Content recommendation method and content recommendation system

Abstract

The embodiment of the specification provides a content recommendation method and system, wherein the content recommendation method comprises the following steps: obtaining a target topic behavior of a target user; querying a user preference label of the target user from a feature database based on a behavior feature of the target topic behavior, wherein the feature database is constructed based on a behavior feature and a user preference label of a historical topic behavior of a historical user; recalling preference recall content from a content database based on the user preference label of the target user; recalling topic recall content from the content database based on the behavior feature of the target topic behavior; and sorting the preference recall content and the topic recall content to determine recommended content for the target user. The long-term preference interest and the short-term topic intention of the user are considered, so that the recommended content with personalization and timeliness can be output in the case that the long-term preference behavior of the user is sparse, the accuracy and diversity of the recommended result are improved, and the adaptation capability for new users is enhanced.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of content recommendation technology, and in particular to a content recommendation method and a content recommendation system. Background Technology

[0002] With the rapid development of the Internet and content recommendation technology, various platforms need to provide users with personalized content distribution services in order to improve user experience and platform stickiness.

[0003] Currently, content recommendation systems for users typically employ collaborative filtering based on long-term historical user behavior data to retrieve candidate content from a massive content library. However, for new users with sparse behavior or users with unclear long-term preferences, recommendations can only be retrieved based on short-term user behavior data.

[0004] However, for new users with sparse behavior or users with unclear long-term preferences, it is impossible to recall them based on long-term historical behavioral data. Furthermore, the aforementioned techniques for recall based on short-term behavioral data struggle to directly obtain stable interest representations, resulting in insufficient accuracy and diversity in recommendations, leading to lower overall quality and negatively impacting the user's browsing experience. Therefore, there is an urgent need for a content recommendation method that can achieve more accurate and higher-quality content recommendations. Summary of the Invention

[0005] In view of this, embodiments of this specification provide a content recommendation method. One or more embodiments of this specification also relate to a content recommendation system, a computing device, a computer-readable storage medium, and a computer program product, to address the technical deficiencies existing in the prior art.

[0006] According to a first aspect of the embodiments of this specification, a content recommendation method is provided, comprising: Acquire the target user's target topic behavior, wherein the target topic behavior includes at least one of click behavior, browsing behavior and interaction behavior; Based on the behavioral characteristics of the target topic behavior, the user preference tags of the target user are queried from the feature database. The feature database is constructed based on the behavioral characteristics and user preference tags of the historical topic behavior of historical users. Based on the user preference tags of the target users, retrieve preference-retrieved content from the content database; Based on the behavioral characteristics of target topic behaviors, retrieve topic-related content from the content database; The content retrieved based on preferences and topics is sorted to determine the recommended content for the target users.

[0007] According to a second aspect of the embodiments of this specification, a content recommendation system is provided, comprising: The acquisition unit is used to acquire the target user's target topic behavior, wherein the target topic behavior includes at least one of click behavior, browsing behavior and interaction behavior; The query unit is used to query the user preference tags of the target user from the feature database based on the behavioral characteristics of the target topic behavior. The feature database is constructed based on the behavioral characteristics and user preference tags of the historical topic behavior of historical users. The preference recall unit is used to recall preference-recall content from the content database based on the user preference tags of the target user. The topic recall unit is used to recall topic-related content from the content database based on the behavioral characteristics of target topic behavior. The sorting unit is used to sort the preference-recalled content and topic-recalled content to determine the recommended content for the target user.

[0008] According to a third aspect of the embodiments of this specification, a computing device is provided, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer program / instructions, which, when executed by the processor, implement the steps of the recommended method described above.

[0009] According to a fourth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores a computer program / instructions that, when executed by a processor, implement the steps of the method recommended above.

[0010] According to a fifth aspect of the embodiments of this specification, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the steps of the method recommended above.

[0011] One embodiment of this specification implements a content recommendation method, comprising: obtaining target user's target topic behavior, wherein the target topic behavior includes at least one of click behavior, browsing behavior, and interaction behavior; querying the target user's user preference tags from a feature database based on the behavioral characteristics of the target topic behavior, wherein the feature database is constructed based on the behavioral characteristics of historical users' historical topic behavior and user preference tags; recalling preference-recalled content from a content database based on the target user's user preference tags; recalling topic-recalled content from the content database based on the behavioral characteristics of the target topic behavior; and sorting the preference-recalled content and the topic-recalled content to determine recommended content for the target user.

[0012] By acquiring target users' target topic behaviors and querying their preference tags from a pre-built feature database based on historical user data, reliable interest representations can be obtained even if the target users' long-term preference behaviors are sparse. By recalling preference content based on preference tags and topic content based on topic behaviors, a two-pronged recall mechanism is formed, taking into account both users' long-term stable preference interests and short-term dynamic topic intentions. Furthermore, by jointly ranking the content recalled from both sources, recommended content that combines user personalization and content timeliness can be output, effectively improving the accuracy, diversity, and overall quality of the recommendation results, and enhancing the adaptability of the content recommendation method to new users. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating a content recommendation method provided in one embodiment of this specification; Figure 2 This is a flowchart illustrating the processing procedure of a content recommendation method provided in one embodiment of this specification; Figure 3 This is a flowchart of a content recommendation method provided in one embodiment of this specification; Figure 4 This statement is a schematic diagram of the structure of a content recommendation system provided in one embodiment; Figure 5 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0014] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0015] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0016] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0017] Furthermore, it should be noted that the user information (including but not limited to user device information, user registration information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in one or more embodiments of this specification are obtained through open-source datasets or public datasets that comply with their license agreements, or are obtained with full authorization from the relevant parties. Moreover, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0018] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0019] An Application Programming Interface (API) is a set of predefined functions, protocols, and tools that enable communication and interaction between different software components. APIs hide the underlying implementation details, allowing developers to access the functionality or data of external systems by calling the interface, thereby improving development efficiency and facilitating system integration. They are the foundation for modularity and service-oriented architecture in modern software design.

[0020] A Software Development Kit (SDK) is a resource released by a platform, hardware, or service provider, containing API libraries, code examples, documentation, debugging tools, and other resources. With the help of an SDK, developers can efficiently build applications on a specific operating system or framework, lowering the development threshold and ensuring good compatibility with the target platform.

[0021] The Hypertext Transfer Protocol (HTTP) is used to transfer hypertext, images, audio, and video data between web browsers and servers. HTTP employs a client-server model and a request-response mechanism, supports stateless transmission, and is one of the core protocols for data communication on the Internet, laying the technological foundation for the development of the World Wide Web.

[0022] An Internet Protocol address (IP address) is a unique numerical identifier assigned to each device connected to the Internet or a local area network (LAN) for locating and identifying network nodes. IP addresses can be in two formats: IPv4 (32-bit) and IPv6 (128-bit), ensuring that data packets can be correctly routed to their destination devices in complex network topologies.

[0023] Bayesian optimization is a global optimization method based on probabilistic models, primarily used to optimize black-box functions that are costly to evaluate, non-convex, or non-differentiable. It approximates the objective function by constructing surrogate models such as Gaussian processes and dynamically selects the next sampling point using a sampling function (e.g., desired improvement), thereby finding an approximate optimal solution in a relatively small number of iterations. It is widely used in fields such as hyperparameter tuning.

[0024] The Jaccard Similarity Coefficient is a metric used to measure the similarity between two finite sets. It is defined as the ratio of the size of the intersection to the size of the union of the two sets, and its value ranges from 0 to 1. A higher value indicates greater similarity between the sets; a value of 1 indicates that the two sets are completely identical, and a value of 0 indicates that they have no common elements. This metric is commonly used for text deduplication, recommendation systems, and clustering performance evaluation.

[0025] A / B testing is a system tool for online experiments that allows users to be randomly assigned to different versions (e.g., control group A and experimental group B), collect and analyze user behavior data, and assess the statistical significance of strategy changes on key metrics such as click-through rate and conversion rate. This framework helps teams make data-driven product optimization decisions and reduce launch risks.

[0026] This specification provides a content recommendation method, and also relates to a content recommendation system, a computing device, a computer-readable storage medium, and a computer program product, which are described in detail in the following embodiments.

[0027] See Figure 1 , Figure 1 A flowchart of a content recommendation method according to an embodiment of this specification is shown, such as... Figure 1 As shown, the specific steps include:

[0028] Step 102: Obtain the target user's target topic behavior, wherein the target topic behavior includes at least one of click behavior, browsing behavior and interaction behavior.

[0029] The content recommendation method provided in one or more embodiments of this specification can be widely applied to recommendation systems of various content distribution platforms, especially suitable for cold start scenarios where user behavior is sparse or lacks historical preferences. Specifically, it can be applied to mobile applications, web-based content communities, and local lifestyle service products. This content recommendation method can effectively improve the click-through rate, dwell time, and long-term retention performance of target users, optimize their content acquisition experience, and solve problems such as inaccurate interest matching and poor content timeliness caused by long-term data sparsity.

[0030] Target users are specific individuals in a content recommendation system who require personalized content recommendation services. Specifically, target users can include newly registered users, users with short login times, users with sparse behavioral data, and users in the cold start phase. Target users represent the specific objects of the recommendation service, and their behavioral data and preference characteristics serve as important inputs to the content recommendation algorithm. Furthermore, target user identification can be based on various indicators such as user registration time, login frequency, and the amount of historical behavioral data to ensure the targeting of the recommendation strategy.

[0031] Target topic behavior refers to a series of actions performed by a target user during their interaction with a content recommendation system, all related to a specific topic. Specifically, target topic behavior can include various interactive actions such as clicking and browsing of topic content. Target topic behavior reflects a target user's level of interest and attention to specific topic content, providing real-time, dynamic behavioral signals for interest modeling. Furthermore, target topic behavior typically exhibits timeliness, representing a target user's short-term, immediate interest intentions, complementing long-term, stable user preference tags.

[0032] Click behavior refers to the selective triggering action performed by a target user on specific topic content presented by a content recommendation system using a mouse, touchscreen, or other input device. Specifically, click behavior can manifest as a user clicking on a topic title, a topic cover image, or a link related to the topic. Click behavior can characterize a target user's active interest and exploration intent regarding specific topic content. Furthermore, the frequency of click behavior, click sequences, and the types of topics clicked can serve as important behavioral indicators for analyzing user interests and preferences.

[0033] Browsing behavior refers to the visual viewing and information acquisition operations performed by target users on specific topic content presented by a content recommendation system. Specifically, browsing behavior can include users lingering to view topic details pages, scrolling through topic content, or viewing related image and text information. Browsing behavior can characterize the depth of a target user's attention to specific topic content and their willingness to acquire information. Furthermore, the duration of browsing behavior, the browsing path, and whether other operations are accompanied during browsing can serve as important references for measuring the intensity of user interest.

[0034] Interactive behavior refers to proactive actions taken by target users in response to specific topics presented by a content recommendation system, going beyond simple viewing. Specifically, interactive behavior can include actions such as liking, commenting, sharing, saving, or forwarding topic content. Interactive behavior can characterize a target user's level of acceptance, willingness to participate, and intention to spread the information regarding specific topic content. Furthermore, the type and frequency of different interactive behaviors can reflect a user's emotional inclination towards the topic content and the depth of their interaction, providing multi-dimensional behavioral signals for interest modeling.

[0035] In practical applications, obtaining target user behavior related to target topics can be achieved through various methods.

[0036] One alternative approach is to collect target user action logs in real time from a user behavior log system. Specifically, the content recommendation system can deploy specific tracking code on the front-end page to capture various action events performed by the target user on the client interface, and send this event data, along with user identifiers, timestamps, action types, topic identifiers, and other information, to the back-end server for storage and processing. This method enables real-time collection and high-precision recording of behavioral data.

[0037] Another alternative approach is to extract historical behavioral data of the target user from server access logs. Specifically, the content recommendation system's server can record detailed information about all user requests, including request URLs, request parameters, user identifiers, and response statuses. By parsing and processing these log files, the target user's access trajectory and operation sequence for specific topics can be reconstructed. This method is suitable for retrospective analysis and offline modeling of historical behavioral data.

[0038] Another alternative approach is to directly retrieve the target user's target topic behavior features from a pre-built feature database via a feature query interface. Specifically, the content recommendation system can pre-standardize and encapsulate various user behavior data to form a unified feature layer and provide a feature query interface. When it's necessary to retrieve the target user's target topic behavior, the corresponding behavioral feature vector or behavioral statistical indicator can be directly queried through this interface. This method improves the efficiency and consistency of feature acquisition.

[0039] Furthermore, target user behavior related to specific topics can be obtained through a hybrid data source approach. Specifically, real-time event tracking data can be merged with historical log data to ensure both the timeliness and integrity of the behavioral data; alternatively, client-side collected behavioral data can be compared and verified with server-side recorded behavioral data to improve the accuracy and reliability of the behavioral data. This method can be flexibly configured according to actual service needs and system architecture characteristics.

[0040] In this step, by acquiring the target user's target topic behavior, which includes at least one of click behavior, browsing behavior, and interaction behavior, we can comprehensively and multidimensionally capture the target user's real-time operation and interaction intent in the content recommendation system. Click behavior reflects the user's active selection tendency, browsing behavior represents the user's information acquisition depth, and interaction behavior reflects the user's emotional participation. The comprehensive acquisition of these behavioral data provides a data foundation for subsequent user interest modeling and content recommendation. Furthermore, by systematically collecting and analyzing these target topic behaviors, we can effectively overcome the difficulty of interest representation for new users or users with sparse behavior due to the lack of long-term historical data, and provide data conditions for accurate recommendations in cold start scenarios.

[0041] Step 104: Based on the behavioral characteristics of the target topic behavior, query the user preference tags of the target user from the feature database. The feature database is constructed based on the behavioral characteristics and user preference tags of historical users' historical topic behavior.

[0042] Behavioral features of target topic behaviors are numerical representations or feature vectors extracted from the target users' target topic behaviors, used to quantitatively characterize user interaction patterns and interest tendencies. Specifically, behavioral features of target topic behaviors can include click frequency features, browsing duration features, interaction type features, behavior sequence features, topic association features, etc. These behavioral features can transform user actions into machine-processable data, providing standardized feature inputs for subsequent interest modeling and tag queries. Furthermore, behavioral features of target topic behaviors typically undergo feature engineering processing, including normalization, discretization, and encoding, to ensure feature comparability and computational efficiency.

[0043] A feature database is a structured data storage system built upon historical user behavior patterns and user preference tags to store the mapping relationship between user behavior features and interest tags. Specifically, a feature database can take various forms, including relational databases, key-value stores, vector databases, or feature storage platforms. A feature database can centrally manage multi-dimensional user feature data and provide efficient feature retrieval and query services, supporting the real-time invocation needs of online recommendation systems. Furthermore, feature databases typically provide standardized feature query interfaces, enabling downstream recall and ranking modules to access user feature data in a unified manner.

