filtering layer, and an output layer; encoding user features and content features separately through the input layer to obtain corresponding user sparse variables and content sparse variables; mapping user sparse variables and content sparse variables to vector spaces through the embedding layer to obtain user latent vectors and content latent vectors; concatenating the user latent vectors and content latent vectors and transmitting them to the collaborative filtering layer, where a multilayer perceptron in the collaborative filtering layer calculates the nonlinear interaction pattern between the user and the information content to obtain nonlinear features; and outputting the nonlinear features as the first feature through the output layer.

[0073] It's important to note that input features in recommender systems often exist in categorical form, such as user IDs and news categories. The process of converting these categorical features into numerical forms that computers can process is called encoding. One-hot encoding is a viable encoding method, where each category feature is treated as an independent binary variable, with each variable representing a category. The variable value is 1 when a category is represented, and 0 for the rest. One-hot encoding preserves the independence of category features, but it increases feature dimensionality when the number of categories is large. The encoded feature variables are then sparse variables. Due to the characteristics of one-hot encoding, most elements in this variable are 0, with only a few categories corresponding to elements of 1. Using sparse variables for computation can lead to difficulties in model training, as a large number of 0 values ​​weakens the representational power of the feature vector. Therefore, further processing to a vector space is necessary.

[0074] By using an embedding layer, high-dimensional sparse features are mapped to a low-dimensional continuous vector space. Similar users or similar information in this space are mapped to nearby locations. This mapping method helps to capture the potential correlations between features while reducing computational cost. The latent vector is the representation of a user or information in the embedding space, capable of representing the complex relationships between users and information.

[0075] In existing technologies, traditional collaborative filtering algorithms are often based on linear relationships, assuming that user preferences for information can be linearly predicted by the direct relationship between user and information ratings or click frequencies. This approach has limitations when dealing with non-linear and complex user preferences. This embodiment, however, leverages the non-linear learning capabilities of neural networks to more accurately capture the patterns of user preferences changing with information attributes (such as type, length, and difficulty), improving the personalization and accuracy of recommendations. Furthermore, the neural collaborative filtering model performs better in handling the cold start problem (i.e., new users or new information), and can predict unknown interactions to some extent through the learning of latent vectors.

[0076] Step S103: The historical reading sequence, corresponding reading completion rate and information content category in the historical behavior data are processed by the Long Short-Term Memory Network model to obtain the second feature, which is used to reflect the temporal changes of user interests.

[0077] Specifically, the historical reading sequence can contain a list of information read by the user in the past. Each piece of information includes specific content, a reading timestamp, a corresponding completion rate, and the content category. Before processing, this information needs to be converted into a numerical vector form, such as using word embeddings to convert the content into a vector and using one-hot encoding to represent the content category. LSTM, through its special gating mechanism (input gate, output gate, forget gate), can "remember" important sequence information and "forget" irrelevant information. In this embodiment, LSTM adjusts the memory units according to the completion rate and content category, retaining the user's recent high interest in a certain type of information while gradually forgetting earlier reading behaviors, reflecting the natural decay of user interest over time. When reading information continuously, a user's interest is often influenced by previously read content. For example, after reading information about blockchain technology, a user may be more inclined to read content related to cryptography. LSTM can capture these potential reading order dependencies by analyzing the historical reading sequence, providing temporal features of the evolution of user interests. After training, LSTM can generate a comprehensive output based on historical behavior data, namely the second feature, which contains an abstract representation of the temporal changes in user interests. For example, the second feature could be a vector, where each element represents a user's current level of interest in a particular information category and how that interest changes over time.

[0078] In one specific implementation scenario, a user might have enjoyed reading real estate news a week ago, but recently prefers reading news about emerging technologies. Therefore, the reading types from a week ago would be of low interest to the user now. Similarly, after reading an article related to artificial intelligence, the user might be more inclined to read similar articles. In this case, this embodiment uses a Long Short-Term Memory (LSTM) network to capture the sequential dependencies of reading behavior. The LSTM network transmits memories in a precise way: deciding which information needs to be remembered, which needs to be updated, and which memories need to be forgotten. In this embodiment, the input is the user's historical reading sequence sorted by time, along with features such as the user's reading completion rate and content category. After the embedding layer transforms each reading content into an embedding vector, the reading completion rate is used as the sequence feature. Then, the sequence data is processed in the LSTM network layer. A forgetting gate in the gating mechanism filters out irrelevant history, filtering out reading sequences with low completion rates from a certain period, retaining key interests and reading sequences—that is, reading sequences with high completion rates in recent times. The output layer uses the hidden state of the last time step to represent the user's current interests, modeling the dynamic changes in user interests.

