Financial article recommendation method and device and readable storage medium

Abstract

The invention discloses a financial article recommendation method and device and a readable storage medium, and relates to the field of big data, and the method comprises the steps: firstly constructing a financial article content library containing a stock tag and a theme tag, and compressing an article dimension to a preset attribute dimension; obtaining user attribute data and historical interaction data, and constructing interaction characteristics of the user and the tag; processing the interaction characteristics by adopting association rule mining, time sequence prediction, interest attenuation, deep learning and a user self-selected label model, and fusing an output result to predict a target label; and matching the articles based on the target labels, and generating a recommendation list after duplicate removal according to the release time in an inverted order. Through cooperation of a multi-dimensional label system and multiple models, user interest changes are dynamically captured, recommendation accuracy is improved, and the method is suitable for accurate information recommendation scenes in the financial field.

Description

Technical Field

[0001] The present invention relates to the field of big data, and in particular to a method, device and readable storage medium for recommending financial articles. Background Art

[0002] With the explosive growth of financial market information, users face the challenge of sifting through a vast amount of financial articles. Personalized recommendations are crucial for improving information acquisition efficiency. Existing technologies often rely on single-label systems or traditional collaborative filtering algorithms. These suffer from a single label dimension and inadequate capture of dynamic user interests. For example, recommendations based solely on a user's historical clicks struggle to uncover potential connections between labels, fail to account for the time-dependent decay of interest, and lack interaction data for new users, resulting in poor recommendation effectiveness. Furthermore, the financial sector places high demands on label accuracy, and existing models are unable to adequately classify and integrate specialized labels, resulting in poorly relevant and timely recommendations. Summary of the Invention

[0003] The object of the present invention is to provide a method, device and readable storage medium for recommending financial articles.

[0004] In a first aspect, an embodiment of the present invention provides a method for recommending financial articles, comprising: Constructing a financial article content library, which stores financial articles and corresponding stock tags and theme tags. The stock tags and theme tags are obtained by classifying the article content to compress the article dimensions into preset attribute dimensions composed of the stock tags and theme tags. The total number of the preset attribute dimensions is determined by the tag classification system. Obtaining user attribute data and historical interaction data between the user and financial articles in the financial article content library, wherein the user attribute data includes a user's self-selected stock list and a self-selected subject list, and the historical interaction data includes the user's interaction behavior with the financial articles and the corresponding article identifiers and timestamps; Based on the user attribute data and the historical interaction data, constructing interaction features between the user and the stock tags and interaction features between the user and the subject tags, respectively, wherein the interaction features are used to reflect the association relationship and interest level of the user in each stock tag and subject tag; Processing the interaction features using multiple prediction models and fusing the outputs of the models to predict target stock tags and target subject tags of interest to the user, wherein the multiple prediction models include at least an association rule mining model, a time series prediction model, an interest decay model, a deep learning model, and a user-selected tag model; Based on the predicted target stock label and target subject label, corresponding financial articles are matched from the financial article content library, and after deduplication of the matched financial articles in reverse order of publication time, a user-personalized financial article recommendation list is generated and output.

[0005] In a second aspect, an embodiment of the present invention provides a financial article recommendation device, comprising: An acquisition module is configured to construct a financial article content library, wherein the financial article content library stores financial articles and corresponding stock tags and subject matter tags, wherein the stock tags and subject matter tags are obtained by classifying the article content to compress the article dimensions into preset attribute dimensions composed of the stock tags and subject matter tags, wherein the total number of the preset attribute dimensions is determined by a tag classification system; obtain user attribute data and historical interaction data between users and financial articles in the financial article content library, wherein the user attribute data includes a user's self-selected stock list and a self-selected subject matter list, and the historical interaction data includes the user's interaction behavior with the financial articles and the corresponding article identifiers and timestamps; and construct, based on the user attribute data and the historical interaction data, interaction features between the user and the stock tags and interaction features between the user and the subject matter tags, respectively, wherein the interaction features are used to reflect the user's association and interest in each stock tag and subject matter tag; A recommendation module is used to process the interaction features using multiple prediction models and fuse the output results of each model to predict the target stock tags and target subject tags that the user is interested in. The multiple prediction models include at least an association rule mining model, a time series prediction model, an interest decay model, a deep learning model, and a user-selected tag model; based on the predicted target stock tags and target subject tags, corresponding financial articles are matched from the financial article content library, and the matched financial articles are deduplicated in reverse order of publication time to generate and output a user-personalized financial article recommendation list.

[0006] In a third aspect, an embodiment of the present invention provides a readable storage medium, wherein the readable storage medium includes a computer program, and when the computer program runs, the computer device where the readable storage medium is located controls the method described in the first aspect.

[0007] Compared with the existing technology, the beneficial effects provided by the present invention include: using a financial article recommendation method, device and readable storage medium disclosed by the present invention, involving the field of big data, including: first constructing a financial article content library containing stock tags and subject tags, compressing the article dimensions to preset attribute dimensions; obtaining user attribute data and historical interaction data, and constructing interaction features between users and tags; using association rule mining, time series prediction, interest decay, deep learning and user-selected tag models to process the interaction features, and fusing the output results to predict the target tag; matching articles based on the target tag, and generating a recommendation list after deduplication in reverse order of publication time. The present invention dynamically captures changes in user interests and improves recommendation accuracy through the collaboration of a multi-dimensional tag system and multiple models, and is suitable for precise information recommendation scenarios in the financial field. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly describes the drawings required for use in the embodiments. It should be understood that the following drawings illustrate only certain embodiments of the present invention and should not be construed as limiting the scope of the present invention. Those skilled in the art can, without inventive effort, derive other relevant drawings from these drawings.

[0009] Figure 1 A schematic diagram of the steps of the financial article recommendation method provided by an embodiment of the present invention; Figure 2 A schematic diagram of the overall process framework of the financial article recommendation method provided by an embodiment of the present invention; Figure 3 A schematic block diagram of the structure of a financial article recommendation device provided by an embodiment of the present invention; Figure 4 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0010] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the embodiments of the present invention. It should be understood that the described embodiments are only a portion of the embodiments of the present invention, not all of them. Generally, the components of the embodiments of the present invention described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

[0011] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0012] In order to solve the technical problems in the above background technology, Figure 1 This is a flow chart of a method for recommending financial articles provided in an embodiment of the present disclosure. The method for recommending financial articles is introduced in detail below.

[0013] Step S201: Construct a financial article content library. The financial article content library stores financial articles and corresponding stock tags and theme tags. The stock tags and theme tags are obtained by classifying the article content to compress the article dimensions into preset attribute dimensions composed of the stock tags and theme tags. The total number of the preset attribute dimensions is determined by the tag classification system. Step S202: Obtain user attribute data and historical interaction data between the user and financial articles in the financial article content library, wherein the user attribute data includes a user-selected stock list and a user-selected subject list, and the historical interaction data includes the user's interaction behavior with the financial articles and the corresponding article identifiers and timestamps; Step S203: Based on the user attribute data and the historical interaction data, constructing interaction features between the user and the stock tags and interaction features between the user and the subject tags, respectively. The interaction features are used to reflect the user's association and interest in each stock tag and subject tag. Step S204: Processing the interaction features using multiple prediction models and fusing the output results of each model to predict target stock tags and target theme tags of interest to the user. The multiple prediction models include at least an association rule mining model, a time series prediction model, an interest decay model, a deep learning model, and a user-selected tag model. Step S205 , based on the predicted target stock label and target subject label, corresponding financial articles are matched from the financial article content library, and after deduplication of the matched financial articles in reverse order of publication time, a user-personalized financial article recommendation list is generated and output.

[0014] In an embodiment of the present invention, illustratively, the server first executes the step of constructing a financial article content library, which is used to centrally store various financial articles and their corresponding tag information, and converts the high-dimensional content of the articles into computable low-dimensional attributes through tagging processing, providing a data basis for subsequent recommendations.

[0015] In practice, the server collects financial article resources through various legitimate channels, including but not limited to brokerage research reports, listed company announcements, industry analysis minutes, and financial news information. For example, the server retrieved "Commentary on the Third Quarter Financial Report of a Power Battery Company in YYYY" and "In-Depth Research Report on the New Energy Equipment Industry" from the database of a partner financial data service provider, crawled "Announcement of the Annual Performance Forecast of a Listed Company in YYYY" from the official website of the stock exchange, and obtained "Interpretation of the Policy of a Certain Emerging Industry" from the API interface of financial media. When each article is stored, the server assigns it a unique article identifier, such as "REPORT00012345", and records basic information such as the title, publication time, source organization, and text content.

[0016] The core process is to automatically add stock and topic tags to each article. For stock tags, the server uses text parsing technology to identify the stock names and tickers mentioned in the article. For example, in the article "Commentary on the Third Quarter Financial Report of a Power Battery Company in YYYY," it extracts "a certain power battery company (123456.SH)" and matches it with the built-in stock database, ultimately assigning the article the stock tag "a certain power battery company (123456.SH)." For topic tags, the server uses natural language processing technology for automated classification. Specifically, the server uses a pre-trained language model that supports zero-shot or small-shot classification as its foundation. It fine-tunes the model using historical financial industry article data (such as research reports and industry trends from the past three years) to enable it to understand the specialized terminology and semantics of the financial sector. For example, for the article "Analysis of the Application Prospects of a Certain Energy Storage Technology," the fine-tuned model analyzes its core content and determines that it falls under the "energy storage technology" category under the "new energy" sector, thereby assigning the corresponding topic tag.

[0017] Through this tagging process, the server transforms the original textual content of articles into structured attributes consisting of a "stock tag + theme tag." For example, an article about the technological progress of a new energy vehicle company might be tagged with the stock tag "New Energy Vehicle Company (654321.SZ)" and the theme tags "New Energy Vehicles" and "Power Batteries." These tags together constitute the pre-defined attribute dimensions, the total number of which is determined by the tag classification system. For example, stock tags encompass all tradable stocks on the market (approximately thousands to tens of thousands), while theme tags are divided into primary, secondary, and tertiary tags based on common industry classification standards (e.g., "Power Equipment," "New Energy Passenger Vehicles," "Intelligent Driving," etc., totaling hundreds to thousands). The server stores these tagged articles in a unified financial article repository, typically using a relational database to store structured information such as tags and a distributed file system to store unstructured data such as the article body, facilitating efficient subsequent query and access.

[0018] In order to achieve personalized recommendations, the server needs to collect and integrate user-related data, mainly including attribute data actively provided by users and historical interaction data generated by users on the platform.