[0044] User preference tags are sets of categorized identifiers or semantic tags used to describe a user's long-term, stable interests. Specifically, user preference tags can include tags based on content topics, such as "food," "travel," and "fashion"; tags based on behavioral patterns, such as "frequent browser" and "deeply engaged user"; or tags based on demographic attributes, such as "young people" and "residents of first-tier cities." User preference tags can be used to abstractly represent a user's interest profile, providing stable and reliable interest-guided signals for content recommendations. Furthermore, user preference tags typically have a long lifespan, reflecting a user's persistent interests across multiple sessions.

[0045] Historical users are a group of users with a long usage history and rich behavioral data accumulated in the content recommendation system. Specifically, historical users include individuals whose registration time exceeds a certain threshold, whose behavioral records reach a certain number, or who have formed stable interest profiles. Historical users can serve as a source of training data, providing sufficient labeled samples for model training and feature calculation. Furthermore, the behavioral data and interest tags of historical users constitute the core data foundation of the feature database, providing referenceable similarity samples for the interest prediction and tag completion of new users.

[0046] Historical topic behavior is a collection of user activity records related to specific topics generated during interactions with the content recommendation system over a past time period. Specifically, historical topic behavior can include user click records, browsing records, and interaction records related to topic content. These records typically contain metadata information such as timestamps, topic identifiers, and behavior types. Historical topic behavior can be used to analyze long-term user behavior patterns, uncover the evolution of user interests, and provide raw behavioral data for building a feature database. Furthermore, after feature extraction and standardization, historical topic behavior can form behavioral features, which, together with user preference tags, constitute a complete record in the feature database.

[0047] Historical user preference tags are a set of tags that characterize the long-term interests of historical users, extracted or inferred from their historical topic behavior data through statistical analysis methods or machine learning models. Specifically, historical user preference tags can be obtained through aggregated analysis of historical user behavior logs, such as counting the frequency of clicks on specific topics or calculating user preference scores for different content categories. These tags can serve as standard labeled data in a feature database, providing a benchmark for similarity matching in target user preference tag queries. Furthermore, historical user preference tags typically have high confidence levels because they are derived from sufficient historical behavioral data and accurately reflect users' true interests.

[0048] In practical applications, querying user preference tags of target users from feature databases based on behavioral characteristics of target topic behaviors can be achieved through various methods.

[0049] One alternative approach is to query user preference tags using feature similarity matching. Specifically, the similarity metric between the behavioral features of the target topic and the behavioral features of historical topic behaviors stored in a feature database can be calculated, such as cosine similarity, Euclidean distance, or Jaccard similarity. One or more historical users whose behavioral features are most similar to the target user are retrieved from the feature database. Then, the user preference tags of these similar historical users are aggregated, for example through voting mechanisms, weighted averaging, or tag fusion algorithms, to generate the target user's user preference tags. This method can fully leverage the similarity relationships of historical users for tag transfer.

[0050] Another alternative approach is to directly predict user preference tags using a pre-trained feature mapping model. Specifically, a regression or classification model can be trained based on historical data from a feature database. This model takes behavioral feature vectors as input and user preference tags or tag probability distributions as output. When querying the user preference tags of a target user, simply input the behavioral features of the target topic behavior into the trained model to obtain the predicted user preference tags. This method enables end-to-end tag querying, reducing intermediate computational steps.

[0051] Another alternative approach is to access the standard services of the feature database through a feature query interface. Specifically, the feature database can provide a unified application programming interface (API) that receives query requests containing behavioral feature vectors of target topic behaviors, internally performs processing logic such as similarity calculation, model inference, or rule matching, and then returns the corresponding user preference tag results. This method encapsulates the query logic within the feature database, providing a simple and easy-to-use service interface, which is beneficial for system decoupling and function reuse.

[0052] Furthermore, user preference tags can be queried through a combination of multi-stage retrieval and fine-grained ranking. Specifically, a candidate set of historical users can be quickly retrieved from massive feature data using approximate nearest neighbor search. Then, more refined feature matching and tag evaluation are performed on the historical users in the candidate set. Finally, the user preference tags of the target user are determined by combining the matching scores from multiple dimensions. This method can improve the query efficiency of large-scale feature databases while ensuring query accuracy.

[0053] In this step, user preference tags of target users are queried from the feature database based on the behavioral characteristics of target topic behavior. The feature database is constructed based on the behavioral characteristics and user preference tags of historical users' historical topic behavior, which can effectively solve the technical problem that it is difficult to accurately represent the interests and preferences of new users or users with sparse behavior due to the lack of long-term historical data. By using the rich behavioral data and labeled tags of historical users to construct the feature database, a transferable knowledge foundation is provided for the interest modeling of target users. By matching or inferring the short-term behavioral characteristics of target users with the historical features in the feature database, the long-term interest tags of target users can be quickly and reliably inferred and predicted. This achieves an effective mapping from short-term dynamic behavior to long-term stable interests, avoiding the noise interference and representation instability caused by simply relying on the sparse behavioral data of target users themselves. It significantly improves the accuracy and robustness of interest modeling in the cold start scenario of new users, and lays a solid interest representation foundation for subsequent accurate content recall.

[0054] Step 106: Retrieve preference-based content from the content database based on the target user's user preference tags.

[0055] Preference-retrieved content is a collection of content retrieved from a content database based on the target user's user preference tags, matching the user's long-term stable interests. Specifically, preference-retrieved content can include various formats such as articles, videos, images, and audio, which are highly correlated with the user's preference tags in terms of topic category, style, or sentiment. Preference-retrieved content can satisfy users' demand for personalized, high-quality content and is provided to target users as an important component of the recommendation results. Furthermore, preference-retrieved content typically has high content quality and user relevance, reflecting the user's persistent interests and preferences, providing target users with a stable and reliable content recommendation experience.

[0056] In practical applications, retrieving content from a content database based on the user preference tags of the target user can be achieved through various methods.

[0057] One alternative approach is to directly retrieve content through tag matching. Specifically, user preference tags can be used as query criteria. The tag field in the content metadata of the content database can be precisely or fuzzily matched to retrieve content items whose tags match or are related to the user's preference tags, forming a preference-recall content set. This method is simple and direct, and can quickly obtain content directly associated with the user's interest tags.

[0058] Another alternative approach is to utilize a pre-trained preference retrieval model for content retrieval. Specifically, training samples can be built based on historical user preference tags and their associated content. This allows for the training of a machine learning model that maps user preference tags to relevant content. When it's necessary to retrieve preference-recalled content for a target user, the target user's preference tags are input into the trained preference retrieval model. The model can then output a corresponding content recommendation list or content relevance score, thereby retrieving preference-recalled content from the content database. This method can capture the complex non-linear relationship between user preferences and content.

[0059] Another option is to combine collaborative filtering algorithms for content retrieval. Specifically, based on the similarity of user preference tags, historical users with similar preference tags to the target user can be found in the feature database. Then, content that these similar historical users have interacted with or preferred is aggregated, forming a candidate pool. Detailed information is then retrieved from the content database to create preference-retrieved content. This method can utilize collective data from a user group to discover content that the user may be interested in but which was not directly matched by tags.

[0060] Furthermore, content retrieval can be achieved through multi-dimensional feature combinations. Specifically, user preference tags can be combined with other user characteristics such as geographic preference tags and demographic attributes to form composite query conditions, enabling multi-condition joint retrieval in the content database. Alternatively, a ranking function for content retrieval can be constructed based on user preference tags, comprehensively considering multiple factors such as the relevance of content to tags, content quality indicators, and content timeliness to rank and filter the search results, obtaining high-quality preference-retrieved content. This method enables more refined content filtering and quality control.

[0061] In this step, by retrieving preference-based content from the content database using the target user's user preference tags, a foundation for content recommendations that highly match the target user's long-term stable interests is provided. By utilizing user preference tags obtained from the feature database as the key basis for content retrieval, consistency between the retrieved content and the user's interests in terms of topic dimension is ensured. This provides a high-quality, highly relevant candidate content pool for subsequent recommendation ranking, effectively improving the overall recommendation system's performance in terms of interest matching accuracy and content quality. As a result, more accurate and higher-quality content recommendations are achieved, enhancing the content recommendation experience for the target user.

[0062] Step 108: Based on the behavioral characteristics of the target topic, retrieve the topic-related content from the content database.

[0063] Topic-based content recall is a collection of content retrieved from a content database based on the behavioral characteristics of target users' behavior on target topics, and is associated with the user's short-term dynamic interests. Specifically, topic-based content recall can include fresh, trending, or popular content similar to topics recently clicked, viewed, or interacted with by the user. This content is highly correlated with the user's short-term behavioral characteristics in terms of topic, style, or timeliness. Topic-based content recall can satisfy users' consumption needs for timely and fresh content, complementing preference-based content recall as an important component of recommendation results. Furthermore, topic-based content recall typically has strong timeliness characteristics, reflecting current hot topics or trends, providing target users with a timely and novel content recommendation experience.

[0064] In practical applications, retrieving topic-related content from a content database based on the behavioral characteristics of the target topic can be achieved through various methods.

[0065] One alternative approach is to expand content retrieval using topic similarity. Specifically, a pre-trained topic similarity model can be used, taking the behavioral features of the target topic as input. The model outputs one or more related topics similar to the input topic. These related topics are then used as query conditions to retrieve content from a content database, forming a topic-retrieved content set. This method can expand the user's interest range based on topic semantic similarity, improving the diversity of content retrieval.

[0066] Another alternative approach is to directly match behavioral features with content features for content recall. Specifically, the behavioral features of the target topic can be transformed into feature vectors, and the topic features of the content in the content database can also be transformed into feature vectors. By calculating the similarity measure between feature vectors, such as cosine similarity or inner product similarity, the content most similar to the user's behavioral features is recalled as the topic recall content. This method achieves direct matching between user behavioral features and content features without the need for an intermediate topic expansion step.

[0067] Another option is to combine regional trending topics for collaborative content retrieval. Specifically, the target user's geographic preference tags can be used to determine their region, and a real-time set of trending topics for that region can be obtained. Then, the intersection of the target user's behavioral characteristics with the regional trending topics set can be used to obtain a subset of topics that aligns with both the user's short-term interests and regional hot topics. Finally, content related to these topics can be retrieved from the content database. This method can balance personalized user interests with regional trends, improving the regional adaptability of content.

[0068] Furthermore, dynamic content retrieval can be achieved through real-time behavioral sequence analysis. Specifically, the behavioral sequence patterns of target users over a recent period can be analyzed to identify the trends in topics the user is currently interested in or the evolution path of their interests. Based on these trends, future topics that the user might be interested in can be predicted, and then content related to the predicted topics can be retrieved from the content database as topic retrieval content. This method can capture the dynamic changes in user interests and achieve more accurate real-time content recommendations.

[0069] In this step, topic-based content is retrieved from the content database based on the behavioral characteristics of the target topic. This allows for real-time content recommendations that highly match the short-term dynamic interests of the target user. By utilizing the real-time behavioral characteristics of the target user as the basis for content retrieval, the consistency between the retrieved content and the user's current interests in terms of timeliness is ensured. This provides the recommendation system with elements of freshness and timeliness, effectively compensating for the potential update lag problem of the retrieved content. As a result, the recommendation results include both stable and reliable high-quality content and real-time, novel, and trending content, significantly improving the overall diversity, freshness, and user satisfaction of the recommendation system. This leads to more accurate and higher-quality content recommendations, enhancing the content recommendation experience for the target user.

[0070] Step 110: Sort the preference-recalled content and topic-recalled content to determine the recommended content for the target users.

[0071] Ranking is a computational process that prioritizes and arranges content retrieved based on preferences and topics. Specifically, ranking can include various methods such as scoring prediction based on machine learning models, reordering adjustments based on service rules, and sequence generation based on diversity control. Ranking can filter out a subset of content from the retrieved content that meets user needs and enhances the user experience, and determine the presentation order of this content. Furthermore, the ranking process can comprehensively consider multiple dimensions of factors, including the relevance of content to user interests, content quality metrics, timeliness characteristics, and diversity balance, and determine the final recommendation sequence through a multi-objective optimization algorithm.

[0072] Recommended content is the final content presentation result received by the target user after preference-recalled content and topic-recalled content have been sorted. Specifically, recommended content can include a list of content, a content stream, or a collection of content cards displayed to the target user in the content recommendation system interface. This content can be arranged in descending order of ranking score. Recommended content can meet the target user's information consumption and interest exploration needs, providing a personalized content browsing experience. Furthermore, recommended content is usually highly user-adaptable, including both high-quality content that matches the user's long-term stable interests and fresh content that is related to the user's short-term dynamic intentions, achieving a balance between interest stability and content timeliness.

[0073] In practical applications, sorting preference-recalled content and topic-recalled content to determine recommended content for target users can be achieved through various methods.

[0074] One alternative approach is to employ a two-stage ranking strategy. Specifically, a coarse-ranking model is first used to initially filter and rank the content retrieved based on preferences and topics using preset coarse-ranking metrics, resulting in coarse-ranked recommended content. These preset coarse-ranking metrics can include at least one of the following: estimated click-through rate, estimated browsing time, and estimated interaction rate. Then, a fine-ranking model is used to refine the ranking of the coarse-ranked recommended content based on preset fine-ranking metrics, yielding the final recommended content tailored to the target user. This method can quickly filter low-quality candidate content in the coarse-ranking stage, reducing the processing scale of the fine-ranking stage, thus improving overall processing efficiency while maintaining ranking quality.

[0075] Another option is to combine regional protection mechanisms with ranking optimization. Specifically, during the ranking process, a safety net strategy can be applied to content retrieved from both preference-based and topic-based sources. For example, based on a preset regional safety net quota, this ensures that a certain number of content from these two sources makes it into the final recommendation list, preventing high-quality content from being prematurely eliminated due to low scores in a single ranking. This method is suitable for new user cold start scenarios, ensuring that recall strategies specifically designed for new users can be effectively implemented and improving the stability of the recommendation system's strategies.

[0076] Another alternative approach is to introduce a multi-source content fusion and ranking mechanism. Specifically, preference-based and topic-based content can be merged with content from other basic recall sources, such as vector recall, like recall, click recall, and real-time recall, to form a unified candidate content pool. This pool is then ranked to determine the final recommended content. This method integrates the advantages of multiple recall strategies, providing a more comprehensive and diverse basis for content selection.

[0077] In addition, a personalized weight adjustment ranking method can be adopted. Specifically, the weight parameters in the ranking model can be dynamically adjusted according to the characteristic attributes of the target users. For new users or users with sparse behavior, the ranking weight of preferred content and topic-recalled content can be appropriately increased; or, differentiated ranking rules can be designed based on factors such as content type, topic popularity, and regional characteristics to achieve more refined content ranking control. This method can enhance the flexibility and scenario adaptability of the ranking strategy.