[0079] Figure 3 This is a model diagram of an optional long short-term memory network model according to an embodiment of the present invention, such as... Figure 3 As shown, recurrent neural networks use the output of the previous step as the input of the next step, while long short-term memory networks incorporate mechanisms such as forget gates to better balance long-term and short-term features. Representing long-term memory, it determines which long-term features need to be retained and which need to be discarded. For example, if a user read a technology-related article and an agriculture-related article last week, and the completion rate for the technology article was 0.9 while that for the agriculture article was only 0.2, the Long Short-Term Memory (LSTM) network would consider the act of reading the technology article to be retained in long-term memory, while the act of reading the agriculture article should be forgotten. Representing working memory, it learns how to focus attention on relevant information in long-term memory. Furthermore, four key vectors in the memory network determine the final output: `remember` (also known as the forgetting gate, where 1 represents complete retention and 0 represents complete forgetting); `save` (also known as the input gate, determining how many features are allowed into the memory network); and `focus` (also known as the output gate, outputting the final feature result through the attention vector). By leveraging the temporal modeling capabilities of LSTM while considering time decay, Long Short-Term Memory (LSTM) comprehensively captures the dynamic changes in user interests, enabling more accurate personalized information recommendations.

[0080] In one optional approach, the step of processing historical reading sequences, corresponding reading completion rates, and information content categories in historical behavioral data using a Long Short-Term Memory (LSTM) network model to obtain a second feature includes: inputting historical reading sequences, user features, and content features into the LTM network model, wherein the LTM network model at least includes an embedding layer and an LTM network layer; converting each historical reading information in the historical reading sequence into an embedding vector through the embedding layer, and determining the reading completion rate corresponding to each historical reading information as a vector feature of the embedding vector to obtain an embedding feature sequence; performing temporal analysis on the embedding feature sequence through the LTM network layer to obtain the second feature, wherein the LTM network layer at least includes a forgetting gate, the forgetting gate being used to filter historical reading information with a reading completion rate less than a first preset threshold, and filtering such historical reading information in the embedding feature sequence.

[0081] It's important to note that the embedding layer transforms discrete historical reading information into continuous numerical vector representations, i.e., embedding vectors. In this embodiment, the overall embedding layer converts information categories and related attributes (such as length and difficulty) into low-dimensional vectors, facilitating model understanding and processing. Through the embedding layer, information content is no longer a complex structure difficult to calculate directly, but rather transformed into vectors capable of mathematical operations. This allows the model to represent the similarity of content attributes between pieces of information as vector distance or angle, enabling it to better capture the potential connections between different pieces of information. Determining the reading completion rate as a vector feature means directly incorporating this numerical value as part of the embedding vector's features. Integrating the quantitative indicator of reading time directly into the information representation allows the model to directly obtain quantitative information about the user's depth of interest in the information. The reading completion rate, as an important indicator of user interest in information, helps the model more accurately identify and predict user preferences for different types of information.

[0082] The resulting embedded feature sequence is a series of vector representations encoded from the user's historical reading information, arranged in chronological order. Each vector contains the embedded features of the information and the reading completion rate feature, serving as direct input to the LSTM model, enabling the model to capture patterns in user reading behavior over time. The construction and arrangement of the sequence ensures that the LSTM can process information in chronological order, thereby identifying dynamic trends in user interests. In the Long Short-Term Memory (LSTM) network layer, the main operation of the model's temporal analysis is to regulate and transmit information through internal forget gates, input gates, and output gates. The forget gate acts as a filter, especially for historical reading information with a reading completion rate below a preset threshold (i.e., the first preset threshold), as these may represent content that the user was not interested in in the past or information whose user interest has waned. The forget gate reduces or eliminates the influence of this information in the sequence, ensuring that the model focuses on learning the patterns of recent user attention and changes in interest.

[0083] The temporal analysis process specifically includes: Information input: Through the input gate, the model determines which information needs to be added to the current unit state; Information forgetting: The forgetting gate determines which historical information no longer affects the current state, especially information with low completion rates; Information output: The output gate determines which information needs to be passed to the next time step, i.e., which information is most critical to the user's current interest. LSTM can effectively process sequential data and identify the evolution of user interests through the above gating operations, thereby obtaining a second feature that reflects the temporal changes in the user's interest in information.

[0084] Step S104: The historical behavior data and information content data are processed by a graph convolutional network model to obtain a third feature, which is used to characterize the user's potential indirect interests.