[0019] User attribute data is primarily obtained through user registration and system setup. For example, when a user first uses the recommendation platform, they are required to provide information such as their institution (e.g., "asset management company"), occupation type (e.g., "investment manager"), and trading role (e.g., "buyer" or "seller"). More importantly, users can manually add lists of selected stocks and topics to their wishlists, directly expressing their interests. For example, User A, an investment manager focused on the new energy sector, might add "a power battery company (123456.SH)" and "a new energy vehicle company (654321.SZ)" to their wishlists, and "intelligent manufacturing" and "energy storage technology" to their wishlists. This data is encrypted and stored in the user information database on the server, serving as a key basis for understanding users' explicit interests.

[0020] Historical user interaction data is a record of user behavior while using the platform. The server captures this data in real time using front-end tracking technology. This data includes user clicks, reads, favorites, shares, and comments on financial articles, along with the article ID and timestamp corresponding to each action. For example, if user A clicks and reads the article titled "Minutes of the 2019 New Energy Industry Chain Summit" at HH:MM on MM / DD / YYYY, the server will record the user ID, article ID, action type ("read"), and time of occurrence ("YYYY-MM-DDHH:MM:SS"), and associate it with the article's stock tag (e.g., "a certain power battery company," "a certain new energy vehicle company") and topic tag (e.g., "intelligent manufacturing"). This interaction data is aggregated daily into the data warehouse. After deduplication and cleaning (e.g., removing accidental clicks), a user behavior sequence is formed, reflecting the user's implicit interests and preferences.

[0021] Based on the collected user attribute data and historical interaction data, the server needs to further construct the interaction features between users and stock tags, and users and subject matter tags, and convert the raw data into a mathematical representation that can quantify the degree of user interest.

[0022] When constructing interaction features between users and stock tags, the server first integrates the user's historical interaction behavior and selected stock information. For example, for user A, the server will count the frequency of their interaction with each stock tag over the past three months: suppose they clicked on articles related to "a certain power battery company" five times and saved them twice, and clicked on articles related to "a certain new energy vehicle company" three times and shared them once. At the same time, the server assigns different weights to different interaction behaviors (e.g., saving a tag has a higher weight than clicking, and sharing has a higher weight than saving a tag), and performs attenuation based on the time when the interaction occurred—recent interactions have a greater impact on current interests, while long-term effects gradually weaken. For example, a user's saving of "a certain power battery company" a week ago would have a higher weight than a click from a month ago. Furthermore, stock tags in the user's selected stock list are given an additional high weight to highlight the user's explicit interests. Through the above calculations, the server ultimately generates a vector for the user. Each dimension of the vector corresponds to a stock label, and the value represents the user's interest in that label. For example, the stock interaction feature vector for user A might be "A certain power battery company: 8.5 points, a certain new energy vehicle company: 6.2 points, a certain semiconductor company: 3.1 points..." The logic for constructing interaction features between users and topic tags is similar to that for stock tags. The server counts the frequency of user interactions with each topic tag (e.g., reading an article on "Smart Manufacturing" eight times), the weight of the interaction type (e.g., adding the article to favorites twice), and combines this with time decay and the weighting of the selected topic list to generate a topic interaction feature vector. For example, "Smart Manufacturing: 9.3 points, Energy Storage Technology: 5.7 points, New Energy Passenger Vehicles: 4.8 points..." These interaction feature vectors directly reflect the user's associations and interest in different tags and serve as the core input for subsequent prediction models.

[0023] The server uses multiple prediction models to comprehensively process the interaction features and fuses the output results of each model to accurately predict the target stock tags and target theme tags that the user is interested in. The following is a detailed explanation using the case of user A: Association rule mining model: This model aims to discover potential associations between stock tags or theme tags in a user's historical interactions. For example, the server analyzes user A's past stock browsing history and finds that after following "a certain new energy vehicle company," they are likely to continue following "a certain auto parts company" (both companies are in the new energy vehicle industry chain). This model then mines the association rule "a certain new energy vehicle company → a certain auto parts company." Based on this rule, if a user recently interacted with "a certain new energy vehicle company," the model will recommend "a certain auto parts company" as a stock tag of potential interest.

[0024] Time Series Prediction Model: This model is primarily used to predict short-term trending tags, particularly for new users in cold-start scenarios (when new users have no historical interaction data). The server counts the daily visits to each stock tag or theme tag over a period of time (e.g., the past 14 days) and uses the time series prediction model to predict popularity trends over the next few days. For example, if the model predicts that "a leading consumer goods company" or "an industrial automation company" will become trending stock tags in the coming day, the server will recommend these popular tags to new users or as supplementary recommendations for existing users.

[0025] Interest decay model: This model simulates how user interests change over time and dynamically adjusts the weight of historical interactions using an "interest retention factor." The server divides the user's historical interaction data into multiple windows (hourly granularity for recent windows and daily granularity for future windows) and calculates the degree of overlap between the user's interacted tags within each window and the current window's tag set (the ratio of the number of elements in the intersection to the total number of elements in the historical windows). This is the interest retention factor. For example, if a user interacted frequently with the "photovoltaics" tag three months ago, the interest retention factor would be high, and the tag's weight would be retained. Conversely, if there has been no interaction for a long time, the weight would gradually decrease.

[0026] Deep Learning Model (Deep Interest Network): This model uses a neural network to capture dynamic changes in user interests. The server inputs user attribute features (such as organization and occupation), behavior sequence features (such as the time series of clicks and favorites), and self-selected tag features (such as the encoding vector of self-selected stocks) into the deep learning model. The model uses an attention mechanism to dynamically weight user behavior features at different time points. For example, a user's recent multiple clicks on "intelligent manufacturing" will be given higher attention. This generates a feature vector that reflects the user's current interest and outputs the user's interest probability for each tag. For example, the model predicts a user's interest probability of 0.89 for "a certain power battery company" and a 0.76 for "intelligent manufacturing."

[0027] User-selected tag model: This model directly utilizes the user's manually added list of selected stocks and topics, and recommends the tags within them as high-priority items. For example, "a certain power battery company" and "a certain new energy vehicle company" in User A's list of selected stocks will be directly added to the candidate tag list.

[0028] Model Fusion: The server assigns weights to each model based on its historical prediction accuracy (the higher the accuracy, the greater the weight). The server then performs a weighted fusion ranking of the candidate tags output by the models, removing duplicates. For example, the association rule mining model has an accuracy of 0.85 and is given the highest weight; the self-selected model has an accuracy of 0.78 and is the next highest weight. After fusion, User A's target stock tags are ranked as "a power battery company, a new energy vehicle company, an auto parts company, a leading consumer company, and a semiconductor company," and the target theme tags are ranked as "intelligent manufacturing, energy storage technology, and new energy passenger vehicles."

[0029] Based on the predicted target stock labels and target subject labels, the server matches corresponding financial articles from the financial article content library, and generates the final personalized recommendation list after screening and sorting.

[0030] Specifically, the server searches the financial article database based on target stock tags (such as "a certain power battery company" and "a certain new energy vehicle company") and target theme tags (such as "intelligent manufacturing"), matching articles containing all of these tags or at least one core tag. For example, user A is matched with articles such as "Commentary on a Certain Power Battery Company's Q3 Financial Report" (containing the stock tag "a certain power battery company"), "Intelligent Manufacturing Industry Outlook for YYYY" (containing the theme tag "intelligent manufacturing"), and "Analysis of the Progress of Cooperation between a Certain New Energy Vehicle Company and a Certain Power Battery Company" (containing the stock tags "a certain new energy vehicle company" and "a certain power battery company" and the theme tag "new energy passenger vehicles").

[0031] After matching, the server deduplicates the articles (eliminating duplicate records of the same article) and sorts them by publication time, ensuring that users see the latest content first. For example, the sorted articles might be: "Smart Manufacturing Industry Outlook for YYYY" published on MM / DD / YYYY, "Analysis of the Progress of Cooperation between a New Energy Vehicle Company and a Power Battery Company" published on MM / DD / YYYY, and "Commentary on the Q3 Financial Report of a Power Battery Company" published on MM / DD / YYYY.

[0032] Finally, the server sets a refresh interval for the recommendation list based on article length and the user's average reading speed (based on historical user reading data, e.g., an average reading speed of 1,000 words per minute). For example, if the current recommendation list contains five articles totaling approximately 5,000 words, with an estimated reading time of 5 minutes, the server will automatically refresh the list after 10 minutes to add the new articles. Finally, the sorted article list is presented to the user through the client interface, completing the personalized financial article recommendation process.

[0033] Through the above method, the server can comprehensively consider the user's explicit interests and implicit behaviors, combine the advantages of multiple models, and achieve accurate and real-time financial article recommendations, effectively improving the user's reading experience and information acquisition efficiency.

[0034] In an embodiment of the present invention, the article content is classified into subject tags, including: Use a pre-trained model that supports zero-shot or small-shot classification as the base model; Collect financial industry article corpus, clean the financial industry article corpus and expand the sample. The sample expansion includes supplementing related articles through keyword search for low-frequency subject tags; Generate labels for training samples using a multi-model voting mechanism that includes the outputs of at least two pre-trained language models and keyword rule matching results; Fine-tune the base model on the cleaned and sample-expanded corpus of financial industry articles to obtain a subject label classification model; The fine-tuned subject label classification model is used to classify financial articles and output candidate labels; the candidate labels are re-ranked using the refined ranking model and the final subject label is output.

[0035] In the financial article recommendation method, classifying article content by subject tags is the core step in building the content library. The server performs the following process: The server first uses a pre-trained language model that supports zero-shot or small-shot classification as its base model. Because the financial sector is rife with niche topics (such as "green finance" and "supply chain finance"), some low-frequency topics lack sample size. Zero-shot / small-shot models can adapt to low-sample scenarios through transfer learning. Therefore, the server selects a pre-trained model based on the Transformer architecture (such as a general language model that supports cross-domain transfer). This model has been pre-trained on a large-scale general text corpus and possesses basic semantic understanding capabilities.