[0078] In this step, by sorting the content retrieved based on preferences and topics, the recommended content for the target user is determined. This enables effective conversion and optimization from multiple retrieved candidate contents to the final user presentation result. The sorting process comprehensively evaluates and prioritizes the content retrieved based on preferences and topics, ensuring that the recommended results include both high-quality content that highly matches the user's long-term stable interests and fresh content that is closely related to the user's short-term dynamic intentions. This achieves a synergistic balance between long-term preferences and short-term interests, significantly improving the overall performance of the recommendation system. The final content list presented to the target user achieves an optimized balance in multiple dimensions such as relevance, quality, timeliness, and diversity, enhancing the recommendation accuracy and user satisfaction in the cold start scenario for new users.

[0079] In this embodiment, by acquiring the target user's target topic behavior and querying the target user's preference tags from a feature database pre-built based on historical user data, a reliable interest representation can be obtained even if the target user's long-term preference behavior is sparse. By recalling preference content based on preference tags and topic content based on topic behavior, a two-pronged recall mechanism is formed, taking into account both the user's long-term stable preference interests and short-term dynamic topic intentions. Furthermore, by jointly ranking the content recalled from both sources, even when the user's long-term historical behavior is insufficient, it can still output recommended content that combines user personalization and content timeliness, effectively improving the accuracy, diversity, and overall quality of the recommendation results, and enhancing the adaptability of the content recommendation method to new users.

[0080] In an optional embodiment of this specification, before querying the target user's user preference tags from the feature database based on the behavioral characteristics of the target topic behavior, the method further includes: Obtain historical user topic behavior and historical user preference tags, wherein historical topic behavior includes at least one of click behavior, browsing behavior and interaction behavior; The behavioral characteristics and user preference tags of historical topics are standardized, encapsulated to obtain a feature database, and a feature query interface for the feature database is generated. Based on the behavioral characteristics of the target topic, query the user preference tags of the target user from the feature database, including: Based on the behavioral characteristics of the target topic, access the feature query interface to query the user preference tags of the target user from the feature database.

[0081] Feature standardization is a data preprocessing process that standardizes and unifies the behavioral features of historical topics and user preference tags. Specifically, feature standardization can include normalization of numerical features, one-hot encoding of categorical features, word vector embedding of textual features, and fixed-length truncation or padding of sequence features. Feature standardization can transform raw feature data from different sources, with varying dimensions and formats, into standardized feature representations with consistent scale and comparability, providing a unified input format for subsequent feature storage, retrieval, and computation. Furthermore, feature standardization can also include auxiliary operations such as outlier detection and handling, missing value imputation, and feature dimensionality reduction to improve the quality and usability of feature data.

[0082] Encapsulation is a systematic construction process that organizes, integrates, and persistently stores standardized behavioral characteristics and user preference tags of historical topics according to predefined data structures and storage formats. Specifically, encapsulation may include designing the table structure or key-value mapping relationship of the feature database, determining the storage format and indexing method of feature vectors, and dividing the feature data into storage partitions and replica strategies. Encapsulation can integrate discrete feature data items into a logically unified feature database entity, hiding internal storage details and implementation complexity from the outside, and providing a simple and consistent data access abstraction. Furthermore, the encapsulation process is often accompanied by data compression, encryption, or anonymization measures to ensure accurate storage and efficient access to feature data.

[0083] The feature query interface is a programmatic access point provided by the feature database, allowing external systems to query feature data through standardized protocols and formats. Specifically, the feature query interface can be represented as a set of well-defined application programming interfaces (APIs) that specify the parameter format of query requests, data interaction protocols, authentication mechanisms, and encapsulation specifications for returned results. The feature query interface enables downstream recommendation recall modules, ranking modules, or other service systems to retrieve the required user feature data in a unified and efficient manner, without needing to concern themselves with the internal implementation details of the feature database. Furthermore, the feature query interface typically supports multiple query modes such as batch queries, asynchronous queries, and conditional filtering to meet diverse data access needs in different service scenarios.

[0084] In practical applications, obtaining historical user topic behavior and user preference tags can be achieved through various methods.

[0085] One alternative approach is to extract relevant historical user data in batches from the offline data warehouses of user behavior log systems and user profiling systems. Specifically, a scheduled data synchronization task can be used to read historical users' click records, browsing records, and interaction records within a specific time range from the distributed file system storing user behavior logs. Simultaneously, corresponding user preference tags can be queried from the database of the user profiling system. These data are then merged to form a complete behavior-tag mapping dataset for historical users. This method leverages existing data infrastructure to acquire large-scale, high-quality historical user annotation data in a batch processing manner.

[0086] Another alternative approach is to continuously collect historical user behavior data and tag updates through a real-time data stream processing platform. Specifically, a real-time data acquisition pipeline can be deployed to monitor user behavior event streams and profile update event streams. When a historical user exhibits new behavior or their tag changes, the relevant events can be immediately captured and written to a temporary buffer, which is then aggregated and persistently stored by downstream processing programs. This method enables near real-time collection and updating of historical user data, ensuring the timeliness of data in the feature database.

[0087] Another alternative approach is to integrate historical user data from different service systems through multi-source data fusion. Specifically, metadata interaction information about user topics can be obtained from content management systems, user attention and interaction network data can be obtained from social relationship systems, and user regional activity trajectories can be obtained from geolocation services. This multi-dimensional data can then be correlated and fused with core behavioral logs and tag data to form a richer and more comprehensive historical user feature dataset. This method can expand the dimensions and information content of historical user data, providing a richer modeling foundation for the feature database.

[0088] Furthermore, historical user data can be optimized through data sampling and augmentation techniques. Specifically, oversampling techniques can be used to increase the representativeness of historical users with sparse behavior, reweighting techniques can be used to adjust the sample weights of historical users with unbalanced label distributions, and data synthesis techniques can be used to generate simulated historical user data that conforms to the real distribution to supplement insufficient data volume. This approach can improve the quality and balance of historical user datasets at the data level, thereby enhancing the effectiveness of subsequent feature modeling.

[0089] Once the historical topic behaviors and user preference tags of historical users are obtained, the behavioral characteristics of historical topic behaviors and user preference tags can be standardized, encapsulated to obtain a feature database, and a feature query interface for the feature database can be generated.

[0090] One alternative approach is to employ a phased pipelined processing architecture to achieve feature standardization and encapsulation. Specifically, a feature preprocessing pipeline can be designed, sequentially executing standardized operations such as data cleaning, feature extraction, normalization, and encoding conversion to transform raw historical data into standard feature vectors. Then, the processed feature vectors can be organized according to user identifiers and imported in batches into selected database storage engines, such as relational databases, key-value stores, or dedicated feature storage platforms. Furthermore, a feature query interface service can be developed and deployed on top of the database. This service can receive query requests including user identifiers and feature types, retrieve data from the database, and return the corresponding feature data. This approach achieves modularity and scalability in the processing flow.

[0091] Another alternative approach is to leverage the various automated tools provided by the feature computation platform to complete standardized processing and encapsulation. Specifically, the feature definition language provided by the platform can be used to declare the behavioral features of historical topic behaviors and the calculation logic for user preference tags. The platform then automatically generates and schedules feature computation tasks based on these declarations, storing the results in its built-in feature storage system. This system typically provides standardized feature query interfaces, supporting feature retrieval via a unified API or SDK. This approach reduces the technical complexity of feature engineering and improves development efficiency.

[0092] Furthermore, an independent feature service system can be designed based on a microservice architecture. Specifically, a feature standardization processing microservice can be developed, which can receive raw historical data input and output standardized features accordingly. Simultaneously, a feature storage management microservice can be used to persistently store standardized features and provide basic CRUD (Create, Read, Update, Delete) capabilities. Further, a feature query interface microservice can provide a unified feature query API externally, while internally calling the feature storage management microservice to retrieve data. This approach achieves separation of concerns through service-oriented decomposition, enhancing the system's maintainability and scalability.

[0093] Once a feature database has been pre-built and a feature query interface for the feature database has been generated, users can access the feature query interface based on the behavioral characteristics of the target topic behavior and query the user preference tags of the target user from the feature database.

[0094] An alternative approach is to invoke the feature query interface via a direct HTTP request. Specifically, the content recommendation system can construct a query request that includes behavioral feature vectors of the target topic's behavior, encapsulate the request according to the protocol format defined by the feature query interface, and send it to the feature query interface server via HTTP. After receiving the request, the server can parse the feature vectors, perform similarity matching or model inference calculations in the feature database, obtain the target user's user preference tags, and encapsulate them into a response message to return to the recommendation system.

[0095] Another alternative approach is to access the feature query interface through a client-side software development kit (SDK). Specifically, the feature query interface provides SSDs for different programming languages. After integrating this SSD, the recommendation system can directly call the query methods encapsulated within the SSD, passing in the behavioral feature vector of the target topic. The SSD internally handles details such as request serialization, network communication, and response deserialization, ultimately returning the user preference tag results. This method encapsulates network communication and protocol details, providing a more convenient programming interface.

[0096] Another alternative approach is to perform feature queries using asynchronous message queues. Specifically, the content recommendation system can publish a query message containing behavioral feature vectors of the target topic's behavior to a message queue. The feature query interface service can then retrieve the corresponding message from this queue for processing, writing the retrieved user preference tags to another result queue or directly calling back the callback interface provided by the recommendation system. This method is suitable for scenarios with relatively relaxed real-time requirements or extremely high query throughput, enabling decoupling and load balancing between systems.

[0097] Furthermore, feature query efficiency can be improved through caching optimization mechanisms. For example, recommendation systems can cache recently queried feature-tag mappings in a local or distributed cache. When querying a target user's preference tags, the system can first check the cache for existing query results with the same or similar behavioral feature vectors. If a match is found, the cached result is returned directly; otherwise, the feature query interface is accessed. This approach reduces the pressure of direct access to the feature database and lowers query latency.

[0098] In this embodiment, before querying the target user's user preference tags from the feature database based on the behavioral characteristics of the target topic behavior, historical user topic behaviors and user preference tags are obtained. These historical user topic behavior behavioral characteristics and user preference tags are then standardized and encapsulated to obtain a feature database. A feature query interface for the feature database is generated, and the user preference tags of the target user are then queried based on the behavioral characteristics of the target topic behavior. This achieves the systematic construction and standardized service of feature data. Through the collection and standardization of historical user data, a high-quality feature database is constructed, providing a rich and reliable reference benchmark for querying the target user's interest tags. Through feature encapsulation and query interface generation, the management and access of feature data are serviced, providing unified and efficient feature retrieval capabilities. Through standardized interface access methods, the integration between the recommendation system and the feature database is achieved, improving the modularity and maintainability of the entire recommendation architecture and providing a solid data foundation and infrastructure support for the core function of querying user preference tags based on behavioral characteristics.

[0099] In an optional embodiment of this specification, before performing feature standardization processing on the behavioral characteristics and user preference tags of historical topic behaviors and encapsulating them to obtain a feature database, the following steps are also included: Obtain historical users' geographic preference tags; The behavioral characteristics and user preference tags of historical topics are standardized, encapsulated to obtain a feature database, and a feature query interface for the feature database is generated, including: The behavioral characteristics, geographic preference tags, and user preference tags of historical topics are standardized, encapsulated to obtain a feature database, and a feature query interface for the feature database is generated. Before accessing the feature query interface to retrieve the target user's user preference tags from the feature database based on the behavioral characteristics of the target topic, the process also includes: Obtain the geographic preference tags of the target users; Based on the behavioral characteristics of the target topic, access the feature query interface to query the target user's user preference tags from the feature database, including: Based on the behavioral characteristics of the target topic and geographical preference tags, access the feature query interface to query the user preference tags of the target user from the feature database.

[0100] Geographic preference tags are a set of categorized identifiers or semantic labels used to describe users' interests and behavioral preferences across geographical locations. Specifically, geographic preference tags can include tags based on administrative divisions, such as "users in North China" or "users in the Yangtze River Delta region"; tags based on city groups, such as "users in first-tier cities" or "users in new first-tier cities"; tags based on cities of residence, such as "users residing in Beijing" or "users residing in Shanghai"; and tags based on geographical location characteristics, such as "users in coastal areas" or "users in inland areas". Geographic preference tags can reflect users' cultural identification with, interest in, and consumption tendencies regarding content from different regions, providing a basis for making decisions on the regional suitability of content recommendations. Furthermore, geographic preference tags are usually associated with multi-dimensional information such as users' actual geographical location, historical activity areas, and content interaction preferences, and can characterize users' stable or dynamic interest characteristics across geographical dimensions.

[0101] In practical applications, obtaining historical users' geographic preference tags can be achieved through various methods.

[0102] One alternative approach is to directly extract geographic preference tags from user registration information and user profiles. Specifically, the content recommendation system can parse the information about the user's permanent residence city, geographical location, or manually set geographic preference tags during registration, and convert this explicit geographic information into standardized geographic preference tags. For example, the user's "Beijing" can be mapped to the tag "Beijing resident user," and "Shanghai" can be mapped to the tag "Shanghai resident user."

[0103] Another alternative approach is to infer geographic preference tags by analyzing users' historical behavioral data. Specifically, this involves statistically analyzing the geographic distribution of content accessed by historical users within specific time periods. For example, calculating the proportion of content from different regions within user interactions, identifying the most frequently viewed regional topics, or analyzing high-frequency geographic locations in user activity patterns. Based on these behavioral patterns, geographic preference tags can be derived. For instance, if a user frequently browses content related to "Guangdong cuisine," a "South China preference" tag could be added. This approach allows for the uncovering of implicit geographic preferences from users' actual behavior.

[0104] Another alternative approach is to generate geographic preference tags using geolocation service data. Specifically, real-time geographic location information of historical users can be obtained through mobile device location services, or network access locations can be obtained through IP address resolution. This location data can then be used for cluster analysis, trajectory mining, and semantic annotation to generate geographic preference tags describing users' geographic activity patterns. For example, frequently visited geographic locations can be clustered into "work areas," "residential areas," and "leisure areas," and corresponding geographic preference tags can be generated based on these regional characteristics. This method provides dynamic, real-time geographic preference information.

[0105] Furthermore, geographic preference labels can be predicted using regional preference models. For example, a regional preference prediction model can be trained based on other characteristics of historical users, such as interest tags, behavioral patterns, and demographic attributes. This model can predict a user's possible geographic preference tendencies based on non-geographical characteristics, generating supplementary geographic preference labels for historical users. This method is particularly suitable for users with sparse or inaccurate geographic information, as it can complete their geographic dimension features.

[0106] Once the geographic preference tags of historical users are obtained, the behavioral characteristics, geographic preference tags, and user preference tags of historical topic behaviors can be standardized, encapsulated to obtain a feature database, and a feature query interface for the feature database can be generated.