[0085] Specifically, graph convolutional network models are deep learning models specifically designed for processing graph-structured data. They can effectively learn the association features between nodes and the structural information of the graph. In this embodiment, GCN is used to construct an indirect interest association network between users and information, thereby capturing users' potential indirect interests in information. This is a deep personalized feature that is difficult for traditional recommendation algorithms to reach.

[0086] In practice, users and news items are treated as nodes in a graph, while reading behaviors between users and news items, and similarities between news items, form edges, creating a heterogeneous graph. This graph structure not only contains explicit user behavior information but also reflects potential connections between news items through the associations between news node associations. GCN allows features between news node associations to be passed along edges through graph convolution operations. After multiple passes, the features of each news item will contain information about its neighboring news items, establishing indirect interest associations between news items. Through multi-layer graph convolution operations, GCN can learn the complex relationships between user nodes and news node associations, thereby inferring that users may be interested in other news items that they have not directly read but are related to the news items they have previously read in terms of topic and type. This is the third feature, representing a quantitative representation of the user's potential indirect interests.

[0087] In one alternative approach, the step of processing historical behavior data and information content data using a graph convolutional network model to obtain the third feature includes: constructing a heterogeneous graph based on the historical behavior data and information content data, wherein the heterogeneous graph contains user nodes and information nodes; in the heterogeneous graph, for the edges between user nodes and information nodes, setting a first type of weight as the user's completion rate of reading information content; in the heterogeneous graph, for the edges between an information node and another information node, setting a second type of weight as information similarity; inputting the heterogeneous graph into the graph convolutional network model, performing link analysis on the heterogeneous graph through convolutional layers to obtain the link probability from user nodes to each information node; and generating the third feature by using user nodes and information nodes as link objects and combining the link probabilities.

[0088] It's important to note that after capturing nonlinear and temporal features, it's also necessary to consider the correlations of potential indirect interests. These features can be captured through the construction of heterogeneous graphs and the training of graph neural networks. In practical applications, this manifests as follows: if user A is interested in the content of information X, and information X and information Y have high similarity, then user A is likely also interested in the content of information Y.

[0089] First, a heterogeneous graph needs to be established. In this embodiment, the nodes in the graph structure have two types of meanings: representing users and representing specific information. The edges in the graph structure correspond to two types: user-information edges and information-information edges. For user-information edges, the weight is set to the user's completion rate of reading the information. For information-information edges, the weight is set to the similarity of the information, including category, length, etc., which can be calculated using cosine similarity. Then, the graph convolutional neural network learns the feature representations of the nodes through multiple convolutional layers. Each layer's operation consists of feature aggregation and feature update. Finally, the link prediction function of the graph convolutional network is used to predict the probability of a link from a user node to a new information node, as a prediction of the user's likelihood of reading the information. This captures the user's potential indirect interest associations with the information.

[0090] Figure 4 This is a model diagram of an optional graph convolutional network model according to an embodiment of the present invention, such as... Figure 4 As shown, the heterogeneous graph modeled in the above steps is input, and then information aggregation of neighboring nodes is achieved through multi-layer stacking. In the first layer, information of direct neighbors (first-order neighbors) is aggregated; in the second layer, information of neighbors' neighbors (second-order neighbors) is aggregated, and so on. Through multi-layer stacking of convolutions, information from all nodes is aggregated, capturing the dependencies of more distant nodes. In practice, this can be understood as user A reading news X, news X being similar to news Y in category and length, and news Y belonging to the same field as news Z. Therefore, the probability that user A read news Z can be predicted. The calculation method is as follows: assuming the input node feature matrix, after one layer, the formula for calculating the output feature matrix is: ;in , It is an identity matrix, and by adding self-loops, the nodes include their own information during aggregation. yes The degree matrix is ​​calculated, and the adjacency matrix is ​​symmetrically normalized to avoid excessively large eigenvalues ​​due to differences in node degree. It is a learnable weight matrix used for feature dimension transformation. As the activation function, this embodiment can use the ReLU function to introduce nonlinearity. This embodiment can leverage the mining capabilities of graph convolutional networks to uncover the user's potential indirect interests.

[0091] Step S105: Generate a multi-dimensional user profile based on the first feature, the second feature, and the third feature, and generate an information recommendation scheme based on the multi-dimensional user profile.

[0092] Specifically, multi-dimensional user profiles are achieved through comprehensive analysis and fusion of the first feature (non-linear interaction features between users and information), the second feature (temporal changes in user interests), and the third feature (potential indirect interests of users). These features are extracted from three perspectives: deep learning, temporal analysis, and graph structure analysis, respectively, to explore the evolution of users' reading preferences and interests, thus forming user profiles.