[0036] Next, the server collects, cleans, and expands the corpus of financial industry articles. The server obtains a batch of financial articles from the past five years from compliant financial databases, institutional research reports, and financial information platform interfaces. These articles range from research reports to industry reports and policy interpretations, totaling one million original articles. The server cleans the raw corpus to remove noisy data: using a text deduplication algorithm (based on SimHash) to remove articles with a duplication rate exceeding 95%, filtering out short texts under 500 words to prevent insufficient information, and using regular expressions to remove formatting noise such as HTML tags and special symbols, retaining 800,000 valid articles. For low-frequency subject labels (such as "carbon trading financial instruments", "cross-border digital RMB settlement" and other labels with less than 10 samples), the server supplements related articles through keyword search: using the label name and core keywords (such as "carbon trading" + "financial instruments", "carbon futures") as search conditions, search for articles in the company's internal document library whose titles or abstracts contain these keywords, and select the top 20 articles based on text similarity (cosine similarity ≥ 0.7) to supplement the corresponding subject, so that the sample size of each subject label reaches more than 30 articles to ensure balanced model training.

[0037] The server then uses a multi-model voting mechanism to generate labels for training samples. To ensure label accuracy, the server uses two pre-trained language models with different architectures (Model X and Model Y) and a keyword rule system to predict labels for the cleaned corpus. For example, for the article "Analysis of the Policy Impact of the 2024 Carbon Trading Market," Model X outputs the labels "carbon finance" and "green finance," while Model Y outputs the labels "carbon finance" and "environmental policy." The keyword rule system matches keywords such as "carbon trading" and "policy" to output the labels "carbon finance" and "policy research." The server sets a voting rule: when at least two predictions (models or rules) agree, that label is selected as the final label for the sample. In this example, "carbon finance" is predicted by all three results and is therefore determined to be the article's subject label. "Green finance," "environmental policy," and "policy research" are excluded because they were predicted by only a single model or rule. Through this mechanism, the server generates high-quality training labels for 800,000 articles, achieving a label accuracy of 92%.

[0038] After preparing the training samples, the server fine-tuned the base model on the cleaned and expanded corpus of financial industry articles, resulting in a subject label classification model. The server partitioned the 800,000 labeled articles into a training set (640,000 articles) and a validation set (160,000 articles) at an 8:2 ratio. Using the labels as the classification target, the server used the training set to fine-tune the base model parameters: an initial learning rate of 2e-5, a gradient descent optimizer, and five training epochs. The validation set was used to evaluate model accuracy after each epoch. When the validation set accuracy stopped improving for two consecutive epochs (e.g., 0.88 in the third epoch, 0.87 in the fourth epoch, and 0.88 in the fifth epoch), the server stopped training and saved the model parameters from the third epoch as the final subject label classification model. The model now achieved an accuracy of 0.88 on the validation set, meeting the classification requirements.

[0039] Finally, the server uses a fine-tuned subject label classification model to classify financial articles, outputting candidate labels, and re-ranking them using a refined ranking model to obtain the final labels. For newly added financial articles (such as the "Progress Report on the Digital RMB Cross-border Payment Pilot Program"), the server first extracts the article body text and inputs it into the fine-tuned subject label classification model. The model then outputs the top five candidate labels and their corresponding confidence scores, such as "digital RMB" (0.92), "cross-border payment" (0.85), "financial technology" (0.78), "digital currency" (0.72), and "payment and settlement" (0.65). To further improve label relevance, the server uses a refined ranking model to re-rank the candidate labels. The refined ranking model uses the cosine similarity between the article text vector and the candidate label vector as a feature, combined with the timeliness weight of the label in the financial field (for example, "digital RMB" is a recent hot topic, so its weight is increased by 10%), to score and rank the candidate labels, ultimately outputting the top three labels as the article's final subject label: "digital RMB," "cross-border payment," and "financial technology."

[0040] Through the above process, the server realizes the automatic and high-precision classification of financial article subject labels, providing a reliable labeling basis for subsequent recommendations.

[0041] In the embodiment of the present invention, the association rule mining model is the FP-Growth model. The process of mining association rules using the FP-Growth model can be implemented through the following examples.

[0042] Taking users as units, based on the historical interaction data, extracting the browsing sequence of stock tags or theme tags of users within a preset time period; Dividing the browsing sequence into multiple subsequences according to a preset sequence length threshold and an adjacent interaction time interval threshold; Calculate the high-frequency factor of each stock tag or theme tag every day, where the high-frequency factor is the ratio of the total number of user interactions for the stock tag or theme tag on that day to the total number of online users on that day; High-frequency items in the multiple subsequences are eliminated based on the high-frequency factor, and the elimination rule is as follows: if the high-frequency factor of a stock label or theme label is greater than a first preset threshold, and the difference between the high-frequency factor of the stock label or theme label and the high-frequency factor of the adjacent stock label or theme label is greater than a second preset threshold, then it is determined to be a high-frequency item and eliminated; A transaction data set is constructed based on the subsequences after removing high-frequency items, and frequent item sets are mined through FP-tree. Association rules are generated based on the frequent item sets. The mining parameters of the association rules include a minimum support value set based on the total number of sequences and a preset minimum confidence value. The association rules are used to predict stock tags or subject tags that the user is interested in.

[0043] In an embodiment of the present invention, for example, in a financial article recommendation method, an association rule mining model (FP-Growth model) is used to mine potential associations of tags in user historical interactions. The server executes the following process: The server first extracts tag browsing sequences based on historical interaction data, taking each user as a unit. This historical interaction data includes user clicks, readings, and other behaviors on financial articles, as well as the corresponding stock / theme tags. The server groups the interactions by user ID, selects interaction records within a preset time period (e.g., the past 30 days), and sorts them in ascending order by timestamp. It then extracts the stock tag or theme tag associated with each interaction, forming the user's tag browsing sequence. For example, if user A has interacted with articles containing "Stock Tag A," "Theme Tag B," "Stock Tag C," "Stock Tag A," and "Theme Tag D" in the past 30 days, the server will extract their stock tag browsing sequence as [Stock Tag A, Stock Tag C, Stock Tag A] and their theme tag browsing sequence as [Theme Tag B, Theme Tag D].

[0044] Next, the server divides the sequence into subsequences based on a preset threshold. A sequence length threshold (e.g., a maximum length of 8) and a threshold for the time interval between adjacent interactions (e.g., 24 hours) are set. If the browsing sequence length exceeds 8, or the time interval between interactions between two adjacent tags exceeds 24 hours, the sequence is split at the interval. For example, if user A's stock tag browsing sequence is 3 in length (within the threshold), but the time interval between interactions between the second tag (stock tag C) and the third tag (stock tag A) is 25 hours (exceeding the threshold), the server will split it into two subsequences: [stock tag A, stock tag C] and [stock tag A].

[0045] The server then calculates the high-frequency factor for each tag each day. The high-frequency factor is the ratio of the total number of user interactions for a tag on that day to the total number of online users that day. It is used to identify tags with widespread high interaction due to market trends (not genuine user interest). For example, if the total number of online platform users on a given day is 10,000 and the total number of interactions for stock tag E is 5,000, its high-frequency factor is 5,000 / 10,000 = 0.5; if the total number of interactions for stock tag A is 800, its high-frequency factor is 800 / 10,000 = 0.08.

[0046] Remove high-frequency items from the subsequence based on the high-frequency factor. Set a first preset threshold (such as 0.3) and a second preset threshold (such as 0.15): If a tag's high-frequency factor is greater than 0.3 and the difference between it and the high-frequency factor of the adjacent tag in the sequence is greater than 0.15, it is determined to be a high-frequency item and removed. For example, user B's subsequence is [stock tag E, stock tag A, stock tag F], where stock tag E has a high-frequency factor of 0.5 (greater than 0.3) and the difference between it and the high-frequency factor of the adjacent stock tag A of 0.08 is 0.5-0.08 = 0.42 (greater than 0.15). The server removes stock tag E and adjusts the subsequence to [stock tag A, stock tag F].

[0047] The server constructs a transaction dataset based on subsequences after removing high-frequency items and uses an FP-tree to mine frequent itemsets and association rules. Each subsequence is considered a transaction, and all subsequences from all users constitute the transaction dataset. Mining parameters are set: minimum support based on the total number of sequences (for example, if the total number of sequences is 100,000, the minimum support is 50 / 100,000 = 0.0005, meaning that itemsets with ≥50 occurrences are considered frequent itemsets), and minimum confidence is 0.3 (rule reliability ≥ 30%). For example, in the transaction dataset, the subsequence containing [stock label A, stock label C] appears 60 times (support = 60 / 100,000 = 0.0006 > 0.0005), and the subsequence containing [stock label A] appears 200 times. The server mines the frequent itemset {stock label A, stock label C} using the FP-tree and generates the association rule "stock label A → stock label C" with a confidence of 60 / 200 = 0.3 (≥ 0.3), satisfying the rule conditions.

[0048] Finally, the server uses association rules to predict the user's tags of interest. For the target user, the server extracts their most recent interaction tag sequence, matches it to the mined association rules, and recommends the tag that follows the rule. For example, if user A's most recent interaction sequence is [Stock Tag A], and the server matches the association rule "Stock Tag A → Stock Tag C" (confidence level 0.3), it predicts Stock Tag C as the stock tag of interest to user A and adds it to the target tag candidate list.

[0049] Through the above process, the server effectively mines the potential associations between user interest tags and improves the relevance and accuracy of recommendations.

[0050] In an embodiment of the present invention, the time series prediction model is an ARIMA model, which is used to solve the cold start problem of new users. The prediction process of the ARIMA model can be implemented through the following example.

[0051] Based on the historical interaction data, obtain the daily visit sequence of each stock tag or theme tag within a preset time window; Performing a stationary test on the daily visit volume sequence, and if it is not stationary, making the sequence stationary through differential processing to obtain a stationary daily visit volume sequence; Determining the order parameter of the ARIMA model based on the autocorrelation function and partial autocorrelation function of the stabilized daily visit volume series; An ARIMA model is trained based on the determined order parameter and the stabilized daily visit sequence, and the trained ARIMA model is used to predict the popularity of each stock tag or theme tag within a preset time period in the future. The top N items in popularity ranking are taken as popular tags, which are used as the target stock tags or target theme tags for new users.

[0052] In an embodiment of the present invention, for example, in a method for recommending financial articles, a time series prediction model (ARIMA model) is primarily used to solve the cold start problem for new users (when new users have no historical interaction data). Initial recommendations are achieved by predicting popular tags. The server executes the following process: The server first obtains a daily visit count sequence for each tag within a preset time window based on historical interaction data. This historical interaction data includes all past tag interaction records of all platform users (such as clicks or reads of articles containing a certain tag). The server selects a preset time window (such as the past 14 days) and calculates the daily visit count for each stock tag or theme tag (the total number of times all users interacted with the tag that day) on a calendar day to form a daily visit count sequence. For example, for stock tag X, the server counts its daily visits over the past 14 days (from YYYY-MM-DD to YYYY-MM-DD+13 days) and obtains the sequence: [120, 150, 180, 220, 190, 250, 300, 280, 320, 350, 330, 380, 420, 400] (unit: times / day). For theme tag Y, the daily visit sequence is [80, 95, 110, 105, 130, 150, 140, 160, 175, 190, 185, 210, 230, 220].