[0107] One alternative approach is to employ a standardized processing flow for multimodal feature fusion. Specifically, standardized processing strategies can be designed for three different types of data: behavioral features of historical topic behaviors, geographic preference tags, and user preference tags. For numerical behavioral features, normalization and discretization are performed; for categorical geographic preference tags and user preference tags, one-hot encoding or embedding representation is used. Then, the processed feature vectors are concatenated, weighted, or fused using attention mechanisms to form a unified user feature representation. Subsequently, these fused feature vectors are organized according to user identifiers and stored in a feature database. Finally, a feature query interface is developed based on this feature database. This interface supports querying complete user feature vectors by user identifier or performing conditional filtering queries by feature dimension. This achieves unified processing and integrated access to multi-source heterogeneous features.

[0108] Another alternative approach is to construct a hierarchical feature standardization and encapsulation architecture. Specifically, this can be designed with three layers: a basic feature layer, a derived feature layer, and an application feature layer. In the basic feature layer, basic cleaning and standardization can be performed on the original historical topic behavior features, geographical preference tags, and user preference tags. In the derived feature layer, cross-features, statistical features, or model-generated features can be calculated based on the basic features. In the application feature layer, features can be combined into specific feature sets according to downstream service requirements. Features at each layer can be independently encapsulated and stored, and the feature query interface can select to access feature data from different layers based on the application scenario. This hierarchical design enables flexible combination and on-demand access to features.

[0109] Another alternative approach is to automate the processing and encapsulation using feature pipeline orchestration tools. Specifically, the visual orchestration tools provided by feature engineering platforms can be used to orchestrate data processing tasks such as extracting behavioral features from historical topics, generating geographic preference tags, and acquiring user preference tags, along with subsequent standardization, feature fusion, storage, and encapsulation operations, into an executable workflow. This workflow can be scheduled and run on a timer, automatically completing the entire process from raw data to the feature database. A feature query interface can be automatically generated as an output component of this workflow, providing standardized feature retrieval services. This method reduces the development and maintenance costs of feature processing.

[0110] Furthermore, a versioned feature management mechanism can be employed to support feature iteration. For example, an independent feature version can be created for each execution of the feature standardization process, with feature data from different versions coexisting in the feature database and bearing version identifiers. The feature query interface supports querying feature data by specifying a version number or automatically selecting the latest version of feature data. Simultaneously, a feature metadata management system is established to record information such as the processing logic, data source, and quality metrics for each feature version. This approach supports feature rollback, comparison, and iterative optimization, enhancing the maintainability of the feature system.

[0111] In practical applications, for details on obtaining the geographic preference tags of target users, please refer to the specific methods for obtaining the geographic preference tags of historical target users.

[0112] Once the geographic preference tags of the target user are obtained, the feature query interface can be accessed based on the behavioral characteristics of the target topic behavior and the geographic preference tags to query the user preference tags of the target user from the feature database.

[0113] One alternative approach is to concatenate the behavioral features of the target topic behavior with geographical preference tags to obtain a composite feature vector for querying. Specifically, the behavioral feature vector of the target topic behavior can be concatenated with the encoded vector of the geographical preference tags to form an extended feature representation. This composite feature vector is then used as query input and submitted to the feature database through a feature query interface. The feature database can then perform similarity matching between this composite feature vector and the composite feature vectors of historical users to find the most similar historical users and return their user preference tags. This method incorporates geographical preference information as part of the features in similarity calculation, achieving joint matching of multi-dimensional features.

[0114] Another alternative approach is to design a feature query interface that supports multi-condition queries. Specifically, the feature query interface can accept behavioral features of the target topic and geographic preference tags as independent query conditions. Within the feature database, initial filtering is performed based on geographic preference tags to narrow down the range of candidate historical users. Then, fine-tuning is performed on the behavioral features within the filtered subset, ultimately returning the matched user preference tags. This method improves query efficiency through two-step filtering and is particularly suitable for scenarios with large feature databases.

[0115] Another alternative approach is to reorder or weightedly fuse feature query results based on geographic preference tags. Specifically, one can first obtain a preliminary set of candidate user preference tags based on the behavioral characteristics of the target topic's behavior through a feature query interface. Then, this candidate set can be reordered based on geographic preference tags, increasing the priority of user preference tags that match geographic preferences. Alternatively, one can perform independent queries based on behavioral characteristics and geographic preference tags respectively, obtaining two sets of user preference tag results, and then weighting and combining them according to a preset fusion strategy. This method achieves synergy between behavioral characteristics and geographic preference tags at the result level.

[0116] Furthermore, more refined tag queries can be achieved through customized query strategies. For example, the weight of geographic preference tags in the query can be dynamically adjusted based on the strength of those tags. For users with strong geographic preferences, the weight of the geographic dimension in similarity calculation can be increased. Alternatively, feature query optimization based on geographic partitions can be designed, storing the feature database in partitions according to geographic preference tags. During a query, the partition corresponding to the target user's geographic preference tag is located first, and then behavioral feature matching is performed. This approach allows for query optimization based on specific service needs and data characteristics.

[0117] In this embodiment, before standardizing the behavioral characteristics and user preference tags of historical topic behaviors, the geographic preference tags of historical users are obtained. These geographic preference tags are then standardized and encapsulated together with the behavioral characteristics and user preference tags of historical topic behaviors. Simultaneously, before accessing the feature query interface to query the user preference tags of the target user based on the behavioral characteristics of the target topic behavior, the geographic preference tags of the target user are obtained. Finally, the feature query interface is accessed together with the behavioral characteristics and geographic preference tags of the target topic behavior to query the user preference tags. This achieves deep integration and effective utilization of geographic dimension features in user interest modeling, enriching the dimensions of user feature representation. It enables the feature database to simultaneously characterize user feature patterns at the behavioral, interest, and geographic levels, enhancing the accuracy and relevance of tag queries. The query results not only reflect user behavioral similarities but also demonstrate the consistency of geographic preferences. This allows for more accurate and relevant interest tags for new users or users with sparse behavior, better matching their actual regional background and cultural context. This significantly improves the accuracy, personalization, and overall quality of the content, enhancing the user's content acquisition experience.

[0118] In an optional embodiment of this specification, before retrieving content from the content database based on the target user's user preference tags, the method further includes: Obtain user preference tags and historical preference content from historical users; Train a preference recall model based on historical user preference tags and historical preference content; Based on the target user's user preference tags, retrieve preference-retrieved content from the content database, including: By using a preference recall model, preference-based content is retrieved from a content database based on the user preference tags of the target user.

[0119] Historical preference content refers to the collection of content in a content recommendation system in which users have historically engaged positively, expressed clear interests, or consumed deeply. Specifically, historical preference content can include various forms of content such as articles clicked, videos viewed, images liked, audio comments made, or posts shared by historical users. Historical preference content reflects a user's actual interest and consumption preferences regarding specific topics, styles, or types of content, providing high-quality labeled sample data for training preference recall models. Furthermore, historical preference content typically undergoes certain quality screening and interest intensity filtering; for example, it may only include content with interaction duration exceeding a threshold, interaction frequency meeting a standard, or interaction type indicating strong interest, to ensure the accuracy and representativeness of the training data.

[0120] Preference recall models are machine learning models trained on historical user preference tags and content preferences, capable of predicting content a user might be interested in based on those tags. Specifically, preference recall models can take various forms, including matrix factorization-based collaborative filtering models, deep learning-based neural recommendation models, tree-based ranking models, and multi-task learning-based recall models. These models learn the complex mapping between user preference tags and content features, outputting a corresponding content recommendation list or content relevance score when given user preference tags, thus achieving efficient matching from user interest tags to personalized content. Furthermore, preference recall models are typically deployed using offline training and online services, continuously learning historical user behavior patterns and optimizing recall performance.

[0121] In practical applications, training a preference recall model based on historical user preference tags and historical preference content can be achieved through various methods.

[0122] One alternative approach is to train a preference recall model using a supervised learning framework. Specifically, historical user preference tags can be used as input features, and historical preference content can be used as the model's prediction target to construct a supervised training sample set. Then, an appropriate machine learning algorithm, such as logistic regression, gradient boosting decision trees, or deep neural networks, is selected to optimize the model parameters on this training sample set, enabling the model to accurately predict the content preference probability under a given user preference tag. During training, negative sampling techniques can be introduced to handle massive amounts of candidate content, cross-entropy loss functions can be used to optimize the model's prediction accuracy, and early stopping on the validation set can prevent overfitting. This method can directly learn the association between user tags and content.

[0123] Another alternative approach is to build a preference recall model based on collaborative filtering. Specifically, a user-content interaction matrix can be constructed from historical user preference tags and content preferences, where matrix elements represent the intensity of a user's interest in the content. Then, matrix factorization is used to decompose this interaction matrix into user latent vector matrices and content latent vector matrices, learning the representations of users and content in the latent space. The trained model can then calculate an interest score based on the inner product of the latent vectors corresponding to user preference tags and content latent vectors, thus achieving content recall. This method can uncover potential association patterns between users and content.

[0124] Another alternative approach is to train a preference recall model according to a pre-defined training strategy. This strategy can include optimizing the model parameters based on the correlation between historical user preference tags and historical preference content, and adjusting the strategy parameters through offline evaluation. Specifically, a multi-stage training process can be designed. First, an initial model is built based on the co-occurrence relationship between user preference tags and historical preference content. Then, the model's performance is evaluated using offline evaluation metrics such as precision, recall, or normalized cumulative gain. Based on the evaluation results, model parameters, feature engineering strategies, or sampling strategies are adjusted. The training and evaluation process can be iteratively executed until the model performance converges. This method achieves continuous optimization of model performance through a systematic training and evaluation cycle.

[0125] Furthermore, transfer learning or domain adaptation techniques can be used to train preference recall models. Specifically, a basic recall model can be pre-trained based on large-scale general recommendation data, and then fine-tuned using user preference tags and content from historical users within the domain to adapt it to specific service scenarios. Alternatively, a multi-task learning architecture can be designed to simultaneously learn the mapping from user preference tags to content and the mapping from user behavior sequences to content, improving the model's generalization ability through knowledge sharing between tasks. This approach can leverage external knowledge or relevant task information to enhance model performance.

[0126] Once a preference recall model is trained, it can be used to recall preference-recall content from the content database based on the user preference tags of the target user.

[0127] One alternative approach is to directly generate a content recommendation list through preference recall model inference. Specifically, the target user's preference tags can be input into a trained preference recall model, which outputs interest scores or ranking scores for all candidate content in the content database. Then, the content is sorted from highest to lowest score, and the top-ranked content is selected as the preference recall content. To improve recall efficiency, approximate nearest neighbor search technology can be used to accelerate the retrieval process of a large-scale content database. This method achieves end-to-end content recall and simplifies the recall process.

[0128] Another alternative approach is to combine preference recall models with inverted index retrieval for content retrieval. Specifically, based on the preference recall model, a relevance score between each piece of content in the content database and different user preference tags can be calculated, and a tag-content inverted index can be built. When it is necessary to recall preference-recalled content for a target user, the corresponding inverted index list can be queried based on the target user's user preference tags to obtain an initial set of relevant content. Then, the preference recall model is used to refine or reorder this initial set to obtain the final preference-recalled content. This method, through index retrieval, can narrow down the candidate range and improve the retrieval efficiency of large-scale content databases.

[0129] Furthermore, the recall strategy of the preference recall model can be dynamically adjusted based on user characteristics. Specifically, for new users or users with sparse behavior, the recall weight of the preference recall model can be appropriately increased or the recall threshold can be relaxed to ensure a sufficient number of recalls; for active users or users with clear interests, the number of recalls can be reduced but the accuracy requirement can be increased; alternatively, the output results of the preference recall model can be post-filtered or re-ranked by combining information such as users' geographical preference tags and time context. This method achieves personalized adaptation of the recall strategy.

[0130] In this embodiment, before recalling preference-based content from the content database based on the target user's user preference tags, historical user preference tags and historical preference content are obtained. A preference recall model is trained based on this data, and then the model is used to recall preference-based content based on the target user's user preference tags. This achieves intelligent content recall based on a machine learning model, deeply exploring the complex non-linear relationship between user preference tags and content, enabling more accurate and personalized content matching. By using the preference recall model, the retrieval problem of large-scale content databases can be effectively handled, improving recall efficiency while ensuring recall quality. This enhances the intelligence level and stability of the preference recall process, making the preference-based content recalled for the target user more closely match their true interests and preferences. This provides a higher quality and more relevant candidate content foundation for subsequent ranking and recommendation, thereby improving the accuracy and user satisfaction of the entire content recommendation system.

[0131] In one optional embodiment of this specification, a preference recall model is trained based on historical user preference tags and historical preference content, including: Based on historical users' user preference tags and historical preference content, a preference recall model is trained according to a preset training strategy. The preset training strategy includes optimizing the model parameters of the preference recall model based on the correlation between historical users' user preference tags and historical preference content, and adjusting the strategy parameters of the preference recall model through offline evaluation.

[0132] A pre-defined training strategy is a set of rules, procedures, methods, and configurations followed when training a preference-based recall model. Specifically, a pre-defined training strategy may include training data partitioning strategies (such as the ratio of training, validation, and test sets), model optimization strategies (such as the selection of learning rate scheduling algorithms and gradient descent optimizers), regularization strategies (such as weight decay and dropout rate settings), and training termination strategies (such as early stopping conditions and the setting of the maximum number of iterations). A pre-defined training strategy guides the systematic execution of the training process, ensuring the repeatability, controllability, and stability of model training results. Furthermore, pre-defined training strategies are typically determined based on domain experience, experimental results, or automatic hyperparameter optimization techniques and can be customized according to specific service scenarios and data characteristics.

[0133] The association between historical user preference tags and historical preferred content encompasses various forms of relationships, including statistical correlation, semantic similarity, and causal connections. Specifically, the association can manifest as a high co-occurrence frequency between specific user preference tags and certain content categories (e.g., the "food" tag and "restaurant recommendations" content frequently appearing together); a close proximity in the latent space between user preference tags and content feature vectors, indicating semantic similarity; or the ability of the occurrence of user preference tags to predict the probability of user interaction with specific content. This association reflects the concrete manifestation of user interest tags in actual content consumption, providing crucial learning signals for training preference recall models. Furthermore, the strength and type of this association can be quantitatively evaluated using various metrics such as correlation coefficients, mutual information, and attention weights.

[0134] Model parameters are numerical variables such as learnable weights and biases within the preference recall model. These variables are continuously adjusted during training using optimization algorithms to minimize the loss function. Specifically, model parameters may include the weight matrix and bias vector of fully connected layers in a neural network, the values ​​of each dimension of the user latent vector and content latent vector in a matrix factorization model, and the feature selection threshold for split nodes and the predicted values ​​of leaf nodes in a tree model. Model parameters determine the specific mapping function from input features to output predictions, and their magnitude and distribution directly affect the model's predictive ability and generalization performance. Furthermore, model parameters are typically initialized randomly to start the training process, and then iteratively updated using optimization algorithms such as backpropagation and gradient descent, eventually converging to a numerical state that can well fit the training data.