[0093] Further steps in generating a news recommendation scheme involve matching the generated user profile with the currently recommended news database to find the news list that best matches the user's current interests and needs. Specifically, this includes: calculating the similarity between the user profile and the news to be recommended, which can be based on Euclidean distance, cosine similarity, or other distance / similarity measures; for potential indirect interest features, additional graph algorithms or deep learning models are needed to calculate the indirect similarity between news items or between users and news items; ranking the news items in the database according to the similarity score, and then selecting the top candidate news items as the recommendation list, which may also include filtering out previously read news, avoiding duplicate recommendations, and considering the timeliness of the news; dynamically adjusting the recommendation strategy based on user feedback (such as clicks, reading time, favorites, etc.) to ensure that the recommendation system can adapt to changes in user interests and continuously provide high-quality personalized recommendations.

[0094] In one optional approach, the steps of generating an information recommendation scheme based on a multi-dimensional user profile include: determining the information content category of the information to be recommended; calculating the target user's reading completion rate score for each information content category based on the multi-dimensional user profile; sorting the information to be recommended for each information content category according to the reading completion rate to obtain an information recommendation sequence; and generating an information recommendation scheme based on the information recommendation sequence.

[0095] It should be noted that in the aforementioned steps, this embodiment models the non-linear characteristics of user-information interaction, the temporal characteristics of user reading behavior, and the characteristics of potential indirect user interests from the data features of user historical behavior data and information content. In practice, this can be understood as extracting the parts of deep interest (non-linear characteristics) from the user's past behavior, such as interest in long articles in the technology field, as a recommendation indicator. However, even when interested in the same type of information, a user's interests a week ago differ from their recent interests, requiring modeling of temporal characteristics to obtain information the user is more likely to be interested in recently. Furthermore, it is necessary to consider the user's potential indirect interest characteristics; for example, different research reports on the same event may present different information, and a user interested in one report may also be more inclined to read other research reports on the same event. Through these steps, a deep user behavior profile is mined, enabling a better prediction of a user's level of interest in a particular piece of information. The final output of the reading completion rate is used to characterize the user's estimated level of interest in a piece of information. After fusing features, deep learning and other methods are used to predict the final reading completion rate for each piece of information. The prediction results are then ranked, with information with higher predicted reading completion rates being recommended first.

[0096] Through steps S101 to S105, user features and content features can be constructed based on the target user's historical behavior data and information content data. The user features include at least the reading completion rate, and the content features include at least the information content category. Then, the user features and content features are processed by a neural collaborative filtering model to obtain the first feature, which is used to characterize the interaction between the target user and the information content. Next, the historical reading sequence, the corresponding reading completion rate, and the information content category in the historical behavior data are processed by a long short-term memory network model to obtain the second feature, which is used to reflect the temporal changes in user interests. Then, the historical behavior data and information content data are processed by a graph convolutional network model to obtain the third feature, which is used to characterize the user's potential indirect interests. Finally, a multi-dimensional user profile is generated based on the first, second, and third features, and an information recommendation scheme is generated based on the multi-dimensional user profile.

[0097] In this embodiment of the invention, a method of deeply mining the multi-dimensional features of user reading behavior and information content is adopted. By integrating the innovative methods of neural collaborative filtering model, long short-term memory network model and graph convolutional network model, the goal of comprehensively understanding user interests and preferences is achieved. This results in a significant improvement in the personalization level and technical accuracy of the information recommendation system, thereby solving the technical problem of low information recommendation accuracy in related technologies.

[0098] Specifically, this invention first constructs user and content features using the target user's historical behavior data and information content data. User features focus on reading completion rate, while content features focus on information classification systems. Then, a neural collaborative filtering model processes the user and content features to uncover non-linear interaction relationships and capture the deep association between users and information content, forming the first feature. Next, a long short-term memory network model is used to perform temporal analysis on the user's reading sequence, combining reading completion rate and information category to extract the second feature reflecting the dynamic changes in user interests, enhancing the timeliness and adaptability of the recommendation scheme. Finally, a graph convolutional network model is used to analyze links on the heterogeneous graph of user-information, revealing potential indirect interest points of users to obtain the third feature, expanding the scope of the recommendation system. Ultimately, the multi-dimensional user profile constructed based on these three features becomes the core of generating customized, high-precision information recommendation schemes, ensuring that information recommendations not only meet users' immediate needs but also reveal their potential interests. This fundamentally solves the problem of low accuracy in information recommendations caused by single user profiles and insufficient capture of interest changes in related technologies.