[0053] Next, the server tests the daily visit count series for stationarity. If it's not stationary, it uses differencing to make the series stationary. Stationarity is a core assumption of the ARIMA model (the series mean and variance remain constant over time). The server uses the ADF (Augmented Dickey-Fuller) test. If the p-value is greater than 0.05, the series is nonstationary and requires differencing (usually first- or second-order differencing). Taking the series of stock label X as an example, the ADF test results in a p-value of 0.32 (greater than 0.05), indicating it's nonstationary. The server performs first-order differencing (number of visits per day minus number of visits per day) to obtain the differenced series: [30, 30, 40, -30, 60, 50, -20, 40, 30, -20, 50, 40, -20]. The differenced series is again tested with an ADF test, and the p-value is 0.02 (less than 0.05), indicating it's stationary. This results in the stationary daily visit count series.

[0054] The server then determines the order parameters of the ARIMA model based on the autocorrelation function (ACF) and partial autocorrelation function (PACF) of the stationary series. The ARIMA model is expressed as ARIMA(p,d,q), where d is the number of differencing (d=1 is set because it is stationary after first-order differencing), p is the autoregressive (AR) order (the lag order at which the PACF is truncated), and q is the moving average (MA) order (the lag order at which the ACF is truncated). The server plots the ACF and PACF of the stationary series: the PACF is significantly truncated at lag 2 (the correlation coefficient falls within the confidence interval after lag 2), so p=2; the ACF is significantly truncated at lag 1, so q=1. Therefore, the ARIMA model parameters for stock label X are determined to be ARIMA(2,1,1).

[0055] The server trains an ARIMA model based on a predetermined order parameter and a stationary sequence to predict tag popularity within a preset timeframe and select the top N tags as popular tags. The server takes the stationary daily visit count sequence (e.g., the first-order difference sequence of stock tag X) as input and trains the model with ARIMA(2,1,1) parameters to minimize the prediction error (e.g., mean squared error (MSE)). After training, the server predicts daily visit count (i.e., popularity) for a preset timeframe (e.g., one day). The same process is repeated for all stock tags and theme tags to obtain their respective predicted popularity values. For example, the server predicts that the popularity of stock tag X will be 430, stock tag Z will be 390, theme tag Y will be 240, and theme tag W will be 210 for the next day. After sorting by descending popularity, the server selects the top N tags (e.g., N=5) as popular tags, ultimately resulting in the target tag list for the new user: [stock tag X, stock tag Z, theme tag Y, theme tag W, stock tag M].

[0056] Through the above process, the server uses the ARIMA model to capture the temporal trend of tag access volume, providing popular tag recommendations for new users without historical data, effectively alleviating the cold start problem.

[0057] In an embodiment of the present invention, the interest decay model dynamically adjusts the weight of the user's historical interaction behavior through the interest retention factor. The processing flow of the interest decay model can be implemented through the following example.

[0058] Divide the historical interaction data into multiple time windows according to dynamic time granularity, where the recent window adopts hourly granularity and the long-term window adopts daily granularity; Extract the stock tags or theme tags set that the user interacted with in each time window, as well as the stock tags or theme tags set that the user interacted with in the current time window; Calculate the interest retention factor for each time window, where the interest retention factor is the ratio of the number of elements in the intersection of the stock label or theme label set of the time window and the stock label or theme label set of the current time window to the total number of elements in the stock label or theme label set of the time window; The weight of the user's historical interaction behaviors within the corresponding time window is weighted based on the interest retention factor, so that the weight of recent interaction behaviors is higher than that of long-term interaction behaviors.

[0059] In an embodiment of the present invention, for example, in a method for recommending financial articles, the interest decay model dynamically adjusts the weight of a user's historical interaction behavior through an interest retention factor to capture changes in user interest over time. The server executes the following process: The server first divides the user's historical interaction data into multiple time windows based on dynamic time granularity. The recent window uses hourly granularity, and the long-term window uses daily granularity. Historical interaction data includes the user's interaction records (such as clicks, readings, and favorites) on financial articles in the past 90 days and the corresponding stock / theme tags. The server uses the current time (such as YYYY-MM-DD 15:00) as the benchmark to trace back. The division rule is: the last 24 hours are divided into 24 recent windows with 1 hour / window (such as (15:00, 14:00], (14:00, 13:00], ..., (16:00, 15:00] the day before); 2 The period from 4 hours to 90 days is divided into 89 long-term windows (e.g., [YYYY-MM-DD 15:00, YYYY-MM-DD-115:00], ..., [YYYY-MM-DD-8815:00, YYYY-MM-DD-8915:00]). For example, user C's historical interaction data covers the past 30 days. The server divides it into 24 hourly near-term windows and 29 day-long far-term windows, for a total of 53 time windows.

[0060] Next, the server extracts the tag set of user interactions in each time window, as well as the tag set in the current time window. The current time window is the most recent hourly window (such as (15:00, 14:00]). The server traverses the user interaction records in the window, extracts all stock tags and subject tags, and forms the current window tag set after deduplication. For the historical time windows (the remaining 52 windows), the same operation is performed: traverse the interaction records in the window, extract tags and deduplication to obtain the tag set of each historical window. For example, user C interacted with articles containing "stock tag A" and "subject tag B" in the current window (14:00-15:00). , the current window label set is {stock label A, theme label B}; in the recent window from 13:00 to 14:00, "stock label A", "stock label C" and "theme label B" were interacted, and the window label set is {stock label A, stock label C, theme label B}; in the forward window 3 days ago (YYYY-MM-DD-3 15:00 to YYYY-MM-DD-2 15:00), "stock label D" and "theme label E" were interacted, and the window label set is {stock label D, theme label E}.

[0061] The server then calculates the interest retention factor for each time window using the formula: "Interest retention factor = number of elements in the intersection of the historical window label set and the current window label set / total number of elements in the historical window label set." The number of elements in the intersection is the number of labels contained in both sets, while the total number is the number of elements in the historical window label set. For example, the intersection of user C's recent window set from 13:00 to 14:00 {stock label A, stock label C, theme label B} and the current window set {stock label A, theme label B} is {stock label A, theme label B}, the number of intersection elements = 2, and the total number of historical windows = 3, so the interest retention factor of this window = 2 / 3≈0.67; the forward window set {stock label D, theme label E} 3 days ago has no intersection with the current window set (number of intersection elements = 0), and the interest retention factor = 0 / 2=0; and the recent window set from 12:00 to 13:00 {stock label A, theme label B} is exactly the same as the current window set, the number of intersection elements = 2, the total number = 2, and the interest retention factor = 1.0.

[0062] Finally, the server weights the user's historical interactions within the corresponding time window based on the interest retention factor, giving more weight to recent interactions than to more distant ones. The server sets a base time-decay weight for each time window (recent windows have higher base weights than more distant ones, e.g., base weight for the last hour = 1.0, for the previous 1-2 hours = 0.9, ..., for the previous day = 0.5, for the previous 2 days = 0.4, ..., for the previous 30 days = 0.1). The server then multiplies this base weight by the interest retention factor to obtain the final interaction weight. For example, user C's window from 13:00 to 14:00 has a base time-decay weight of 0.9 and an interest retention factor of 0.67, resulting in a final weight of 0.9 × 0.67 ≈ 0.60; the window from 12:00 to 13:00 has a base weight of 0.8 and an interest retention factor of 1.0, resulting in a final weight of 0.8 × 1.0 = 0.80; the window from three days ago has a base weight of 0.4 and an interest retention factor of 0, resulting in a final weight of 0.4 × 0 = 0. The contribution value of the user's interactive behavior in each window (such as clicks and favorites) is calculated based on the final weight of the window. For example, the weight of the click on "Stock Label A" in the window of 12:00-13:00 is 0.80, while the click on "Stock Label D" three days ago is no longer included in the current interest calculation because of the weight = 0.

[0063] Through the above process, the server dynamically adjusts the weight of the user's historical interactions, retaining the long-term behavioral impact related to the current interest while highlighting the dominant role of recent behavior, making the interest calculation more in line with the user's real-time preferences.

[0064] In an embodiment of the present invention, the deep learning model is a deep interest network, and the processing flow of the deep interest network can be implemented through the following examples.

[0065] Obtaining user attribute features, behavior type sequence features, and self-selected tag features, wherein the user attribute features are extracted based on the user attribute data, the behavior type sequence features are extracted based on the historical interaction data, and the self-selected tag features are obtained through multi-hot encoding based on the user's self-selected stock list and self-selected theme list; Inputting the user attribute features, behavior type sequence features and self-selected label features into a deep interest network; Dynamically weighting the behavior type sequence features using the attention mechanism through the deep interest network to generate a user dynamic interest feature vector to capture the user's interest changes at different time points; Based on the user's dynamic interest feature vector, the user's interest probability for each stock tag or theme tag is output through the deep interest network, and the top M items in interest probability ranking are taken as recommended tags.

[0066] In an embodiment of the present invention, for example, in a method for recommending financial articles, a deep learning model (deep interest network, DIN) is used to capture dynamic changes in user interests. The server executes the following process: The server first obtains user attribute features, behavior type sequence features, and self-selected label features. User attribute features are extracted based on user attribute data: user attribute data includes institution affiliation, occupation type, trading role, etc., which the server converts into structured features. For example, user D's attribute data is "asset management company," "investment manager," and "buyer." The server uses one-hot encoding to convert the institution into a 100-dimensional vector (corresponding to 100 types of financial institutions, with "asset management company" corresponding to position 1 and all other positions 0), the occupation type into a 50-dimensional vector (with "investment manager" corresponding to position 1), and the trading role into a 2-dimensional vector (with "buyer" corresponding to position 1). The three are concatenated into a user attribute feature vector of length 152.

[0067] The behavior type sequence feature is extracted based on historical interaction data: historical interaction data contains the user's interaction behaviors (reading, clicking, favorites, etc.) and timestamps over the past 30 days. The server sorts them in ascending order by timestamp, extracts the behavior types and corresponding labels, and forms a behavior type sequence. For example, user D's interaction record is: YYYY-MM-DD 09:30 clicks on the article "Stock Label A", 10:15 reads the article "Subject Label B", 14:20 collects the article "Stock Label C", ..., YYYY-MM-DD+29 16:45 reads the article "Stock Label A". The server converts it into a behavior type sequence: [(click, stock label A, 09:30), (read, subject label B, 10:15), (collect, stock label C, 14:20), ..., (read, stock label A, 16:45)], and uses the embedding layer to convert the behavior type (click / read / collect, etc.) and label into a vector of fixed dimension (such as 64 dimensions), which is then spliced into a behavior type sequence feature matrix of length 30×64 (assuming 30 interactions).