[0135] Policy parameters refer to configurable parameters used to control training behavior and influence training performance during the training preference recall model process, but they are not integral parts of the model itself. Specifically, policy parameters can include hyperparameters such as learning rate, batch size, number of training epochs, regularization coefficient, dropout rate, optimizer type, and loss function weights, as well as data processing policy parameters such as negative sampling ratio, feature engineering methods, and data augmentation intensity. Policy parameters can adjust the speed, stability, and final model performance of the training process, but unlike model parameters, they cannot be directly optimized through gradient descent. Instead, they need to be adjusted using hyperparameter tuning methods such as grid search, random search, and Bayesian optimization, or based on empirical rules. Furthermore, the appropriate setting of policy parameters has a significant impact on whether the model can fully learn effective patterns in the data and avoid problems such as overfitting or underfitting.

[0136] In practical applications, the preference recall model can be trained using a variety of methods based on historical user preference tags and historical preference content, according to a preset training strategy.

[0137] One alternative approach is to design a phased training process and implement a pre-defined training strategy. Specifically, initial training data samples can be constructed first based on the association between historical user preference tags and historical preference content. For example, each user preference tag can be combined with historical preference content interacted with by the user to form positive samples, while negative samples are generated using negative sampling techniques. Then, the model parameters of the preference recall model are initialized, and the model parameters are iteratively updated to minimize prediction error according to optimization algorithms in the pre-defined training strategy, such as stochastic gradient descent or adaptive moment estimation. During training, model performance can be periodically evaluated on an independent validation set. Based on changes in offline evaluation metrics in the pre-defined training strategy, such as accuracy, recall, or mean precision, strategy parameters, such as learning rate decay rate or regularization strength, are dynamically adjusted. Finally, when the model's performance on the validation set reaches a pre-defined threshold or no longer improves, training is terminated, and the trained preference recall model is obtained.

[0138] Another alternative approach is to employ an automated machine learning framework to execute a pre-defined training strategy. Specifically, automated machine learning tools can be used, taking historical user preference tags and historical preference data as input, defining a model search space that includes different types of model architectures, such as matrix factorization models, deep neural network models, or gradient boosting tree models, and simultaneously defining a policy parameter search space, which may include learning rate ranges, batch size options, regularization coefficient ranges, etc. The automated machine learning framework can automatically sample different model configurations and policy parameter combinations within the search space, train preference-recall models for each combination, and perform offline evaluation on a validation set. Based on the evaluation results, the best-performing model configuration and policy parameter combination is selected, or the prediction results of multiple superior models can be fused using ensemble learning techniques. This method can systematically explore the optimal combination of model and policy, reducing the workload of manual parameter tuning.

[0139] Another alternative approach is to implement an iterative optimization training loop to execute a pre-defined training strategy. Specifically, training can begin with initial model and policy parameters. After each training round, the model's shortcomings in learning associations are analyzed based on offline evaluation results on the validation set, such as low prediction accuracy in identifying certain user preference tags and content associations. Then, the policy parameters are adjusted accordingly, such as increasing sample weights for tag-content pairs with poor prediction performance, adjusting the weight ratio of different association types in the loss function, or modifying the negative sampling strategy to generate more challenging negative samples. The adjusted policy parameters are then used to continue training the model and updating the model parameters. Multiple rounds of iterative training are performed until a satisfactory level of learning for each association is achieved, thus realizing feedback optimization of the training strategy.

[0140] Furthermore, differentiated training strategies can be designed based on the type and strength of the association. Specifically, for strong associations, such as the correspondence between user preference tags and frequently interacted content, higher confidence weights can be used for learning; for weak or implicit associations, more complex model structures, such as attention mechanisms or multi-head self-attention, can be used to capture them; for newly emerging associations or association patterns that evolve over time, online learning or incremental learning strategies can be used to continuously update the model. This approach can more precisely handle the learning needs of different types of associations and improve the model's ability to model complex user interest patterns.

[0141] In the embodiments of this specification, a preference recall model is trained according to a preset training strategy based on historical users' user preference tags and historical preference content. This preset training strategy includes optimizing the model parameters based on the correlation between historical users' user preference tags and historical preference content, and adjusting the strategy parameters of the preference recall model through offline evaluation. This achieves standardization, systematization, and controllability of the training process. The explicit preset training strategy ensures the repeatability and consistency of model training, avoiding randomness in the training process. Optimizing model parameters based on correlations enables the model to deeply learn and capture complex correspondence patterns between user preference tags and content, improving the model's accuracy in modeling user interests. Adjusting the strategy parameters through offline evaluation allows for dynamic optimization of the training configuration based on the model's performance on actual data, improving the model's generalization ability and providing a mechanism to ensure the stability and continuous optimization of the preference recall model's performance. This enables the trained model to more accurately and reliably recall highly relevant content based on the target user's user preference tags, thereby enhancing the accuracy and robustness of the entire content recommendation system in interest matching and providing a strong model foundation for improving recommendation performance in cold start scenarios for new users.

[0142] In an optional embodiment of this specification, before retrieving topic-related content from the content database based on the behavioral characteristics of the target topic behavior, the method further includes: Obtain the historical topic behaviors of historical users and the corresponding historical topic content; Train a topic recall model based on the behavioral characteristics and content of historical topics of historical users. Based on the behavioral characteristics of the target topic, retrieve topic-related content from the content database, including: By using a topic recall model, topic-related content is recalled from the content database based on the behavioral characteristics of the target topic.

[0143] Historical topic content refers to the collection of content that users actually interacted with within a past timeframe, relating to their historical topic-related behaviors. Specifically, historical topic content can include various forms of content such as articles clicked, videos viewed, images interacted with, or audio comments made by historical users. This content typically carries clear topic identifiers or tags. Historical topic content can reflect historical users' interests and consumption preferences regarding specific topics, providing authentic user-content interaction annotation data for training topic recall models. Furthermore, historical topic content usually undergoes certain quality screening, such as including only content with interaction duration exceeding a threshold, interaction types indicating clear interests, or interaction frequency meeting standards, to ensure the accuracy and representativeness of the training data.

[0144] A topic recall model is a machine learning model trained on the behavioral characteristics and content of historical user topic behaviors. It predicts topics a user might be interested in based on their behavioral features. Specifically, topic recall models can take various forms, including recurrent neural network models based on behavioral sequences, transformer models based on attention mechanisms, topic relationship models based on graph neural networks, and topic recommendation models based on multi-task learning. These models learn the complex mapping relationship between user behavioral features and topic content. When given the behavioral characteristics of a target user's target topic behavior, they can output a corresponding list of recommended topic content or a topic relevance score, thus achieving efficient matching from short-term user behavior to relevant topic content. Furthermore, topic recall models are typically deployed using offline training and online services, enabling continuous learning of historical user behavior patterns and optimization of topic recall performance.

[0145] In practical applications, topic recall models can be trained based on the behavioral characteristics and content of historical topics of historical users, which can be achieved through various methods.

[0146] One alternative approach is to train a topic recall model using a supervised learning framework. Specifically, behavioral features of historical users' past topic behaviors can be used as input features to the model, while the corresponding historical topic content can be used as the model's prediction target, constructing a supervised training sample set. Then, an appropriate machine learning algorithm, such as a deep neural network, gradient boosting decision tree, or support vector machine, is selected to optimize the model parameters on this training sample set, enabling the model to accurately predict the topic content that a user might be interested in given behavioral features. During training, negative sampling techniques can be introduced to handle massive amounts of candidate content, a multi-class cross-entropy loss function can be used to optimize the model's prediction accuracy, and early stopping on the validation set can prevent overfitting. This method can directly learn the correlation between behavioral features and topic content.

[0147] Another alternative approach is to train a topic recall model based on sequence modeling techniques. Specifically, historical user topic behaviors can be organized chronologically into behavioral sequences, with the corresponding historical topic content serving as the next prediction target in the sequence. Sequence models such as recurrent neural networks, long short-term memory networks, or transformers can be used to learn the implicit interest evolution patterns and topic shift rules within the behavioral sequences. The trained model can then predict one or more topics that a target user might be interested in based on their recent behavioral sequences. This method can capture the dynamic changes and continuity of user interests.

[0148] Another alternative approach is to train a topic similarity model according to a pre-defined training strategy. This strategy includes optimizing the model parameters of the topic recall model based on the correlation between behavioral features and content related to historical topic behaviors, and adjusting the strategy parameters through offline evaluation. Specifically, a systematic training process can be designed: first, an initial model is built based on the co-occurrence relationship between behavioral features and topic content; then, the model's performance is evaluated using offline evaluation metrics such as accuracy, recall, or mean precision; based on the evaluation results, model hyperparameters, feature engineering strategies, or sampling strategies are adjusted; and the training and evaluation process is iteratively executed until the model performance converges. This method achieves continuous optimization of the model's performance through a cycle of training and evaluation.

[0149] Furthermore, contrastive learning or self-supervised learning methods can be used to train the topic recall model. Specifically, positive sample pairs can be constructed, consisting of behavioral features of the same user and corresponding topic content, while negative sample pairs can consist of behavioral features of different users and topic content, or randomly combined unrelated pairs. The model is trained to make positive sample pairs closer together in the feature space and negative sample pairs farther apart. Alternatively, a mask prediction task can be designed, where random masks are used to mask some behavioral features so that the model can predict the corresponding topic content, learning the intrinsic relationship between behavior and content through self-supervised signals. This method can utilize the structural information of the data itself for training, reducing the dependence on large amounts of labeled data.

[0150] Once a topic recall model is trained, it can be used to recall topic-related content from the content database based on the behavioral characteristics of the target topic.

[0151] One alternative approach is to directly generate a list of recommended topic content through model inference. Specifically, the behavioral characteristics of the target user's target topic behavior can be input into a trained topic recall model. The model outputs interest scores or ranking scores for all candidate topic content in the content database. Then, the content is sorted from highest to lowest score, and the top-ranked content is selected as the topic recall content. To improve recall efficiency, approximate nearest neighbor search techniques can be used to accelerate the retrieval process of a large-scale content database, such as using locality-sensitive hashing or product quantization. This method achieves end-to-end topic content recall and simplifies the recall process.

[0152] Another approach is to combine topic recall models with topic expansion techniques for content retrieval. Specifically, the topic recall model can first identify core topics or sets of topics based on behavioral characteristics of target topic behavior; then, based on topic similarity relationships or topic association graphs, it can expand to include derivative topics or sub-topics related to the core topics; finally, using the expanded set of topics as query conditions, a search can be performed in a content database to retrieve relevant topic-recalled content. This method enables topic expansion based on core interests, improving the diversity and coverage of retrieved content.

[0153] Another alternative approach is to employ a multi-granularity topic recall strategy, where the topic recall model operates at different granularities. Specifically, at a coarse-grained level, relevant topic categories or topic groups can be recalled based on behavioral features, and then at a fine-grained level, specific content items can be recalled based on more specific behavioral features or intra-topic features. The topic recall model can be applied to the coarse-grained and fine-grained recall stages separately, or designed as a multi-task model to handle recall tasks at different granularities simultaneously. This approach enables hierarchical recall from macro-level topic preferences to micro-level content preferences.

[0154] Furthermore, topic recall results can be dynamically adjusted by incorporating real-time behavioral feedback. Specifically, the input to the topic recall model can be updated in real time based on the latest behavioral characteristics of the target users to obtain the latest recall results; alternatively, an online learning version of the topic recall model can be designed to quickly adjust model parameters based on real-time user interaction feedback, achieving dynamic optimization of recall results; or the output of the topic recall model can be integrated with real-time popular topics and trending topics to improve the freshness and timeliness of the recalled content. This approach can enhance the responsiveness of topic recall to changes in users' real-time interests.

[0155] In this embodiment, before recalling topic-related content from the content database based on the behavioral characteristics of target topic behavior, historical user topic behaviors and corresponding historical topic content are obtained. A topic recall model is trained based on this data, and then the model is used to recall topic-related content based on the behavioral characteristics of target topic behavior. This achieves intelligent topic content recall based on a machine learning model, deeply exploring the complex correlation patterns between user behavior characteristics and topic content, enabling more accurate and personalized topic content matching. By using the topic recall model, the retrieval problem of large-scale content databases can be effectively handled, improving recall efficiency while ensuring recall quality. This enhances the intelligence level and stability of the topic recall process, making the topic-related content recalled for the target user more closely aligned with their short-term dynamic interests. This provides a higher quality and more timely candidate content foundation for subsequent ranking and recommendation, thereby improving the freshness, diversity, and user satisfaction of the entire content recommendation system.

[0156] In one optional embodiment of this specification, a topic recall model is trained based on the behavioral characteristics and content of historical topics of historical users, including: Based on the historical topic content corresponding to the historical topic behavior of historical users, a topic similarity model is trained according to a preset training strategy. The preset training strategy includes optimizing the model parameters of the topic recall model based on the correlation between the behavioral features of historical topic behavior and the historical topic content, and adjusting the strategy parameters of the topic recall model through offline evaluation.

[0157] The correlation between the behavioral characteristics of historical topics and their content represents various forms of connection, such as statistical correlation, semantic similarity, temporal continuity, or causal inference, between the behavioral characteristics of historical users and the topic content they actually interacted with. Specifically, this correlation can manifest as a high co-occurrence probability between specific types of behavioral characteristics, such as high-frequency clicking behavior, and specific categories of topic content. For example, a strong correlation exists between a user's frequent clicking behavior on "fitness tutorials" and their subsequent browsing of "healthy eating" topic content. It can also be shown by the close distance between the behavioral feature vector and the topic content feature vector in the latent semantic space, indicating semantic similarity between the behavioral intent and the content theme. Furthermore, it can be shown by specific patterns in the behavioral sequence predicting the likelihood of users interacting with certain topic content. This correlation reflects the intrinsic connection between users' short-term behavioral patterns and content interests, providing crucial learning signals for training topic similarity models. Moreover, the strength, type, and dynamic changes of this correlation can be quantitatively analyzed and modeled using various techniques such as correlation analysis, mutual information calculation, attention weight visualization, or causal inference models.

[0158] In practical applications, topic similarity models can be trained using a variety of methods based on the historical topic content corresponding to the historical topic behavior of historical users and according to a preset training strategy.

[0159] One alternative approach is to train a topic similarity model using a supervised learning-based similarity learning framework. Specifically, behavioral features of historical users' past topic behaviors can be used as model input, while the corresponding historical topic content features are used as the learning objective to construct positive sample pairs. Simultaneously, negative sample pairs combining behavioral features with unrelated topic content are generated using negative sampling techniques. By designing similarity metrics such as cosine similarity or the inverse of Euclidean distance, the model is trained to achieve higher similarity scores for positive sample pairs and lower similarity scores for negative sample pairs. The pre-defined training strategy can define loss functions such as contrastive loss or triplet loss, optimizing model parameters to minimize the loss value, and adjusting strategy parameters such as the negative sampling ratio, margin size, or learning rate scheduling through offline evaluation. This method can directly learn the similarity mapping relationship between behavioral features and topic content.