[0099] The invention will now be described in conjunction with another alternative embodiment.

[0100] Example 2

[0101] The information recommendation device provided in this embodiment includes multiple implementation units, each of which corresponds to a specific implementation step in Embodiment 1 above.

[0102] Figure 5 This is a schematic diagram of an optional information recommendation device according to an embodiment of the present invention, such as... Figure 5 As shown, the device may include: a construction unit 51, a first processing unit 52, a second processing unit 53, a third processing unit 54, and a generation unit 55.

[0103] The construction unit 51 is used to construct user features and content features based on the target user's historical behavior data and information content data. The user features include at least the reading completion rate, and the content features include at least the information content category.

[0104] The first processing unit 52 is used to process user features and content features through a neural collaborative filtering model to obtain a first feature, wherein the first feature is used to characterize the interaction between the target user and the information content.

[0105] The second processing unit 53 is used to process the historical reading sequence, corresponding reading completion rate and information content category in the historical behavior data through a long short-term memory network model to obtain a second feature, wherein the second feature is used to reflect the temporal changes of user interests.

[0106] The third processing unit 54 is used to process historical behavior data and information content data through a graph convolutional network model to obtain a third feature, wherein the third feature is used to characterize the user's potential indirect interests.

[0107] The generation unit 55 is used to generate a multi-dimensional user profile based on the first feature, the second feature, and the third feature, and to generate an information recommendation scheme based on the multi-dimensional user profile.

[0108] The aforementioned information recommendation device can first construct user features and content features based on the target user's historical behavior data and information content data through the construction unit 51. The user features include at least the reading completion rate, and the content features include at least the information content category. Then, the first processing unit 52 processes the user features and content features through a neural collaborative filtering model to obtain the first feature, which is used to characterize the interaction between the target user and the information content. Next, the second processing unit 53 processes the historical reading sequence, the corresponding reading completion rate, and the information content category in the historical behavior data through a long short-term memory network model to obtain the second feature, which is used to reflect the temporal changes in user interests. Then, the third processing unit 54 processes the historical behavior data and information content data through a graph convolutional network model to obtain the third feature, which is used to characterize the user's potential indirect interests. Finally, the generation unit 55 generates a multi-dimensional user profile based on the first, second, and third features, and generates an information recommendation scheme based on the multi-dimensional user profile.

[0109] In this embodiment of the invention, a method of deeply mining the multi-dimensional features of user reading behavior and information content is adopted. By integrating the innovative means of neural collaborative filtering model, long short-term memory network model and graph convolutional network model, the goal of comprehensively understanding user interests and preferences is achieved. This results in a significant improvement in the personalization level and technical accuracy of the information recommendation system, thereby solving the technical problem of low information recommendation accuracy in related technologies.

[0110] Specifically, this invention first constructs user and content features using the target user's historical behavior data and information content data. User features focus on reading completion rate, while content features focus on information classification systems. Then, a neural collaborative filtering model processes the user and content features to uncover non-linear interaction relationships and capture the deep association between users and information content, forming the first feature. Next, a long short-term memory network model is used to perform temporal analysis on the user's reading sequence, combining reading completion rate and information category to extract the second feature reflecting the dynamic changes in user interests, enhancing the timeliness and adaptability of the recommendation scheme. Finally, a graph convolutional network model is used to analyze links on the heterogeneous graph of user-information, revealing potential indirect interest points of users to obtain the third feature, expanding the scope of the recommendation system. Ultimately, the multi-dimensional user profile constructed based on these three features becomes the core of generating customized, high-precision information recommendation schemes, ensuring that information recommendations not only meet users' immediate needs but also reveal their potential interests. This fundamentally solves the problem of low accuracy in information recommendations caused by single user profiles and insufficient capture of interest changes in related technologies.

[0111] Furthermore, the construction unit includes: a first acquisition module, used to acquire N pieces of information content data and classify each piece of information content data according to a preset classification standard to obtain information content categories, where N is a positive integer; and a first generation module, used to generate classification tags based on information content categories and generate content features based on information content data and corresponding classification tags.

[0112] Furthermore, the construction unit also includes: a second acquisition module, used to acquire M historical behavior data, wherein the historical behavior data includes the target user's historical reading information, reading timestamps, and reading duration, where M is a positive integer; a first calculation module, used to perform anomaly filtering on historical reading information based on reading duration and calculate the reading completion rate corresponding to each historical reading information; a first sorting module, used to sort the historical reading information based on reading timestamps to obtain a historical reading sequence; and an association module, used to associate each historical reading information with the target user's user identifier and the reading completion rate corresponding to the historical reading information in the historical reading sequence to obtain user characteristics.