[0068] The self-selected label feature is generated through multi-hot encoding based on the user's self-selected stock list and self-selected theme list. The server generates a full label set based on a preset label system (e.g., 10,000 stock labels and 1,000 theme labels), assigning 1 to the corresponding position of the self-selected label and 0 to the rest. For example, if user D's self-selected stock list is {stock label A, stock label C} and their self-selected theme list is {theme label B}, the server generates a multi-hot vector for the stock labels (length 10,000, with 1s in the corresponding positions of stock labels A and C) and a multi-hot vector for the theme labels (length 1,000, with 1s in the corresponding position of theme label B), concatenating them into a self-selected label feature vector of length 11,000.

[0069] Next, the server inputs user attribute features (152 dimensions), behavior type sequence features (30×64 dimensions), and self-selected tag features (11,000 dimensions) into the Deep Interest Network (DIN). The DIN model consists of an embedding layer, an attention layer, and a fully connected layer. The embedding layer first converts high-dimensional sparse features (such as self-selected tag features) into low-dimensional dense vectors, which are then fused with user attribute features and behavior type sequence features.

[0070] A deep interest network employs an attention mechanism to dynamically weight the sequence features of behavior types to generate a user dynamic interest feature vector. The attention mechanism assigns dynamic weights to different behaviors in the sequence based on the relevance of the current recommendation scenario (e.g., the tag to be recommended) to historical behaviors: behaviors with a higher relevance to the current interest receive a higher weight, while behaviors with a lower relevance receive a lower weight. For example, if user D's current recommended tags include "Stock Tag A," the attention mechanism calculates the relevance of each behavior in the sequence to "Stock Tag A." The most recent behavior (YYYY-MM-DD+2916:45) reading "Stock Tag A" has the highest relevance, receiving a weight of 0.8; the behavior of adding "Stock Tag C" (which belongs to the same new energy industry chain as Stock Tag A) to the favorites has the second highest relevance, receiving a weight of 0.3; and the behavior of reading "Theme Tag B" (which is unrelated to Stock Tag A) has the lowest relevance, receiving a weight of 0.1. The server then sums these behavior vectors according to their weights to generate a 64-dimensional user dynamic interest feature vector, which comprehensively reflects the user's current interest in the new energy industry chain.

[0071] Finally, based on the user's dynamic interest feature vector, a deep interest network outputs the user's interest probability for each tag, selecting the top M items as recommended tags. The server concatenates the dynamic interest feature vector with the user's attribute features and the self-selected tag features, inputs the vector into a fully connected layer, and then passes it through a softmax function. The output then shows the user's interest probability for all stock tags and theme tags (summing to 1). For example, in the interest probability distribution for user D, stock tag A is 0.92, stock tag C is 0.85, theme tag B is 0.78, stock tag D is 0.21, and so on. The server sorts the probability distribution in descending order and selects the top M items (e.g., M = 10) as recommended tags, resulting in a list of [stock tag A, stock tag C, theme tag B, ...].

[0072] Through the above process, the deep interest network dynamically captures the changes in users' interests at different time points, making the recommended tags more in line with users' real-time preferences.

[0073] In an embodiment of the present invention, the interactive features are processed using multiple prediction models and the output results of each model are integrated to predict the target stock tags and target subject tags that the user is interested in. This can be implemented through the following examples.

[0074] Determine the weight of each model based on the accuracy of the association rule mining model, time series prediction model, interest decay model, deep learning model, and user-selected label model, where the accuracy is the ratio of the number of labels correctly predicted by the model to the total number of labels predicted by the model; The weights are used to perform weighted fusion sorting on the stock label and subject label prediction results output by each model to obtain the target stock label and target subject label.

[0075] In an embodiment of the present invention, for example, in a financial article recommendation method, the server uses multiple prediction models to collaboratively process interactive features and outputs target labels through weighted fusion. The specific process is as follows: The server first determines the weight of each model based on its historical accuracy. Accuracy is defined as the ratio of the number of labels correctly predicted by the model to the total number of predicted labels. The server then calculates the prediction performance of each model over the past 14 days. For each model, the server records its daily output list of predicted labels (e.g., the top 20 stock labels, the top 20 theme labels), compares it with the labels that users actually interacted with that day (extracted from historical interaction data), and counts the number of correctly predicted labels (the number of predicted labels that appear in the user's actual interaction labels). This is then divided by the total number of predicted labels (e.g., 40 per day) to determine the accuracy for that day. The 14-day average is then used as the final model accuracy. For example, the association rule mining model (FP-Growth) predicted a total of 560 tags (40 tags / day × 14 days) within 14 days, of which 476 tags were actually interacted with by users, with an accuracy of 476 / 560 = 0.85; the time series prediction model (ARIMA) predicted 560 tags, 364 of which were interacted with, with an accuracy of 364 / 560 = 0.65; the interest decay model predicted 560 tags, 336 of which were interacted with, with an accuracy of 336 / 560 = 0.60; the deep learning model (DIN) predicted 560 tags, 420 of which were interacted with, with an accuracy of 420 / 560 = 0.75; the user-selected tag model directly used tags manually added by users. Within 14 days, it predicted 560 tags, 437 of which were interacted with, with an accuracy of 437 / 560 = 0.78. The server normalizes the accuracy of each model (weight = model accuracy / sum of all model accuracies) and calculates the weights as follows: FP-Growth (0.85 / 3.63≈0.234), ARIMA (0.65 / 3.63≈0.179), interest decay model (0.60 / 3.63≈0.165), DIN (0.75 / 3.63≈0.207), and user-selected model (0.78 / 3.63≈0.215).

[0076] The server then uses the weights to perform a weighted fusion sorting of the prediction results output by each model. Each model outputs a predicted list of stock tags and theme tags for the target user. The server assigns each tag an "in-model ranking score" (assigned according to prediction priority, such as Top 1 = 10 points, Top 2 = 9 points, ..., Top 10 = 1 point, and tags not in the top 10 are scored 0 points). This is then multiplied by the model weight to obtain a "weighted score." The weighted scores of the same tag in different models are accumulated to form a "total score." After sorting in descending order by the total score and removing duplicates, the target tag is obtained. Taking the stock tag prediction of user E as an example: The FP-Growth model outputs the top three labels: stock label A (confidence 0.9, ranking score 10), stock label B (confidence 0.8, ranking score 9), and stock label C (confidence 0.7, ranking score 8). The weighted scores are 10×0.234=2.34, 9×0.234=2.11, and 8×0.234=1.87, respectively. The ARIMA model outputs the top 3 popular tags: stock tag D (popularity 0.9, ranking score 10), stock tag E (popularity 0.8, ranking score 9), and stock tag A (popularity 0.7, ranking score 8). The weighted scores are 10×0.179=1.79, 9×0.179=1.61, and 8×0.179=1.43 respectively. The interest decay model outputs the top three labels: stock label B (interest weight 0.8, ranking score 10), stock label C (interest weight 0.7, ranking score 9), and stock label F (interest weight 0.6, ranking score 8). The weighted scores are 10 × 0.165 = 1.65, 9 × 0.165 = 1.49, and 8 × 0.165 = 1.32, respectively. The DIN model outputs the top three labels: stock label A (interest probability 0.92, ranking score 10), stock label C (interest probability 0.85, ranking score 9), and stock label B (interest probability 0.78, ranking score 8). The weighted scores are 10×0.207=2.07, 9×0.207=1.86, and 8×0.207=1.66, respectively. The user-selected model outputs the top three labels: stock label A (self-selected sort 1, sorting score 10), stock label B (self-selected sort 2, sorting score 9), and stock label G (self-selected sort 3, sorting score 8). The weighted scores are 10×0.215=2.15, 9×0.215=1.94, and 8×0.215=1.72, respectively.

[0077] The server adds up the total scores of each tag: stock tag A (2.34 + 1.43 + 2.07 + 2.15 = 7.99), stock tag B (2.11 + 1.65 + 1.66 + 1.94 = 7.36), stock tag C (1.87 + 1.49 + 1.86 = 5.22), stock tag D (1.79), stock tag E (1.61), stock tag F (1.32), and stock tag G (1.72). After sorting in descending order by total score and removing duplicates, the server obtains the target stock tag list for user E: [stock tag A, stock tag B, stock tag C, stock tag G, stock tag D].

[0078] The process for fusing topic tags is the same as for stock tags. The server ultimately outputs the fused target stock and topic tags, providing a basis for subsequent article matching. By weightedly integrating the strengths of each model, recommendation results take into account historical user connections, real-time hot topics, interest decay trends, and explicit preferences, improving the accuracy and comprehensiveness of tag predictions.

[0079] In an embodiment of the present invention, after generating and outputting the user-personalized financial article recommendation list, the embodiment of the present invention further provides the following implementation methods.

[0080] The financial article recommendation list is refreshed at preset time intervals based on the length of the financial article and the user's reading speed.

[0081] In an embodiment of the present invention, after generating and outputting a user-personalized list of financial article recommendations, the server automatically refreshes the list at preset time intervals based on the length of the financial article and the user's reading speed. The specific process is as follows: The server first reads the length attribute (e.g., the "WORD_COUNT" field, in words) of each article in the current recommendation list from the financial article library and adds them up to get the total length of the list. For example, if user Ji's current recommendation list contains five articles with lengths of 1200, 1500, 800, 2000, and 1000 words, the total length = 1200 + 1500 + 800 + 2000 + 1000 = 6500 words.

[0082] Next, the server obtains the user's average reading speed. This speed is based on historical reading behavior data: the server extracts the user's reading history from the past 30 days, calculates the reading time for each article (the time difference between clicking on the article and closing the page, filtering out invalid readings shorter than 30 seconds), and then divides the article length by the reading time to calculate the single reading speed. The 30-day average is taken as the user's current reading speed. For example, if a user has effectively read 20 articles in the past 30 days, with a total reading time of 120 minutes (7200 seconds), and a total word count of 144,000 words, the average reading speed = 144,000 words / 7200 seconds = 20 words / second (or 1200 words / minute).