[0160] Another alternative approach is to train a topic similarity model using a sequence-to-sequence modeling method. Specifically, historical user topic behaviors can be organized chronologically into behavioral feature sequences, with the corresponding historical topic content sequences serving as the output targets. An encoder-decoder architecture is employed, where the encoder encodes the behavioral feature sequences into context vectors, and the decoder generates a probability distribution of the topic content sequences based on these vectors. The pre-defined training strategy can optimize model parameters based on the alignment relationship between the behavioral feature sequences and the topic content sequences, employing maximum likelihood estimation or reinforcement learning strategies for training. Furthermore, the strategy parameters, such as the learning rate decay rate or regularization strength, are dynamically adjusted based on changes in offline evaluation metrics (e.g., accuracy, recall, or mean precision) within the pre-defined training strategy. This method effectively captures the temporal dependencies between behavioral sequences and topic content.

[0161] Another alternative approach is to train a topic similarity model using graph neural networks. Specifically, a heterogeneous graph network can be constructed, where nodes can include user nodes, behavioral feature nodes, and topic content nodes, and edges can represent various relationships such as user-generated behaviors and related content. A graph neural network model is used to learn the embedding representations of nodes within the graph. Furthermore, similarity is calculated based on the distance between behavioral feature nodes and topic content nodes in the embedding space. The pre-set training strategy can perform link prediction based on existing relationships in the graph, optimizing model parameters to accurately predict missing edges. Offline evaluation is used to adjust strategy parameters such as the number of convolutional layers, the number of attention heads, or the dropout rate. This method leverages higher-order relationships within the graph structure to enhance similarity learning.

[0162] Furthermore, a multi-task learning framework can be used to train the topic similarity model. Specifically, the main task can be designed as similarity prediction between behavioral features and topic content, while auxiliary tasks include behavior classification, topic classification, or user clustering. Multiple related tasks are learned simultaneously by sharing some model parameters, and the generalization ability of the main task is improved through knowledge transfer between tasks. The loss weights of each task can be dynamically adjusted in the preset training strategy, and the multi-task learning strategy is optimized based on offline evaluation results. This method can enhance the learning effect of the similarity model by utilizing the supervision signals of related tasks.

[0163] In the embodiments of this specification, a topic similarity model is trained based on historical topic content corresponding to historical user behavior, according to a preset training strategy. This preset training strategy includes optimizing model parameters based on the correlation between behavioral features and historical topic content, and adjusting the strategy parameters through offline evaluation, thus achieving a systematic and controllable training process. By explicitly using the correlation between behavioral features and topic content as a training signal, the model can deeply learn and quantify the complex correspondence between short-term user behavior patterns and topic interests, improving the accuracy of similarity calculation. The standardized guidance of the preset training strategy ensures the repeatability and stability of the training process. Dynamic adjustment of strategy parameters through offline evaluation optimizes the training configuration based on the model's performance on actual data, effectively balancing the model's fitting ability and generalization performance. This provides a more reliable and detailed similarity measurement basis for topic recall, thereby improving the accuracy, relevance, and timeliness of recalled content, enhancing the accuracy and robustness of the entire content recommendation system in interest matching, and providing a strong model foundation for improving recommendation performance in cold start scenarios for new users. In an optional embodiment of this specification, before sorting the preference-recalled content and topic-recalled content to determine the recommended content for the target user, the method further includes: Obtain the geographic preference tags of the target users; Based on geographic preference tags, the preference recall content is filtered to obtain the filtered preference recall content.

[0164] The filtered preference recall content is a collection of content that better matches the user's geographical location characteristics, obtained by filtering, selecting, or reordering the original preference recall content based on the target user's geographic preference tags. Specifically, the filtered preference recall content can include content retained after geographic tag matching, content ranked high based on geographic relevance scores, or content that meets specific geographic zoning conditions. The filtered preference recall content can meet users' needs for regionally relevant, culturally relevant, or localized content, and is provided to target users as an important component of the recommended content. Furthermore, the filtered preference recall content is usually more regionally targeted, reflecting the cultural characteristics, consumption habits, or interests of the user's geographical region, providing target users with a content recommendation experience more aligned with their regional background.

[0165] In practical applications, filtering preference recall content based on geographic preference tags to obtain the filtered preference recall content can be achieved through various methods.

[0166] One alternative approach is to filter content based on direct matching of geographic tags. Specifically, each content item in the content database can be tagged with geographic tags, such as the associated city, region, province, or cultural sphere. After obtaining the set of preference-recall content, each piece of content in the set is iterated through, checking whether its geographic tags match or are related to the target user's geographic preference tags. Content with matching geographic tags is retained, while content that does not match is filtered out, thus obtaining the filtered preference-recall content. For example, if the target user's geographic preference tag is "North China users," then only content with geographic tags containing "Beijing," "Tianjin," and "Hebei" is retained. This method is simple and direct, and can quickly filter out geographically irrelevant content.

[0167] Another alternative approach is to filter and reorder content based on geographic relevance scores. Specifically, a geographic relevance calculation function can be designed to calculate a relevance score based on the matching degree between the content's geographic features and the user's geographic preference tags. This could be achieved using distance metrics based on geographic hierarchy, statistical probabilities based on geographic tag co-occurrence, or vector calculations based on geographic semantic similarity. A geographic relevance score is then calculated for each piece of content in the preference recall set. A threshold can then be set to filter out content with scores below that threshold, or the content can be reordered from highest to lowest score, selecting the top-ranked content as the filtered preference recall content. This method enables a more refined measurement of geographic relevance.

[0168] Another alternative approach is to employ a tiered filtering strategy based on geographic partitioning. Specifically, the content retrieved based on geographic preference tags can be divided into different geographic partitions, such as "strongly relevant," "moderately relevant," and "weakly relevant" partitions. Then, content is selected from each partition according to a preset quota strategy, for example, more content is selected from the strongly relevant partitions, less from the weakly relevant partitions, or none at all. The selected content is then merged to form the filtered content retrieved. This method can appropriately control content diversity while ensuring geographic relevance.

[0169] Furthermore, comprehensive filtering can be achieved by combining multi-dimensional features. Specifically, geographic preference tags can be combined with other user characteristics such as interest intensity and behavior frequency to design a multi-objective filtering function, considering multiple dimensions such as geographic relevance, content quality, and interest matching. Alternatively, a geographic-aware filtering model can be trained, which takes content features and user geographic preference tags as input and outputs a binary decision or retention probability on whether the content should be retained. The model output is then used to filter content recalled based on preferences. This approach enables more intelligent and comprehensive content filtering.

[0170] In this embodiment, before sorting the preference recall content and topic recall content, the geographic preference tags of the target user are obtained, and the preference recall content is filtered based on these geographic preference tags to obtain filtered preference recall content. This achieves the effective application of geographic dimension features in the quality control of preference recall content. Through the filtering function of geographic preference tags, content that is irrelevant or has low relevance to the user's regional background can be filtered out, while content that is more compatible with the user's region in terms of geographic culture, consumption habits, or interests can be retained. This improves the regional adaptability and cultural relevance of preference recall content, so that the preference recall content that finally enters the sorting stage not only matches the user's long-term interest tags but also conforms to the user's geographic characteristics. This achieves dual optimization of the interest dimension and the geographic dimension, which can effectively enhance the regional targeting of recommended content, improve new users' recognition and satisfaction with the recommendation results, and thus improve the user's overall content acquisition experience.

[0171] In an optional embodiment of this specification, before sorting the preference-recalled content and topic-recalled content to determine the recommended content for the target user, the method further includes: Obtain the geographic topic tags of the target users; Based on geographic topic tags, retrieve popular regional topic content from the content database; The intersection of the topic recall content and the regional popular topic content is used to obtain the filtered topic recall content.

[0172] Geographic hashtags are composite tags that combine geographic location information with topic characteristics, used to describe users' interests and preferences regarding specific topics from a geographical perspective. Specifically, geographic hashtags can include tags combining specific regions and topic categories, such as "Beijing Food," "Shanghai Tourism," "Guangzhou Local Life," and "Chengdu Leisure and Entertainment"; they can also include tags combining broad regions and specialized topics, such as "North China Technology" and "Yangtze River Delta Finance." Geographic hashtags can reflect users' attention to and inclination towards specific topics in different regions, providing a decision-making basis for content recommendations that is both geographically appropriate and topic-specific. Furthermore, geographic hashtags are typically generated based on a combination of user geographic location information, historical topic behavior data, and regional topic popularity, representing users' interests at the intersection of geography and topic.

[0173] Regional trending topics are collections of content that have received widespread attention, generated high-frequency interaction, or shown a significant growth trend in a specific geographical area within a recent timeframe. Specifically, regional trending topics can include various forms of content such as recently discussed local news from a city, reports of popular cultural activities in a region, popular travel guides from a province, or discussions of social events of concern in a major region. Regional trending topics can reflect the collective interests and content consumption trends of user groups in different geographical areas, providing timely and regionally distinctive content sources for content recommendations. Furthermore, regional trending topics are typically identified and updated through real-time statistics of regional user behavior data, monitoring changes in topic popularity, or analyzing content dissemination trajectories, enabling dynamic capture of the evolution of regional hot topics.

[0174] The filtered topic-recall content is a collection of content obtained by intersecting the original topic-recall content with popular regional topics. This content aligns with both the user's personal interests and regional popularity characteristics. Specifically, the filtered topic-recall content can include content retained after intersection operations, content selected after re-ranking based on the intersection results, or content filtered according to dual conditions. The filtered topic-recall content can simultaneously satisfy both the user's personalized interests and the regional group's focus on trending topics, serving as a crucial component of the final recommendation content provided to the target user. Furthermore, the filtered topic-recall content typically possesses high regional relevance and timeliness, reflecting current trending topics in the user's region while maintaining a match with the user's personal interests, providing the target user with a content recommendation experience that combines personalization and regional characteristics.

[0175] In practical applications, obtaining the geographic topic tags of target users can be achieved through various methods.

[0176] One alternative approach is to generate geographic topic tags through correlation analysis between user geolocation information and topic behavior data. Specifically, this involves obtaining the target user's real-time geolocation information, such as GPS coordinates, IP address location, or registered city information, while simultaneously analyzing their recent topic behavior data, such as the categories of topics they clicked, browsed, or interacted with. The geolocation is then mapped to a specific region, and the topic behavior is categorized into topic categories. Finally, the regional information and topic categories are combined to form geographic topic tags. For example, if a user is located in Beijing and frequently browses food content, a tag titled "Beijing Food Topic" is generated. This method can generate personalized geographic topic tags based on the user's actual location and behavior.

[0177] Another alternative approach is to predict geographic topic tags using a regional topic popularity model. Specifically, a regional topic popularity prediction model can be trained based on historical data. This model takes user characteristics such as demographic attributes, interest tags, and behavioral patterns as input and outputs geographic topic tags that the user may be interested in and their probabilities. For a target user, their characteristics are input into the model to obtain a predicted set of geographic topic tags. The model can consider the differences in topic popularity across different regions and recommend popular topic tags from the target user's region. This method leverages machine learning models to uncover users' potential geographic topic interests.

[0178] Another alternative approach is to obtain the target user's geographic topic tags through the migration of geographic topic tags from similar users. Specifically, based on other characteristics of the target user, such as interests, behavioral patterns, and demographic attributes, similar users can be found among historical users. Then, the geographic topic tags of these similar users are aggregated, and estimated geographic topic tags are generated for the target user through a voting mechanism, weighted average, or tag fusion algorithm. This method is particularly suitable for new users or users with sparse behavior, and can make reasonable inferences about geographic topic interests based on user similarity.

[0179] Furthermore, the accuracy of geographic topic tags can be improved through multi-source data fusion. Specifically, multiple information sources such as real-time geographic location data, historical topic behavior data, social relationship data, and content interaction data can be merged, with each data source assigned a confidence weight, and the final geographic topic tags generated through a weighted fusion algorithm. Alternatively, a geographic topic tag generation model can be established, which takes multi-source features as input and outputs the probability distribution of geographic topic tags. This method can comprehensively utilize multiple information sources to improve the accuracy and robustness of geographic topic tags.

[0180] Once the geographic topic tags of the target users are obtained, popular regional topic content can be retrieved from the content database based on the geographic topic tags.

[0181] One alternative approach is to directly retrieve the regional trending content index from the content database using geographic topic tags. Specifically, a regional trending topic index can be pre-built for the content database, organized by geographic topic tags, with each tag storing a list of recently popular content in that region. When it's necessary to retrieve regional trending topic content, the corresponding index entry can be queried based on the target user's geographic topic tag, obtaining the stored content list as the retrieval result. This method is simple and efficient, enabling rapid retrieval of pre-calculated regional trending content.

[0182] Another option is to calculate regional topic popularity in real time and dynamically recall content. Specifically, this involves monitoring user interaction data in each region in real time, and statistically analyzing popularity metrics such as click-through rate, interaction rate, and sharing rate for different topics within a specific region. When it's necessary to recall popular regional topics based on geographic hashtags, the system filters for corresponding regions based on the regional information in the hashtags, filters for relevant topics based on the topic information in the hashtags, and then sorts them according to real-time popularity metrics, selecting the top-ranked content as the recall result. This method provides the latest regional trending content and is highly timely.

[0183] Another alternative approach is to utilize a regional topic popularity model to predict trending content. Specifically, a regional topic popularity prediction model can be trained, taking geographic topic tags and temporal context as input and outputting a list of candidate content that may become popular. Based on the target user's geographic topic tags, the model predicts regional topic content that may be popular within the current time period, and then retrieves this content from a content database as the recall result. This method can predict potential trending content and build up content reserves in advance.

[0184] Furthermore, a comprehensive recall strategy can be implemented by combining multiple dimensions of popularity metrics. Specifically, a comprehensive popularity scoring function can be designed, taking into account various indicators such as absolute popularity (e.g., total interactions), relative popularity (e.g., interaction growth rate), dissemination popularity (e.g., sharing chain length), and timeliness popularity (e.g., freshness decay). Relevant content and regions can then be filtered based on geographic topic tags, and a comprehensive popularity score can be calculated for each piece of content. Popular topics within a region can then be recalled based on these scores. This method provides a more comprehensive assessment of the popularity value of content.

[0185] Once the popular topics in the region have been recalled, the intersection of the recalled topics and the popular topics in the region can be used to obtain the filtered recalled topics.

[0186] One alternative approach is to directly perform set intersection operations using content identifiers. Specifically, separate sets of content identifiers can be created for topic-recall content and regional trending topics; then, the intersection of the two sets is calculated, i.e., the content identifiers that appear in both sets are found; finally, complete content information is retrieved from the content database based on these content identifiers to form the filtered topic-recall content. This method is direct, computationally efficient, and suitable for situations where the content identifier system is unified.