[0113] Further, the first processing unit includes: a first input module, used to input user features and content features into a neural collaborative filtering model, wherein the neural collaborative filtering model includes: an input layer, an embedding layer, a collaborative filtering layer, and an output layer; an encoding module, used to encode the user features and content features respectively through the input layer to obtain corresponding user sparse variables and content sparse variables; a mapping module, used to map the user sparse variables and content sparse variables to a vector space respectively through the embedding layer to obtain user latent vectors and content latent vectors; a concatenation module, used to concatenate the user latent vectors and content latent vectors and transmit them to the collaborative filtering layer, where the multilayer perceptron in the collaborative filtering layer calculates the nonlinear interaction pattern between the user and the information content to obtain nonlinear features; and an output module, used to output the nonlinear features as a first feature through the output layer.

[0114] Further, the second processing unit includes: a second input module, used to input historical reading sequences, user features, and content features into a long short-term memory network model, wherein the long short-term memory network model includes at least an embedding layer and a long short-term memory network layer; a conversion module, used to convert each historical reading information in the historical reading sequence into an embedding vector through the embedding layer, and determine the reading completion rate corresponding to each historical reading information as a vector feature of the embedding vector to obtain an embedding feature sequence; and an analysis module, used to perform time-series analysis on the embedding feature sequence through the long short-term memory network layer to obtain a second feature, wherein the long short-term memory network layer includes at least a forgetting gate, the forgetting gate is used to filter historical reading information with a reading completion rate less than a first preset threshold, and to filter the historical reading information in the embedding feature sequence.

[0115] Furthermore, the third processing unit includes: a construction module for constructing a heterogeneous graph based on historical behavior data and information content data, wherein the heterogeneous graph contains user nodes and information nodes; a first setting module for setting a first type of weight, namely the user's reading completion rate of information content, for the edges between user nodes and information nodes in the heterogeneous graph; a second setting module for setting a second type of weight, namely information similarity, for the edges between information nodes and other information nodes in the heterogeneous graph; a third input module for inputting the heterogeneous graph into a graph convolutional network model, performing link analysis on the heterogeneous graph through convolutional layers to obtain the link probability from user nodes to each information node; and a second generation module for generating a third feature by using user nodes and information nodes as link objects and combining the link probabilities.

[0116] Furthermore, the generation unit includes: a determining module for determining the content category of the information to be recommended; a second calculation module for calculating the reading completion rate score of the target user for each content category based on a multi-dimensional user profile; a second sorting module for sorting the information to be recommended for each content category according to the reading completion rate to obtain an information recommendation sequence; and a third generation module for generating an information recommendation scheme based on the information recommendation sequence.

[0117] The aforementioned information recommendation device may also include a processor and a memory. The aforementioned building unit 51, first processing unit 52, second processing unit 53, third processing unit 54, generation unit 55, etc. are all stored in the memory as program units, and the processor executes the aforementioned program units stored in the memory to realize the corresponding functions.

[0118] The aforementioned processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured. By adjusting kernel parameters, user and content features are constructed based on the target user's historical behavior data and information content data. A neural collaborative filtering model processes these user and content features, a long short-term memory network model processes historical reading sequences, corresponding reading completion rates, and information content categories from the historical behavior data, and a graph convolutional network model processes the historical behavior data and information content data to generate a multi-dimensional user profile and information recommendation scheme.

[0119] The aforementioned memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0120] This application also provides a computer program product, which, when executed on a data processing device, is suitable for executing an initialization program with the following method steps: constructing user features and content features based on the target user's historical behavior data and information content data, wherein the user features include at least a reading completion rate, and the content features include at least an information content category; processing the user features and content features through a neural collaborative filtering model to obtain a first feature, wherein the first feature is used to characterize the interaction between the target user and the information content; processing the historical reading sequence, the corresponding reading completion rate, and the information content category in the historical behavior data through a long short-term memory network model to obtain a second feature, wherein the second feature is used to reflect the temporal changes in user interests; processing the historical behavior data and information content data through a graph convolutional network model to obtain a third feature, wherein the third feature is used to characterize the user's potential indirect interests; generating a multi-dimensional user profile based on the first feature, the second feature, and the third feature, and generating an information recommendation scheme based on the multi-dimensional user profile.

[0121] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute any of the information recommendation methods in the first embodiment described above.