[0083] The server estimates the total reading time for a list based on its total length and reading speed. The formula is "Total reading time = Total length ÷ Reading speed." The user's total reading time = 6,500 words ÷ 1,200 words / minute, which is 5.42 minutes (approximately 325 seconds).

[0084] The server determines the refresh interval based on a pre-set rule: typically 1.5-2 times the total reading time, ensuring that users have ample time to read the current list before accessing new content. For example, if the user's total reading time is 5.42 minutes, the refresh interval is 2 times the total reading time, resulting in a refresh interval of 5.42 × 2, which is approximately 10.84 minutes. The server rounds this down to 10 minutes.

[0085] Finally, the server starts a scheduled task that automatically triggers a refresh 10 minutes after the current recommendation list is output: the target label prediction process is re-called to obtain the latest target stock label and theme label, the latest published articles are matched from the content library (excluding recommended articles), and a new recommendation list is generated after deduplication in reverse order of publication time. The update is pushed through the client interface so that the article list displayed in the user interface is updated with the latest content in real time.

[0086] Through the above process, the server dynamically adapts to the user's reading rhythm and adds new articles in time when the user is likely to complete reading the current list, thereby improving the timeliness of recommendations and the user's continuous reading experience.

[0087] In order to more clearly describe the solution provided by the embodiment of the present invention, a relatively complete implementation method is provided below. Figure 2 , Figure 2 A schematic diagram of the overall process framework provided by an embodiment of the present invention.

[0088] In an embodiment of the present invention, the content library includes: minutes, research reports, announcements, comments, news information and other types, and each article stores attributes such as article ID, article stock list, article subject list, etc.

[0089] The following is an example of a minutes article: { "UNION_ID":"TRANS00000000016936", "TITLE":"Create2021 AI Developer Conference of Company A - Sub - forum on Smart Life of Little X", "TYPE":"TRANSCRIPT", "SHOW_TITLE":"Create2021 AI Developer Conference of Company A - Sub - forum on Smart Life of Little X", "SEC_JSON":"[{\"code\":\"Company A.US\",\"name\":\"Company A\"}]", "IND_JSON":"[{\"code\":\"HINDUSTRY00000001003\",\"name\":\"Media\"}]", "TPC_JSON":"[{\"code\":\"00000000602\",\"name\":\"Smart Home\"},{\"code\":\"00000000405\",\"name\":\"Smart Speaker\"}]", "CONTENT":"roadshow\ / html\ / 2021\ / 12\ / 29\ / Create2021 AI Developer Conference of Company A - Sub - forum on Smart Life of Little X20211229____In - depth Research on Value Investment Conference Minutes and Research Reports____zsxq.html", "WORD_COUNT":1930, "PUBLISH_DATE":"2021 - 12 - 29 00:00:00", "IS_DELETED":0, "HUPDATETIME":"2023 - 12 - 14 11:26:18" } SEC_JSON is the list of stocks to which this article belongs, IND_JSON is the list of industries, and TPC_JSON is the list of themes.

[0090] The user information table includes attributes such as user ID, the institution where the user is located, user occupation, user role (buy - side, sell - side), the list of self - selected stocks, the list of self - selected themes, etc.

[0091] User example: { "USER_ID":" - 3038469999265093123", "USER_PHONE":"12345678910", "USER_NAME":"Zhang San", "USER_INSTITUTION":"b Fund Institution", "USER_ROLE":"Fund Manager", "USER_TYPE":"Buyer", "FAV_SEC":"[{\"code\":\"Company A.US\",\"name\":\"Company A\"}]","[{\"code\":\"Company A.US\",\"name\":\"Company A\"}]", "FAV_TPC":"[{\"code\":\"00000000602\",\"name\":\"Smart Home\"},{\"code\":\"00000000405\",\"name\":\"Smart Speaker\"}]" } The user behavior table contains: behavior id, behavior user id, article id corresponding to the behavior, category of user behavior EVENT_NAME (including 9 categories: reading, searching, playing, Q&A, downloading, uploading, forwarding, favoriting, following), attributes such as behavior occurrence time DT, etc.

[0092] { "ALPHAPIE_USER_BEHAVIOR": { "HUID":3965912, "DT":"2025-01-09", "DATA_SOURCE":"Web", "EVENT_TIME":"2023-01-09 23:07:41.686", "USER_ID":"-3038469999265093123", "ITEM_ID":"TRANS00000000106223", "EVENT_NAME":"Reading",<00002​​​​​​"ITEM_TOPIC":null, "ITEM_INDUSTRY":"Electrical Equipment", "ITEM_DESCRIPTION":null, "ITEM_SOURCE":"Securities Company A", "USER_STATUS": "Professional Edition" "HCREATETIME":"2024-03-2819:21:15", "HUPDATETIME":"2024-04-2615:39:06" } ]} Stocks include A shares, H shares, and US stocks, totaling 16,611 stocks. The stock code, stock name, and abbreviation of each stock are stored, for example: ('600xxx', 'CSNAME': 'V Company').

[0093] Topics refer to the subject matter and main discussion topics of articles, such as 'multimodality', 'industrial internet', 'blockchain', 'carbon neutrality', 'cathode materials', and 'auto parts and accessories', totaling 799 topics. Articles are tagged with multiple categories based on their content.

[0094] Design, training, and evaluation of each model: Taking stocks as an example, the following describes the specific details of each model. The same applies to the training of theme models.

[0095] FPGrowth (FrequentPatternGrowth): Based on user historical behavior data, potential association rules between stocks are mined, and frequent itemsets of stocks are mined.

[0096] Given a set of database transactions, FPGrowth can build a compact data structure called an FP-tree (Frequent Pattern Tree) by scanning the database twice. Compared to the Apriori algorithm, which typically requires multiple scans of the entire database to find frequent itemsets, FPGrowth avoids generating a large number of candidate itemsets. The FP-tree is an extremely efficient data structure that not only reduces the algorithm's memory requirements but also significantly improves execution speed.

[0097] A transaction database is established based on the user's historical stock browsing sequence to mine association rules.

[0098] Processing of stock series: The length of a stock sequence affects the depth of the FP-tree, and thus the complexity of program execution. Therefore, we use users as units and a four-month period as a regular cycle to obtain all stock sequences browsed by each user within three months. We then divide the user's stock sequence into multiple sequences with a maximum length of 7 and a minimum interval of 3 days between interaction records for adjacent stocks.

[0099] Ensure that two consecutive elements in the sequence are not repeated or identical.

[0100] Calculate the daily "hot search" stock list, remove high-frequency stocks in the user's stock sequence, eliminate high-frequency noise in the sequence, avoid interference from daily high-frequency stocks, and help discover potential stocks.

[0101] High frequency stock calculation: We define a high-frequency factor for stocks and sort stocks from large to small based on this indicator. According to the test results of the actual production environment, if the value is greater than 0.5 and the difference between the factor value of the stock and the stock immediately below it is greater than 0.2, it is considered a high-frequency stock.

[0102] ; Example of a user stock interaction sequence: ['Company B', 'Company C', 'Company D', 'Company E', 'Company F', 'Company G', 'Company H', 'Company I'] ['Company X', 'Company Y', 'Company Z', 'Company DD', 'Company AA'] ['BB Company', 'CC Company', 'BB Company', 'CC Company', 'BB Company'] According to the application scenario of this article, a unified evaluation index is defined to evaluate the performance of the model: ; ; ; Coverage: Ensures that the model can predict stocks for every user as much as possible. Precision: Ensures that the model's accuracy in predicting individual stocks is as high as possible. Among them, coverage and precision are inversely proportional. In order to comprehensively consider the performance of the model in two aspects, a comprehensive score Eval_score is set. Eval_score: It is hoped that Precision is slightly larger than coverage, but in order to ensure that coverage is as high as possible, that is, to predict stocks for every user as much as possible, a second penalty factor is set. According to the above sequence processing method, the total number of stock sequences from January to April is: 50,000, and the average sequence length is: 5.03; The stock sequence is generated into training set, validation set and test set in the ratio of 4:1:1. Screen the optimal parameters of the model, take the performance of the model corresponding to the optimal parameters on the test set as the final performance of the model, and subsequently participate in the weight allocation of the hybrid model.

[0103] antecedent: a set of items representing the hypothesis part of the association rule. consequent: an element representing the conclusion part of the association rule. minSupport: the minimum support for identifying an itemset as a frequent itemset. For example, if an item appears in 3 out of 5 transactions, its support is 3 / 5=0.6. minConfidence: the minimum confidence for generating an association rule. Confidence indicates how often an association rule is proven to be true. For example, if item set X appears 4 times and XY appears 2 times at the same time, X=>Y, then the confidence of the rule is 2 / 4=0.5. This parameter does not affect the mining of frequent itemsets, but specifies the minimum confidence for generating association rules from frequent itemsets. lift: a measure of the predictive power of the antecedent on the consequent, calculated as support(antecedent∪consequent) / (support(antecedent)×support(consequent)). In this experiment, minSupport=10 / total number of stock sequences, confidence=0.25. Please refer to Table 1. In this embodiment of the present invention, the number of association rules is: 2w.

[0104] Table 1 antecedent consequent confidence lift support {'J Company', 'K Company'} L Company 0.2 24.15 0.00013 {'Company F'} Company M 0.2 62.18 0.00016 {'N Company'} Company O 0.2 458.91 0.00016 During prediction, the five most recent stocks in the user's sequence are used as the stock sequence. Antecedent rule matching is performed. If a match is successful, the stock with the highest confidence is selected as the recommended stock based on the consequent stock in the rule. Association rules are trained and updated in a 14-day cycle to recommend stocks to the user over the next 14 days.

[0105] Hot search model based on Arima: ARIMA (Autoregressive Integrated Moving Average) is used for non-stationary time series forecasting. It uses differencing to make the series stationary, eliminating trends and seasonality, and then combines autoregressive (AR) and moving average (MA) modeling. Input: 14-day daily visit counts for each stock (collected from the user behavior table). Example time series: P Company: [2025-01-01: 600 visits, 2025-01-02: 1200 visits, ..., 2025-01-14: 1500 visits]; Stationarity test: ADF test (if the p-value is > 0.05, differencing is required).

[0106] Model Training: Parameter Selection: Use ACF / PACF plots to determine p and q values (e.g., ARIMA(2,1,1)). Training Objective: Predict stock popularity (visits) for the next day. For Company P, 2025-01-15: 1600 visits. Sort the predicted popularity of all stocks on January 15, 2025, in reverse order, and select the top 5 as the popular stocks for that day. For example, ['Company Q', 'Company P', 'Company R', 'Company S', 'Company T'].