[0187] Another alternative approach is to perform soft intersection filtering using content similarity calculations. Specifically, similarity scores can be calculated between the retrieved content and popular regional topics, such as cosine similarity based on content feature vectors, Jaccard similarity based on topic tags, or semantic similarity based on semantic embeddings. Then, a similarity threshold is set, retaining content with similarities higher than the threshold as the intersection result; alternatively, the retrieved content can be reordered according to similarity scores, selecting the top-ranked content as the filtered retrieved content. This method can handle situations where the content is not entirely identical but highly related.

[0188] Another alternative approach is to employ a multi-stage filtering strategy to achieve the intersection operation. Specifically, the topic-retrieved content can first be initially filtered based on geographic topic tags, retaining content related to those tags. Then, the filtered content can be further filtered based on regional popularity indicators, retaining content with higher popularity. Finally, the results of the two filtering processes can be integrated to obtain the final filtered topic-retrieved content. This method breaks down the intersection operation into multiple conditional filtering steps, making it easier to understand and implement.

[0189] Furthermore, an approximation of the intersection effect can be achieved based on a weighted fusion strategy. Specifically, an interest relevance score can be calculated for topic-retrieved content, and a popularity score can be calculated for regionally popular topics. Then, a fusion function is designed to combine the two scores into a comprehensive score, such as weighted summation, product, or harmonic average. The content is then sorted according to the comprehensive score, and the top-ranked content is selected as the filtered topic-retrieved content. This method provides a more flexible filtering mechanism while retaining the intersection concept.

[0190] In this embodiment, before sorting the preferred and topic-recalled content, the target user's geographic topic tags are obtained. Based on these tags, popular regional topics are retrieved from the content database. The intersection of the retrieved topic content and the popular regional topics yields the filtered topic-recalled content, achieving dual optimization filtering based on regional hotspots and personal interests. The introduction of geographic topic tags ensures that topic recall not only considers users' personal interests but also incorporates regional hotspot information, enhancing the regional relevance and timeliness of the recalled content. The retrieval of popular regional topics ensures that the recommended content reflects the current hotspots and trends in the user's region. The intersection operation achieves an effective balance between personal interests and group hotspots, ensuring that the topic-recalled content entering the sorting stage not only matches users' personalized interests but also possesses regional hotspot characteristics. This significantly improves the quality and attractiveness of the topic-recalled content, effectively enhancing new users' recognition and participation in the recommended content, and improving the content adaptation capability and user experience satisfaction of the recommendation system during the cold start phase.

[0191] In one optional embodiment of this specification, the preferred recall content and topic recall content are sorted to determine recommended content for the target user, including: Using a coarse ranking model, the preferred recall content and topic recall content are coarsely ranked based on preset coarse ranking indicators to obtain coarsely ranked recommended content. The preset coarse ranking indicators include at least one of the following: estimated click-through rate, estimated browsing time, and estimated interaction rate. By using a fine-ranking model, the coarse-ranked recommended content is fine-ranked based on preset fine-ranking indicators to obtain recommended content for the target users.

[0192] Coarse-ranking models are machine learning models used for initial screening and ranking of recalled content. Specifically, coarse-ranking models can include logistic regression models, gradient boosting decision tree models, deep neural network models, etc. Based on preset coarse-ranking metrics, coarse-ranking models can quickly evaluate the relevance, quality, or user interest matching degree of a large number of candidate contents, filtering out a higher-quality subset of candidates from massive amounts of recalled content, thus reducing the processing scale for the subsequent fine-ranking stage. Furthermore, coarse-ranking models are typically designed with lightweight, high-efficiency model structures to balance ranking effectiveness and computational performance, meeting the response time requirements of real-time recommendation scenarios.

[0193] Predefined coarse-ranking metrics are a set of predefined quality measures used by coarse-ranking models to evaluate and rank candidate content. Specifically, these metrics may include estimated click-through rate, estimated browsing time, and estimated interaction rate. Predefined coarse-ranking metrics can quantify and predict the intensity of users' potential interest in candidate content and the likelihood of interaction, providing the coarse-ranking model with calculable optimization objectives and ranking criteria. Furthermore, these metrics are typically derived from historical user behavior data statistics or predicted through machine learning models, reflecting the user appeal of content across different dimensions.

[0194] Predicted click-through rate (CTR) is a numerical value that predicts the probability of a user clicking on specific content. Specifically, CTR can be calculated using statistical models such as logistic regression or deep learning models such as deep interest networks. CTR characterizes the potential of content to attract users' initial attention and serves as an important indicator for evaluating content appeal.

[0195] Estimated browsing time is a numerical value predicting the duration a user will browse or watch specific content. Specifically, estimated browsing time can be predicted using regression models such as linear regression or sequence models such as recurrent neural networks. Estimated browsing time reflects the depth of a user's interest in the content and its informational value, serving as an important indicator for assessing content stickiness and user engagement.

[0196] The estimated interaction rate is a numerical value that predicts the probability of users performing interactive actions such as liking, commenting, sharing, and saving specific content. Specifically, the estimated interaction rate can be calculated using a multi-classification model or a probabilistic prediction model. The estimated interaction rate can characterize the potential of content to stimulate users' active participation and emotional resonance, serving as one of the important indicators for evaluating the social value of content and user engagement.

[0197] Fine-ranking models are machine learning models used to refine the ranking of candidate content after coarse ranking. Specifically, fine-ranking models can include deep ranking models, multi-task learning models, reinforcement learning models, etc. Based on richer features, more complex model structures, and more refined pre-defined ranking metrics, fine-ranking models can more accurately and personally rank the coarsely ranked recommendations, generating a final user recommendation list. Furthermore, fine-ranking models typically have higher model complexity and computational cost, but they can achieve more accurate user interest modeling and content value assessment, improving the final quality of the recommendation results.

[0198] Predefined ranking metrics are a set of refined quality standards that the ranking model uses to evaluate and rank candidate content. Specifically, these metrics may include comprehensive satisfaction scores, long-term value predictions, diversity measures, and fairness constraints. These metrics comprehensively measure the overall value of content in terms of user satisfaction, platform goals, and ecosystem health, providing a multi-objective optimization basis for the ranking model. Furthermore, these metrics are typically integrated using multi-objective optimization algorithms or weighted fusion strategies to balance the short-term effects and long-term goals of the recommendation system.

[0199] In practical applications, the coarse ranking model can be used to coarsely rank the content recalled by preferences and the content recalled by topics based on preset coarse ranking indicators to obtain coarsely ranked recommended content. This can be achieved through various methods.

[0200] One alternative approach is to perform coarse ranking by combining model scoring with threshold filtering. Specifically, preference-recalled content and topic-recalled content can be merged into a unified candidate content pool. The feature vector of each piece of content is extracted and input into the coarse ranking model. The model outputs a comprehensive score or predicted values ​​for each metric based on at least one of the preset coarse ranking indicators, such as estimated click-through rate, estimated browsing time, and estimated interaction rate. Then, a score threshold or a combination of thresholds for each indicator is set to filter out content below the threshold, retaining content above the threshold to form the coarse ranking recommendation. This method can quickly filter low-quality content and reduce the burden of subsequent processing.

[0201] Another alternative approach is to use a multi-metric weighted fusion ranking method for coarse ranking. Specifically, the predicted value for each candidate content on various preset coarse ranking metrics can be calculated separately, such as estimated click-through rate, estimated browsing time, and estimated interaction rate obtained through independent prediction models. Then, preset weight coefficients are assigned to each metric, and a weighted comprehensive score is calculated. Finally, the content is ranked from highest to lowest according to the comprehensive score, and a certain number of top-ranked content are selected as coarse ranking recommendations. This method comprehensively considers the attractiveness of content across different dimensions, achieving a more balanced coarse ranking effect.

[0202] Another option is to perform coarse ranking based on stratified sampling or quota control strategies. Specifically, preference-recalled content and topic-recalled content can be stratified by source, type, or theme; then, within each stratum, they can be sorted based on preset coarse ranking indicators, and content can be selected from each stratum according to preset quota ratios, such as selecting a certain proportion from the preference-recalled content stratum and a certain proportion from the topic-recalled content stratum; the selected content is then merged to form the coarse-ranked recommended content. This method can consider the diversity and balance of content during the coarse ranking stage, avoiding excessive concentration of a single type of content.

[0203] In addition, a lightweight model can be used for rapid inference in coarse ranking. Specifically, a lightweight neural network model optimized for coarse ranking can be designed, which significantly reduces computational complexity while maintaining a certain level of prediction accuracy; alternatively, model distillation techniques can be used to transfer knowledge from complex fine ranking models to lightweight coarse ranking models; during coarse ranking, this lightweight model is used to quickly score and rank all candidate content. This approach can achieve efficient coarse ranking with limited computing resources.

[0204] Once the coarse-ranked recommended content is obtained, the fine-ranking model can be used to refine the coarse-ranked recommended content based on preset fine-ranking indicators, thereby obtaining recommended content for the target users. This can be achieved through various methods.

[0205] One alternative approach is to employ a deep ranking model for end-to-end fine-tuning. Specifically, multi-dimensional features of the coarse-ranked recommended content, including content features, user features, contextual features, and cross-features, can be extracted and input into the deep ranking model. This model performs multi-objective learning based on preset fine-tuning metrics such as overall satisfaction and long-term value, outputting a final ranking score for each piece of content. The content is then sorted from highest to lowest score to generate a recommended content list tailored to the target user. This method fully leverages the powerful feature learning capabilities of deep models to achieve highly personalized and accurate ranking.

[0206] Another alternative approach is to employ a multi-task learning framework for fine-grained ranking. Specifically, a multi-task fine-grained ranking model can be designed to simultaneously learn multiple related ranking tasks, such as click-through rate prediction, browsing duration prediction, interaction rate prediction, and satisfaction prediction. Each task corresponds to a different preset ranking metric. The model shares the underlying feature representation layer but has an independent task output layer. During training, a multi-task loss function is used for joint optimization, and during inference, the outputs of each task are weighted and fused to generate the final ranking result. This method leverages the correlation between tasks to improve the model's generalization ability and ranking performance.

[0207] Another alternative approach is to employ reinforcement learning strategies for dynamic ranking. Specifically, the ranking process can be modeled as a sequential decision-making problem. The ranking model can select a ranking action based on the current state, such as user characteristics, content characteristics, and historical interactions. It obtains reward signals through interaction with the environment, such as real-time user feedback. These reward signals are designed based on preset ranking metrics, such as user satisfaction and long-term retention. The model optimizes the ranking strategy using reinforcement learning algorithms, such as policy gradient methods, to achieve dynamic ranking that maximizes long-term benefits. This approach can adapt to constantly changing user interests and content ecosystems.

[0208] Furthermore, online fine-tuning optimization can be achieved by incorporating real-time feedback. Specifically, an online learning version of the fine-tuning model can be deployed to collect real-time user interaction feedback on recommended content, such as clicks, views, and interactions. Based on this feedback data, model parameters can be quickly updated, and the ranking strategy adjusted. Alternatively, an A / B testing framework can be designed to deploy multiple variations of the fine-tuning model simultaneously, selecting the optimal model based on real-time performance data. This approach enables the fine-tuning model to quickly adapt to changes in user interests and content trends, maintaining the freshness and accuracy of the ranking results.

[0209] In the embodiments described in this specification, a coarse-ranking model is used to coarsely rank the content recalled by preferences and the content recalled by topics based on preset coarse-ranking indicators to obtain coarse-ranked recommended content. Then, a fine-ranking model is used to fine-rank the coarse-ranked recommended content based on preset fine-ranking indicators to obtain recommended content for the target user. This achieves an efficient and accurate two-stage ranking mechanism. In the coarse-ranking stage, a lightweight model and key indicators are used to quickly filter massive amounts of recalled content, which greatly reduces the scale of content that needs to be finely processed and improves the efficiency and real-time performance of the overall ranking process. In the fine-ranking stage, a more complex model and more comprehensive indicators are used to finely evaluate high-quality candidate content, which achieves a highly personalized and multi-objective optimized final ranking. Through the hierarchical design of preset coarse-ranking indicators and preset fine-ranking indicators, the ranking process can be optimized for the objectives of different stages. Coarse-ranking focuses on rapid screening and efficiency, while fine-ranking focuses on accurate matching and comprehensive value. Together, they improve the overall performance of the recommendation system in terms of user satisfaction, platform goal achievement, and ecosystem health maintenance. Especially in the scenario of new user cold start, it can quickly generate high-quality and personalized recommendation results based on limited user data, which significantly enhances the overall effect of the content recommendation method and the user experience.

[0210] In one optional embodiment of this specification, before using the fine-ranking model to fine-rank the coarse-ranked recommended content based on preset fine-ranking indicators to obtain recommended content for the target user, the method further includes: Based on the preset regional minimum quota, the preferred recall content and topic recall content are filtered to obtain the filtered coarse-ranked recommended content, which includes content that meets the regional minimum quota.

[0211] Regional minimum guarantee quotas are minimum quantity or percentage guarantees set during the coarse-ranking process for content retrieved from two sources: preferred content and topic-based content. Specifically, regional minimum guarantee quotas can include absolute quantity quotas, such as ensuring a certain amount of content is retained for the fine-ranking stage; relative percentage quotas, such as ensuring preferred and topic-based content constitutes a certain percentage of the coarse-ranking recommendations; or hybrid quotas, such as setting both quantity and percentage minimums. Regional minimum guarantee quotas prevent premature elimination of content retrieved from these two sources due to low scores from the coarse-ranking model, ensuring that the recall strategy specifically designed for new users is effectively implemented and enters the subsequent fine-ranking stage. Furthermore, regional minimum guarantee quotas are typically dynamically adjusted based on service strategies, model performance, and user experience needs, ensuring the effectiveness of the recall strategy while avoiding over-protection that leads to low-quality content entering the fine-ranking stage.

[0212] The pre-ranked recommended content refers to the set of content that can enter the fine-ranking stage after screening and protecting preference-recalled content and topic-recalled content based on a preset regional minimum quota. Specifically, the pre-ranked recommended content can include preference-recalled content and topic-recalled content retained through the quota guarantee, as well as other content selected based on the pre-ranking score. The pre-ranked recommended content ensures that candidate content from preference-recalled and topic-recalled content is fully protected and prioritized in the pre-ranking stage, avoiding the impact of scoring bias caused by the sparse behavior of new users on the actual effectiveness of the recall strategy.

[0213] In practical applications, there are several ways to filter preferred and topic-based content based on a preset regional minimum quota to obtain coarsely ranked recommended content.