[0122] Specifically, the information recommendation method includes: constructing user features and content features based on the target user's historical behavior data and information content data, wherein the user features include at least the reading completion rate, and the content features include at least the information content category; processing the user features and content features through a neural collaborative filtering model to obtain a first feature, wherein the first feature is used to characterize the interaction between the target user and the information content; processing the historical reading sequence, corresponding reading completion rate, and information content category in the historical behavior data through a long short-term memory network model to obtain a second feature, wherein the second feature is used to reflect the temporal changes in user interests; processing the historical behavior data and information content data through a graph convolutional network model to obtain a third feature, wherein the third feature is used to characterize the user's potential indirect interests; generating a multi-dimensional user profile based on the first feature, second feature, and third feature, and generating an information recommendation scheme based on the multi-dimensional user profile.

[0123] According to another aspect of the present invention, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the information recommendation method of any one of the above embodiments.

[0124] Specifically, the information recommendation method includes: constructing user features and content features based on the target user's historical behavior data and information content data, wherein the user features include at least the reading completion rate, and the content features include at least the information content category; processing the user features and content features through a neural collaborative filtering model to obtain a first feature, wherein the first feature is used to characterize the interaction between the target user and the information content; processing the historical reading sequence, corresponding reading completion rate, and information content category in the historical behavior data through a long short-term memory network model to obtain a second feature, wherein the second feature is used to reflect the temporal changes in user interests; processing the historical behavior data and information content data through a graph convolutional network model to obtain a third feature, wherein the third feature is used to characterize the user's potential indirect interests; generating a multi-dimensional user profile based on the first feature, second feature, and third feature, and generating an information recommendation scheme based on the multi-dimensional user profile.

[0125] Figure 6This is a structural block diagram of an electronic device for performing an information recommendation method according to an embodiment of the present invention, such as... Figure 6 As shown, the electronic device may include: one or more ( Figure 6 (Only one is shown) Processor 602, memory 604, memory controller, and peripheral interface, wherein the peripheral interface is connected to the radio frequency module, audio module and display.

[0126] The memory can be used to store software programs and modules, such as the program instructions / modules corresponding to the information recommendation method and apparatus in this application embodiment. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby realizing the aforementioned information recommendation method. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0127] Specifically, the information recommendation method includes: constructing user features and content features based on the target user's historical behavior data and information content data, wherein the user features include at least the reading completion rate, and the content features include at least the information content category; processing the user features and content features through a neural collaborative filtering model to obtain a first feature, wherein the first feature is used to characterize the interaction between the target user and the information content; processing the historical reading sequence, corresponding reading completion rate, and information content category in the historical behavior data through a long short-term memory network model to obtain a second feature, wherein the second feature is used to reflect the temporal changes in user interests; processing the historical behavior data and information content data through a graph convolutional network model to obtain a third feature, wherein the third feature is used to characterize the user's potential indirect interests; generating a multi-dimensional user profile based on the first feature, second feature, and third feature, and generating an information recommendation scheme based on the multi-dimensional user profile.

[0128] Those skilled in the art will understand that Figure 6 The structure shown is for illustrative purposes only. Electronic devices can also be smartphones, tablets, handheld computers, mobile internet devices (MIDs), PADs, and other terminal devices. Figure 6 This does not limit the structure of the aforementioned electronic device. For example, electronic devices may also include components that are more... Figure 6 The more or fewer components shown (such as network interfaces, display devices, etc.), or having the same Figure 6 The different configurations shown.

[0129] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0130] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0131] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0132] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0133] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0134] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0135] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0136] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An information recommendation method, characterized in that, include: User features and content features are constructed based on the target user's historical behavior data and information content data, wherein the user features include at least the reading completion rate, and the content features include at least the information content category; The user features and content features are processed by a neural collaborative filtering model to obtain a first feature, wherein the first feature is used to characterize the interaction between the target user and the information content. The historical reading sequence, the corresponding reading completion rate, and the information content category in the historical behavior data are processed by a long short-term memory network model to obtain a second feature, wherein the second feature is used to reflect the temporal changes in user interests. The historical behavior data and the information content data are processed by a graph convolutional network model to obtain a third feature, wherein the third feature is used to characterize the user's potential indirect interests. A multi-dimensional user profile is generated based on the first feature, the second feature, and the third feature, and an information recommendation scheme is generated based on the multi-dimensional user profile.

2. The information recommendation method according to claim 1, characterized in that, The steps for constructing user and content characteristics based on the target user's historical behavior data and information content data include: Obtain N pieces of information content data, and classify each piece of information content data according to a preset classification standard to obtain the information content category, where N is a positive integer; Category tags are generated based on the information content categories, and content features are generated based on the information content data and the corresponding category tags.