[0107] Evaluation method: ; .

[0108] Memory model based on the time series decay of stock interest: Recommend stocks that users have already viewed based on their historical stock interaction sequences. This model introduces an interest retention factor to weight historical user behavior data to reflect the timeliness of user interest. Specifically, the core idea of this method is that user interest in stocks gradually decreases over time. Therefore, newer interactions should have a greater impact on recommendation results than older interactions. Newer stocks have higher interest retention factors, and interest retention factors decay over time.

[0109] Interest retention factor (α): This factor measures how quickly user interest decays over time. This factor determines the degree to which past user behavior influences current interest. Rapidly changing user interests lead to lower interest retention factors, reducing the weight of older data. Experiments have shown that a user's interest decay factor increases with time since the current moment. Specifically, the similarity between currently viewed stocks and stocks viewed yesterday is higher than the similarity between currently viewed stocks and stocks viewed the day before yesterday. In other words, the interest retention factor for the i-th day before the current moment is higher than the interest retention factor for the i+1-th day before the current moment.

[0110] For the current time window cur, the interest retention factor of the user user in the previous i-th time window is calculated as follows: ; The model’s interest retention factor for the i-th time window in the current time window cur, takes all users The mean of .

[0111] Let the current time be 2025-01-15 12:00. Take the stock interaction records of users for the past 7 days to form a stock sequence. Set days and hours as the time units respectively. The time window is: (2025-01-1512:00,2025-01-1511:00],(2025-01-1511:00,2025-01-1510:00],(2025-01-1510:00,2025-01-1412:00],(2025-01-1412:00,2025-01-1312:00],(2025-01-1312:00, 2025-01-12 12:00],(2025-01-12 12:00,2025-01-11 12:00],(2025-01-11 12:00,2025-01-10 12:00],(2025-01-10 12:00,2025-01-09 12:00],(2025-01-09 12:00,2025-01-08 12:00].

[0112] Calculate the interest retention factor of each window, sample the stocks in each window, and use the interest retention factor of each window as the sampling probability to obtain the recommended stocks for each user.

[0113] Evaluation Metrics: ; ; Update the stock recommendation data of the memory model at 1 hour intervals.

[0114] Self-selected stock model: Each user saves a list of stocks of their choice, which are sorted from latest to earliest by stock collection time, and the top 20 are used as the stock list recommended by the model.

[0115] Evaluation Metrics: ; .

[0116] DIN (Deep Interest Network): The DIN model is a deep learning model that uses an attention mechanism to dynamically capture changes in user interests at different points in time and predict the probability of a user's interest in a particular stock. Unlike the aforementioned machine learning models, the DIN model models user interests from a richer perspective.

[0117] Input: User's organization, user position, user category (seller, buyer): whether the user participates in transactions as a seller or buyer. Mean weekly interactions: the average number of times a user interacts with the system each week. Watchlist: a list of stocks that a user has chosen to follow. Watchlist uses multi-hot encoding to obtain the embeddings of all of a user's watchlists, followed by dimensionality reduction using pooling. Sequence of individual stocks associated with articles clicked by the user: the sequence of individual stocks associated with the articles clicked by the user. Sequence of industries associated with articles clicked by the user: the sequence of industries associated with the articles clicked by the user. Sequence of topics associated with articles clicked by the user: the sequence of topics associated with the articles clicked by the user. Pre-trained embeddings combined with dimensionality reduction are used for initialization. Sequence of groups of articles clicked by the user: the sequence of groups of articles clicked by the user, including minutes, comments, and individual stock homepages. Sequence of categories corresponding to user behaviors: the sequence of categories corresponding to user behaviors, including search, reading, and playback. Individual stocks corresponding to recommended articles: the individual stocks corresponding to recommended articles, as well as candidate stocks to be recommended to the user.

[0118] Output: A value between 0 and 1 representing the probability of a user clicking on a particular stock (or recommendation), reflecting the user's interest in the stock. For each user, sort all stocks by their interest scores from highest to lowest, and select the top 50 stocks as the recommended stocks for the user. Evaluation metrics: ; .

[0119] In the present invention, except for the DIN model which differentiates user behavior types (reading, playing, etc.), all other models do not consider the weight of user behavior types.

[0120] Subject label multi-classification model training: Model selection and fine-tuning: Base model: Company A's UTC-roberta-base (fast classification speed, lightweight and suitable for large-scale reasoning); Fine-tuning corpus: 1,200,000+ articles on the financial industry in the past three years: research reports, quarterly reports, quick reviews, in-depth comments, etc.

[0121] Label system: U company v2021, 28 categories at level one, 104 categories at level two, >500 categories at level three.

[0122] This invention uses the UTC-roberta-base as its technical foundation: while maintaining a lightweight parameter scale, the model can achieve an inference throughput of over 100 items per second on a single GPU, making it ideally suited for the company's deployment environment with tens of millions of documents. During the fine-tuning phase, the invention collected approximately 1.2 million publicly available and internally collected financial industry articles from the past three years—including research reports, quarterly reviews, and strategy briefs—and completed label mapping according to the 2021 version of the U Company standard, covering 28 first-level categories, 104 second-level categories, and over 500 third-level categories. To address the significant long-tail effect within financial sub-industries, the invention first conducted a distribution analysis of the training samples and found that less-popular second-level industries such as environmental engineering services and airport shipping had fewer than three articles. To prevent underfitting the model in these categories, the invention used industry names as keywords to search for relevant titles in the internal Elasticsearch cluster, prioritizing the three articles with the highest content relevance as supplementary samples. This expansion reduced the sample variance across industries from 6,700 to 2,400, significantly mitigating the long-tail effect.

[0123] Sample imbalance handling: Statistics show that there are fewer than three articles in unpopular secondary industries such as environmental engineering services and airport shipping. Expansion strategy: Using industry names as keywords, search for matching documents using titles within the company's ES database, and select the three most recent articles based on relevance. Results: The inter-class sample variance decreased from σ² = 6700 to σ² = 2400, significantly mitigating the long tail. Spam filtering rules: 1. Articles with fewer than 500 words: Delete. 2. OCR noise rate exceeding 20%: Delete. 3. Duplicates with existing articles in the database exceeding 95%: Delete. 4. After filtering, 873,000 valid annotated samples remain. Multiple pre-set large models are used to generate labels for each article.

[0124] A simple keyword rule is used as the ninth vote. If five or more votes agree, the tag is added to Ground-Truth.

[0125] This paper designs a voting-based model fusion method to generate groundtruth samples. It also uses a scientific prompt word structure: TASK / ROLE / CONDITION / IO / BACKGROUND, which clearly defines the input and output formats. The following example received a unanimous vote in nine voting nodes: "Photovoltaic Inverters Increase Domestic and International Market Share - In-Depth Report on Leading Inverters" → Level 1: Power Equipment | Level 2: Power Automation Equipment.

[0126] During the online phase, the present invention divides the data into two paths: "historical batches" and "real-time increments." The historical data covers approximately 10 million documents from 2012 to 2022, and the classification is completed solely by the fine-tuned UTC model. It is measured that one hundred tags can be processed in just one second. In the real-time update link, considering that the industry divisions of the latest literature are often accompanied by new concepts and cross-topics, the present invention first uses UTC to give the top five candidate tags for each article, and then calls the DeepSeek model to re-rank these five candidates and output the one with the highest confidence. UTC is responsible for efficient recall, and DeepSeek is used for fine correction. The combination of the two improves the overall F1 score by about 7.8% compared to the UTC single model.

[0127] Based on the results of 50% cross-validation, the macro F1 of the 28 categories in the first-level industry reached 97%, and the macro F1 of the 104 categories in the second-level industry reached 92%. Errors were primarily concentrated in the three adjacent industries of electronics, communications, and power equipment. However, even in these challenging categories, the broad category accuracy remained above 95%.

[0128] Multi-model aggregation sorting strategy: Use precision to measure the performance of each model.

[0129] The number of recommended stocks and update frequency of each model are as follows: FPGrowth: (0,1], updated once on the 1st and 15th of each calendar month; Hot search model: 5, updated once every day at 0:00; Self-selected stock model: (0,20], updated once every day at 0:00; Memory model: (0,10], updated every 2 hours starting from 0:00 every day; DIN: (0,50], updated once every day at 0:00. According to the experimental results > > > > . The stock lists generated by all recommendation models are sorted and deduplicated according to the accuracy of the models. Then, the articles corresponding to these stocks are extracted from the database. If a stock corresponds to multiple articles, these articles are sorted from the latest to the earliest according to the release time, and duplicates are removed to finally generate a list of recommended articles. The processing method for theme recommendation is the same as that for stock recommendation. For the fusion of theme and stock recommendation results, the theme and stock models are fused and sorted according to the accuracy (Precision), and then the articles corresponding to the stocks and themes are extracted from the database and gradually added to the list of recommended articles to ensure that the recommended articles are not repeated. Since the prediction module of stocks and themes is serial with the recommendation of articles, given the recommended list of stocks and themes, the update frequency of articles is once every 10 minutes. According to the strategy of this system, for new users, the popular stock recommendation articles predicted by ARIMA are given priority to avoid the problem of user cold start. Through A / B testing, the system click-through rate increased from 16% to 23%, an increase of 43%.

[0130] Please participate Figure 3 , Figure 3 A financial article recommendation device 110 provided in an embodiment of the present invention includes: Acquisition module 1101 is configured to construct a financial article content library, wherein the financial article content library stores financial articles and corresponding stock tags and subject matter tags, wherein the stock tags and subject matter tags are obtained by classifying the article content to compress the article dimensions into preset attribute dimensions composed of the stock tags and subject matter tags, wherein the total number of the preset attribute dimensions is determined by a tag classification system; obtain user attribute data and historical interaction data between users and financial articles in the financial article content library, wherein the user attribute data includes a user's self-selected stock list and a self-selected subject matter list, and the historical interaction data includes user interaction behaviors with financial articles and corresponding article identifiers and timestamps; and construct interaction features between users and the stock tags and the subject matter tags based on the user attribute data and the historical interaction data, respectively. The interaction features are used to reflect the user's association and interest in each stock tag and subject matter tag. The recommendation module 1102 is used to process the interaction features using multiple prediction models and fuse the output results of each model to predict the target stock tags and target subject tags that the user is interested in. The multiple prediction models include at least an association rule mining model, a time series prediction model, an interest decay model, a deep learning model and a user-selected tag model; based on the predicted target stock tags and target subject tags, corresponding financial articles are matched from the financial article content library, and the matched financial articles are deduplicated in reverse order of publication time to generate and output a user-personalized financial article recommendation list.

[0131] It should be noted that the implementation principles of the aforementioned financial article recommendation device 110 can be referenced from the implementation principles of the aforementioned financial article recommendation method and will not be elaborated upon here. It should be understood that the division of the various modules of the aforementioned device is merely a division of logical functions. In actual implementation, they may be fully or partially integrated into a single physical entity or physically separated. Furthermore, these modules may be implemented entirely as software invoked by a processing element, or entirely as hardware. Alternatively, some modules may be implemented as software invoked by a processing element, while others may be implemented as hardware. For example, the financial article recommendation device 110 may be a separate processing element, or integrated into a chip of the aforementioned device. Furthermore, it may be stored in the form of program code in the memory of the aforementioned device, invoked by a processing element of the aforementioned device to execute the functions of the aforementioned financial article recommendation device 110. The implementation of the other modules is similar. Furthermore, these modules may be fully or partially integrated together, or implemented independently. The processing element described here may be an integrated circuit with signal processing capabilities. During implementation, the steps of the aforementioned method or the aforementioned modules may be performed by hardware integrated logic circuits within a processor element or by software instructions.

[0132] For example, the above modules may be one or more integrated circuits configured to implement the above methods, such as one or more application specific integrated circuits (ASICs), one or more digital signal processors (DSPs), or one or more field programmable gate arrays (FPGAs). For another example, when a module is implemented by scheduling program code on a processing element, the processing element may be a general-purpose processor, such as a central processing unit (CPU) or other processor that can call program code. For another example, these modules may be integrated together and implemented in the form of a system-on-a-chip (SOC).

[0133] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned financial article recommendation device 110. Figure 4 As shown, Figure 4 This is a block diagram of the structure of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a financial article recommendation device 110, a memory 111, a processor 112, and a communication unit 113.

[0134] To enable data transmission or interaction, the memory 111, processor 112, and communication unit 113 are electrically connected to each other, directly or indirectly. For example, these components can be electrically connected via one or more communication buses or signal lines. The financial article recommendation device 110 includes at least one software functional module that can be stored in the form of software or firmware in the memory 111 or embedded in the operating system (OS) of the computer device 100. The processor 112 is used to execute the financial article recommendation device 110 stored in the memory 111, such as the software functional modules and computer programs included in the financial article recommendation device 110.

[0135] An embodiment of the present invention provides a readable storage medium, which includes a computer program. When the computer program is executed, the computer device where the readable storage medium is located is controlled to execute the aforementioned financial article recommendation device 110.

[0136] For illustrative purposes, the foregoing description is made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in light of the above teachings.

Claims

1. A financial article recommendation method, characterized in that: include: Constructing a financial article content library, which stores financial articles and corresponding stock tags and theme tags. The stock tags and theme tags are obtained by classifying the article content to compress the article dimensions into preset attribute dimensions composed of the stock tags and theme tags. The total number of the preset attribute dimensions is determined by the tag classification system. Obtaining user attribute data and historical interaction data between the user and financial articles in the financial article content library, wherein the user attribute data includes a user's self-selected stock list and a self-selected subject list, and the historical interaction data includes the user's interaction behavior with the financial articles and the corresponding article identifiers and timestamps; Based on the user attribute data and the historical interaction data, constructing interaction features between the user and the stock tags and interaction features between the user and the subject tags, respectively, wherein the interaction features are used to reflect the association relationship and interest level of the user in each stock tag and subject tag; Processing the interaction features using multiple prediction models and fusing the outputs of the models to predict target stock tags and target subject tags of interest to the user, wherein the multiple prediction models include at least an association rule mining model, a time series prediction model, an interest decay model, a deep learning model, and a user-selected tag model; Based on the predicted target stock label and target subject label, corresponding financial articles are matched from the financial article content library, and after deduplication of the matched financial articles in reverse order of publication time, a user-personalized financial article recommendation list is generated and output.

2. The method according to claim 1, characterized in that Categorize article content by subject tags, including: Use a pre-trained model that supports zero-shot or small-shot classification as the base model; Collect financial industry article corpus, clean the financial industry article corpus and expand the sample. The sample expansion includes supplementing related articles through keyword search for low-frequency subject tags; Generate labels for training samples using a multi-model voting mechanism that includes the outputs of at least two pre-trained language models and keyword rule matching results; Fine-tune the base model on the cleaned and sample-expanded corpus of financial industry articles to obtain a subject label classification model; The fine-tuned subject label classification model is used to classify financial articles and output candidate labels; the candidate labels are re-ranked using the refined ranking model and the final subject label is output.

3. The method according to claim 1, characterized in that The association rule mining model is the FP-Growth model. The process of mining association rules using the FP-Growth model includes: Taking users as units, based on the historical interaction data, extracting the browsing sequence of stock tags or theme tags of users within a preset time period; Dividing the browsing sequence into multiple subsequences according to a preset sequence length threshold and an adjacent interaction time interval threshold; Calculate the high-frequency factor of each stock tag or theme tag every day, where the high-frequency factor is the ratio of the total number of user interactions for the stock tag or theme tag on that day to the total number of online users on that day; High-frequency items in the multiple subsequences are eliminated based on the high-frequency factor, and the elimination rule is as follows: if the high-frequency factor of a stock label or theme label is greater than a first preset threshold, and the difference between the high-frequency factor of the stock label or theme label and the high-frequency factor of the adjacent stock label or theme label is greater than a second preset threshold, then it is determined to be a high-frequency item and eliminated; A transaction data set is constructed based on the subsequences after removing high-frequency items, and frequent item sets are mined through FP-tree. Association rules are generated based on the frequent item sets. The mining parameters of the association rules include a minimum support value set based on the total number of sequences and a preset minimum confidence value. The association rules are used to predict stock tags or subject tags that the user is interested in.

4. The method according to claim 1, wherein The time series prediction model is an ARIMA model, which is used to solve the cold start problem of new users. The prediction process of the ARIMA model includes: Based on the historical interaction data, obtain the daily visit sequence of each stock tag or theme tag within a preset time window; Performing a stationary test on the daily visit volume sequence, and if it is not stationary, making the sequence stationary through differential processing to obtain a stationary daily visit volume sequence; Determining the order parameter of the ARIMA model based on the autocorrelation function and partial autocorrelation function of the stabilized daily visit volume series; An ARIMA model is trained based on the determined order parameter and the stabilized daily visit sequence, and the trained ARIMA model is used to predict the popularity of each stock tag or theme tag within a preset time period in the future. The top N items in popularity ranking are taken as popular tags, which are used as the target stock tags or target theme tags for new users.

5. The method according to claim 1, wherein The interest decay model dynamically adjusts the weight of the user's historical interaction behavior through the interest retention factor. The processing flow of the interest decay model includes: Divide the historical interaction data into multiple time windows according to dynamic time granularity, where the recent window adopts hourly granularity and the long-term window adopts daily granularity; Extract the stock tags or theme tags set that the user interacted with in each time window, as well as the stock tags or theme tags set that the user interacted with in the current time window; Calculate the interest retention factor for each time window, where the interest retention factor is the ratio of the number of elements in the intersection of the stock label or theme label set of the time window and the stock label or theme label set of the current time window to the total number of elements in the stock label or theme label set of the time window; The weight of the user's historical interaction behaviors within the corresponding time window is weighted based on the interest retention factor, so that the weight of recent interaction behaviors is higher than that of long-term interaction behaviors.

6. The method according to claim 1, characterized in that The deep learning model is a deep interest network, and the processing flow of the deep interest network includes: Obtaining user attribute features, behavior type sequence features, and self-selected tag features, wherein the user attribute features are extracted based on the user attribute data, the behavior type sequence features are extracted based on the historical interaction data, and the self-selected tag features are obtained through multi-hot encoding based on the user's self-selected stock list and self-selected theme list; Inputting the user attribute features, behavior type sequence features and self-selected label features into a deep interest network; Dynamically weighting the behavior type sequence features using the attention mechanism through the deep interest network to generate a user dynamic interest feature vector to capture the user's interest changes at different time points; Based on the user's dynamic interest feature vector, the user's interest probability for each stock tag or theme tag is output through the deep interest network, and the top M items in interest probability ranking are taken as recommended tags.

7. The method according to claim 1, characterized in that The interactive features are processed using multiple prediction models and the output results of each model are integrated to predict the target stock tags and target theme tags that the user is interested in, including: Determine the weight of each model based on the accuracy of the association rule mining model, time series prediction model, interest decay model, deep learning model, and user-selected label model, where the accuracy is the ratio of the number of labels correctly predicted by the model to the total number of labels predicted by the model; The weights are used to perform weighted fusion sorting on the stock label and subject label prediction results output by each model to obtain the target stock label and target subject label.

8. The method according to claim 1, characterized in that After generating and outputting the user-personalized financial article recommendation list, the method further includes: The financial article recommendation list is refreshed at preset time intervals based on the length of the financial article and the user's reading speed.

9. A financial article recommendation device, characterized in that: include: An acquisition module is configured to construct a financial article content library, wherein the financial article content library stores financial articles and corresponding stock tags and subject matter tags, wherein the stock tags and subject matter tags are obtained by classifying the article content to compress the article dimensions into preset attribute dimensions composed of the stock tags and subject matter tags, wherein the total number of the preset attribute dimensions is determined by a tag classification system; obtain user attribute data and historical interaction data between users and financial articles in the financial article content library, wherein the user attribute data includes a user's self-selected stock list and a self-selected subject list, and the historical interaction data includes user interaction behaviors with financial articles and corresponding article identifiers and timestamps; Based on the user attribute data and the historical interaction data, constructing interaction features between the user and the stock tags and interaction features between the user and the subject tags, respectively, wherein the interaction features are used to reflect the association relationship and interest level of the user in each stock tag and subject tag; A recommendation module is used to process the interaction features using multiple prediction models and fuse the output results of each model to predict the target stock tags and target subject tags that the user is interested in. The multiple prediction models include at least an association rule mining model, a time series prediction model, an interest decay model, a deep learning model, and a user-selected tag model; based on the predicted target stock tags and target subject tags, corresponding financial articles are matched from the financial article content library, and the matched financial articles are deduplicated in reverse order of publication time to generate and output a user-personalized financial article recommendation list.

10. A readable storage medium, characterized in that: The readable storage medium includes a computer program, and when the computer program is executed, the computer device where the readable storage medium is located is controlled to execute the method according to any one of claims 1 to 8.

Images