3. The information recommendation method according to claim 1, characterized in that, The steps of constructing user and content features based on the target user's historical behavior data and information content data also include: Obtain M historical behavior data points, wherein the historical behavior data includes the target user's historical reading information, reading timestamps, and reading duration, and M is a positive integer; Based on the reading time, the historical reading information is filtered for anomalies, and the reading completion rate corresponding to each historical reading information is calculated; The historical reading information is sorted based on the reading timestamp to obtain a historical reading sequence; In the historical reading sequence, each piece of historical reading information is associated with the target user's user identifier and the reading completion rate corresponding to the historical reading information to obtain the user characteristics.

4. The information recommendation method according to claim 1, characterized in that, The step of processing the user features and content features using a neural collaborative filtering model to obtain the first feature includes: The user features and the content features are input into the neural collaborative filtering model, wherein the neural collaborative filtering model comprises: an input layer, an embedding layer, a collaborative filtering layer, and an output layer; The user features and content features are encoded by the input layer to obtain corresponding user sparse variables and content sparse variables. The embedding layer maps the user sparse variables and the content sparse variables to the vector space respectively, to obtain the user latent vector and the content latent vector. The user latent vector and the content latent vector are concatenated and transmitted to the collaborative filtering layer. The multilayer perceptron in the collaborative filtering layer calculates the nonlinear interaction pattern between the user and the information content to obtain nonlinear features. The nonlinear feature is output as the first feature through the output layer.

5. The information recommendation method according to claim 1, characterized in that, The step of processing the historical reading sequences, corresponding reading completion rates, and information content categories in the historical behavior data using a long short-term memory network model to obtain the second feature includes: The historical reading sequence, the user features, and the content features are input into the long short-term memory network model, wherein the long short-term memory network model includes at least: an embedding layer and a long short-term memory network layer; The embedding layer converts each piece of historical reading information in the historical reading sequence into an embedding vector, and the reading completion rate corresponding to each piece of historical reading information is determined as the vector feature of the embedding vector, thus obtaining an embedding feature sequence. The second feature is obtained by performing time-series analysis on the embedded feature sequence through the long short-term memory network layer. The long short-term memory network layer includes at least a forget gate, which is used to filter historical reading information with a reading completion rate of less than a first preset threshold and filter the historical reading information in the embedded feature sequence.

6. The information recommendation method according to claim 1, characterized in that, The step of processing the historical behavior data and the information content data using a graph convolutional network model to obtain the third feature includes: A heterogeneous graph is constructed based on the historical behavior data and the information content data, wherein the heterogeneous graph includes: user nodes and information nodes; In the heterogeneous graph, for the edge between the user node and the information node, a first type of weight is set as the user's reading completion rate of the information content; In the heterogeneous graph, for the edge between the information node and another information node, a second type of weight is set as information similarity; The heterogeneous graph is input into the graph convolutional network model, and the convolutional layer performs link analysis on the heterogeneous graph to obtain the link probability from the user node to each information node. The third feature is generated by using the user node and the information node as link objects and combining the link probability.

7. The information recommendation method according to claim 1, characterized in that, The steps for generating an information recommendation scheme based on the multi-dimensional user profile include: Determine the content category of the information to be recommended; Based on the multi-dimensional user profile, the reading completion rate score of the target user for each of the information content categories is calculated; The recommended information is sorted according to the reading completion rate for each of the aforementioned information content categories to obtain an information recommendation sequence; The information recommendation scheme is generated based on the information recommendation sequence.

8. An information recommendation device, characterized in that, include: A construction unit is used to construct user features and content features based on the target user's historical behavior data and information content data, wherein the user features include at least the reading completion rate, and the content features include at least the information content category; The first processing unit is used to process the user features and the content features through a neural collaborative filtering model to obtain a first feature, wherein the first feature is used to characterize the interaction between the target user and the information content. The second processing unit is used to process the historical reading sequence, the corresponding reading completion rate and the information content category in the historical behavior data through a long short-term memory network model to obtain a second feature, wherein the second feature is used to reflect the temporal changes of user interests; The third processing unit is used to process the historical behavior data and the information content data through a graph convolutional network model to obtain a third feature, wherein the third feature is used to characterize the user's potential indirect interests. The generation unit is used to generate a multi-dimensional user profile based on the first feature, the second feature, and the third feature, and to generate an information recommendation scheme based on the multi-dimensional user profile.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the information recommendation method according to any one of claims 1 to 7.

10. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the information recommendation method according to any one of claims 1 to 7.

Images