[0214] One alternative approach is to employ a minimum retention mechanism for selection. Specifically, a minimum number of preferred and topic-based content items can be preset for each to enter the fine-grained ranking. For example, a minimum number of preferred and topic-based content items can be retained. During the coarse-grained ranking process, all candidate content is first sorted according to the coarse-grained ranking model's score. Then, it is checked whether the number of preferred and topic-based content items retained in the ranking results reaches the preset minimum. If the number of retained content items in a particular category is lower than the minimum retention, the remaining candidates from that category are supplemented to reach the minimum retention based on the coarse-grained ranking score from highest to lowest. Finally, the minimum retention content is merged with the content normally selected in the coarse-grained ranking to form the filtered coarse-grained recommended content. This method is simple and direct, and can precisely control the minimum number of content items in each category.

[0215] Another alternative approach is to employ a minimum percentage guarantee mechanism for screening. Specifically, a minimum percentage of preference-recalled content and topic-recalled content can be preset in the coarse-ranked recommended content. For example, it can be set that preference-recalled content and topic-recalled content must together occupy at least a certain percentage. During the coarse-ranking process, all candidate content is first sorted according to the coarse-ranking model score. Then, it is checked whether the combined percentage of preference-recalled content and topic-recalled content in the ranking results reaches the preset percentage. If the percentage is lower than the preset percentage, the remaining candidates from these two sources are supplemented according to the coarse-ranking score from high to low until the preset percentage is reached. Finally, the content that meets the percentage requirement is merged with other coarse-ranked content to form the filtered coarse-ranked recommended content. This method can flexibly adapt to changes in the total amount of content under different coarse-ranking scenarios, maintaining the relative weight of the two sources of recalled content in the overall picture.

[0216] Another option is to use a weighted score compensation mechanism to achieve a guaranteed quota effect. Specifically, during the coarse-ranking stage, score compensation or weight adjustments can be applied to the preferred and topic-based recalled content. For example, these two types of recalled content can be multiplied by a weight coefficient greater than one, or a positive bias term can be added to the original coarse-ranking score, thereby increasing their priority in the coarse-ranking. After the scores are sorted according to the weighted compensation, content is selected according to the regular coarse-ranking filtering rules. Since these two types of recalled content have received a score bonus, they will naturally occupy more slots in the ranking results, thus achieving the guaranteed quota effect. This method does not require explicit quota counting and can achieve guaranteed protection in a smoother way, avoiding the ranking inconsistency problem caused by hard quotas.

[0217] In addition, a combination of tiered quotas and dynamic adjustments can be used for screening. Specifically, based on content quality metrics such as estimated click-through rate or estimated interaction rate, preferred and topic-based recalled content can be divided into high-quality and ordinary quality layers, with different minimum quotas set for each layer, giving higher minimum quotas to the high-quality layer. Alternatively, the strength of the minimum quota can be dynamically adjusted based on the target users' new user stage or the sparsity of their behavior, providing stronger minimum protection for new users with sparser behavior and gradually reducing the minimum quota for users who have accumulated a certain amount of behavioral data. Furthermore, the minimum quota can be dynamically adjusted based on real-time feedback data; when the actual click-through rate or interaction rate of a particular recalled content is high, its minimum quota can be appropriately increased. This method enables a more refined and scenario-adaptive quota protection strategy.

[0218] In this embodiment, before refining the coarse-ranked recommended content using the fine-ranking model, preference-recalled content and topic-recalled content are filtered based on a preset regional minimum quota to obtain the filtered coarse-ranked recommended content. The filtered coarse-ranked recommended content includes content that meets the regional minimum quota, achieving targeted protection and priority guarantee for content on specific recall paths. By setting the regional minimum quota, the problem of preference-recalled content and topic-recalled content being eliminated prematurely in the new user cold start scenario due to sparse behavioral data and inaccurate scoring by the ranking model is effectively solved. This ensures that these two recall strategies specifically designed for new users can play a role in the ranking process, improving the implementation effect and stability of the new user cold start recommendation strategy. This allows the value of the preference-recalled and topic-recalled recall strategies to be fully reflected, thereby effectively improving the recommendation experience for new users and the recommendation system's adaptability to new users.

[0219] In one embodiment of this specification, a process flow description of a content recommendation method is provided. Specifically, see [link to documentation]. Figure 2 , Figure 2 This specification illustrates a flowchart of the processing procedure for a content recommendation method according to one embodiment. Figure 2 As shown.

[0220] First, obtain the target users' target topic behaviors, geographical preference tags, and geographical topic tags.

[0221] Meanwhile, preference recall models and topic recall models are pre-trained based on historical user preference tags and historical topic behaviors.

[0222] Furthermore, a feature database is pre-built based on the behavioral characteristics of historical topic behaviors, geographical preference tags, and user preference tags, and a feature query interface is configured.

[0223] The recall process is conducted in parallel, specifically including: Using a preference recall model, preference recall content is retrieved from the content database based on the user preference tags of the target user; and the preference recall content is filtered based on geographical preference tags to obtain the filtered preference recall content.

[0224] Using a topic recall model, topic recall content is retrieved from the content database based on the behavioral characteristics of target topic behavior; and regional popular topic content is retrieved from the content database based on geographical topic tags; the intersection of topic recall content and regional popular topic content is taken to obtain the filtered topic recall content.

[0225] Furthermore, using a coarse ranking model, the preferred recall content and topic recall content are coarsely ranked based on preset coarse ranking indicators; and based on preset regional minimum quotas, the preferred recall content and the topic recall content are filtered to obtain the filtered coarsely ranked recommended content.

[0226] Finally, using the fine-ranking model, the coarse-ranked recommended content is fine-ranked based on preset fine-ranking indicators to obtain recommended content for the target user.

[0227] In the embodiments of this specification, by acquiring the target user's target topic behavior, geographic preference tags, and geographic topic tags, and by pre-constructing a feature database based on historical user data and pre-training a recall model, the target user's topic behavior, geographic preferences, and geographic topic tags are combined to execute two parallel strategies: preference recall and topic recall. After geographic filtering and regional popular topic intersection processing, and then through regional minimum quota filtering in coarse ranking and fine ranking for refined sorting, the recommended content is finally generated. This realizes the end-to-end processing from offline data support to online collaborative recall, regional adaptation, and minimum ranking, so that even users with sparse behavior can obtain personalized recommendations that take into account both long-term stable interests and short-term dynamic intentions. This effectively improves the accuracy, timeliness, regional fit, and ranking stability of the recommendation results, and enhances the recommendation quality and user experience in the cold start scenario for new users.

[0228] In one embodiment of this specification, a flow sequence description of a content recommendation method is provided. Specifically, see [link to documentation]. Figure 3 , Figure 3 This is a flowchart of a content recommendation method provided in one embodiment of this specification, as shown below. Figure 3 As shown.

[0229] The client responds to a content recommendation request for a target user by sending the target user's target topic behavior to the server.

[0230] The server queries the feature database for user preference tags of the target user based on the behavioral characteristics of the target topic.

[0231] The server sends the search results and interactive content from the first interactive area to the client.

[0232] The feature database returns user preference tags for the target user.

[0233] The server retrieves content based on the user preference tags of the target user from the content database.

[0234] The server retrieves topic-related content from the content database based on the behavioral characteristics of the target topic.

[0235] The content database returns content retrieved based on preferences and topics.

[0236] The server sorts the content retrieved based on preferences and topics.

[0237] The server returns recommended content tailored to the target user to the client.

[0238] In this embodiment, the client responds to a content recommendation request for a target user by sending the target user's target topic behavior to the server. The server then queries the user's preference tags from the feature database based on these behavioral characteristics. Based on the preference tags and topic behavior characteristics, it simultaneously retrieves preference-recalled content and topic-recalled content from the content database. The two retrieved contents are then sorted and returned to the recommendation list. This achieves a complete temporal interaction from user behavior collection, feature query, parallel retrieval, to sorting output. Even if the target user's long-term behavior is sparse, reliable preference and interest tags can be obtained from the feature database built based on historical data. The parallel retrieval mechanism of preferences and topics takes into account both the user's long-term stable interests and short-term dynamic intentions. Finally, personalized recommendation content is generated through sorting, improving the accuracy, timeliness, and overall quality of the recommendation results. This enhances the adaptability and interaction efficiency of the content recommendation method for new user cold start scenarios.

[0239] Corresponding to the above method embodiments, this specification also provides embodiments of a content recommendation system, specifically, see [link to relevant documentation]. Figure 4 , Figure 4 A schematic diagram of the structure of a content recommendation system provided in one embodiment of this specification is shown. Figure 4 As shown, the content recommendation system includes: The acquisition unit 402 is used to acquire the target user's target topic behavior, wherein the target topic behavior includes at least one of click behavior, browsing behavior and interaction behavior; The query unit 404 is used to query the user preference tags of the target user from the feature database based on the behavioral characteristics of the target topic behavior. The feature database is constructed based on the behavioral characteristics and user preference tags of the historical topic behavior of historical users. Preference recall unit 406 is used to recall preference recall content from the content database based on the user preference tags of the target user; The topic recall unit 408 is used to recall topic recall content from the content database based on the behavioral characteristics of the target topic behavior. The sorting unit 410 is used to sort the preference-recalled content and topic-recalled content to determine the recommended content for the target user.

[0240] An acquisition unit is a system functional module or program component in a content recommendation system used to acquire the target user's behavior related to a specific topic. Specifically, an acquisition unit may include a data collection interface, a behavior log parser, a real-time data receiver, or an application programming interface (API). The acquisition unit can interact with front-end pages, mobile clients, or server log systems to capture records of target users' clicks, browsing, and interactions related to topic content. The acquisition unit provides an entry point for user behavior data for the entire content recommendation system, serving as the raw data source for subsequent interest modeling and content recommendation.

[0241] A query unit is a system functional module or program component in a content recommendation system used to query user preference tags of target users from a feature database based on behavioral characteristics of target topic behaviors. Specifically, a query unit may include a feature query interface caller, a similarity calculation engine, a feature mapping model inferencer, or a conditional filtering processor. The query unit can receive behavioral characteristics of the target topic behaviors transmitted by the acquisition unit, and through accessing a pre-encapsulated standardized query interface of the feature database, perform similarity matching, model inference, or multi-condition joint retrieval in the feature database to obtain the user preference tags of the target users.

[0242] A preference recall unit is a system functional module or program component in a content recommendation system used to recall preference-based content from a content database based on the user's preference tags. Specifically, a preference recall unit may include a tag matching engine, a preference recall model inference engine, a collaborative filtering calculator, or a multi-dimensional feature combination retrieval engine. The preference recall unit can receive user preference tags output by the query unit and retrieve a set of content from the content database that matches the user's long-term stable interests through various methods such as direct tag matching, preference recall model inference, or collaborative filtering algorithms.

[0243] A topic recall unit is a system functional module or program component in a content recommendation system used to recall topic-related content from a content database based on the behavioral characteristics of target topic behaviors. Specifically, a topic recall unit may include a topic similarity model inferencer, a behavioral feature direct matching engine, a regional popular topic coprocessor, or a real-time behavioral sequence analyzer. The topic recall unit can receive behavioral characteristics of the target topic behaviors transmitted by the acquisition unit, and retrieve a set of content related to the user's short-term dynamic interests from the content database through various methods such as topic similarity expansion, feature vector matching, or regional popular topic collaboration.

[0244] In a content recommendation system, a ranking unit is a system functional module or program component used to rank preference-recalled content and topic-recalled content to determine recommended content for target users. Specifically, the ranking unit may include a coarse-ranking model inferencer, a fine-ranking model inferencer, a multi-objective optimization ranking engine, a regional quota protection processor, or a multi-source content fusion processor. The ranking unit receives the recalled content output from the preference recall unit and the topic recall unit, and through a two-stage ranking process of coarse and fine ranking, regional quota protection, and multi-source content fusion, comprehensively evaluates and prioritizes candidate content to generate the final list of recommended content for the target user.

[0245] The content recommendation system provided in this specification's embodiments acquires target user topic behaviors through an acquisition unit, queries user preference tags from a feature database based on the behavioral characteristics of the target topic behaviors through a query unit, recalls preference-recalled content based on user preference tags through a preference recall unit, recalls topic-recalled content based on behavioral characteristics through a topic recall unit, and sorts the preference-recalled content and topic-recalled content to determine the recommended content through a sorting unit. This achieves a fully systematic architecture from user behavior collection, interest tag query, two-way parallel recall to final sorting and recommendation. Specifically, the various functional units collaborate to form a complete and modular content recommendation system architecture. Even when the target user's long-term preference behaviors are sparse, reliable interest representations can be obtained. The two-way parallel recall mechanism takes into account both the user's long-term stable preferences and short-term dynamic topic intentions. Finally, the sorting process outputs recommended content that combines user personalization and content timeliness, effectively improving the accuracy, diversity, and overall quality of the recommendation results. This enhances the content recommendation system's adaptability to new users and provides a highly efficient, stable, and scalable cold-start recommendation system for content platforms.

[0246] The above is an illustrative scheme of a content recommendation system according to this embodiment. It should be noted that the technical solution of this content recommendation system and the technical solution of the content recommendation method described above belong to the same concept. Details not described in detail in the technical solution of the content recommendation system can be found in the description of the technical solution of the content recommendation method described above.

[0247] Figure 5 A structural block diagram of a computing device 500 according to one embodiment of this specification is shown. The components of the computing device 500 include, but are not limited to, a memory 510 and a processor 520. The processor 520 is connected to the memory 510 via a bus 530, and a database 550 is used to store data.

[0248] The computing device 500 also includes an access device 540, which enables the computing device 500 to communicate via one or more networks 560. Examples of these networks include a Public Switched Telephone Network (PSTN), a Local Area Network (LAN), a Wide Area Network (WAN), a Personal Area Network (PAN), or a combination of communication networks such as the Internet. The access device 540 may include one or more of any type of wired or wireless network interface (e.g., a network interface controller (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0249] In one embodiment of this specification, the above-described components of the computing device 500 and Figure 5 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 5 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0250] The computing device 500 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 500 can also be a mobile or stationary server.

[0251] The processor 520 is used to execute the following computer program / instructions, which, when executed by the processor, implement the steps of the recommended method described above.

[0252] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the content recommendation method described above belong to the same concept. For details not described in detail in the technical solution of the computing device, please refer to the description of the technical solution of the content recommendation method described above.

[0253] An embodiment of this specification also provides a computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of the method recommended above.

[0254] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium and the technical solution of the content recommendation method described above belong to the same concept. Details not described in detail in the technical solution of the storage medium can be found in the description of the technical solution of the content recommendation method described above.

[0255] An embodiment of this specification also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the method recommended above.

[0256] The above is an illustrative example of a computer program according to this embodiment. It should be noted that the technical solution of this computer program and the technical solution of the content recommendation method described above belong to the same concept. Details not described in detail in the technical solution of the computer program can be found in the description of the technical solution of the content recommendation method described above.

[0257] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0258] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0259] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0260] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0261] The preferred embodiments disclosed above are merely illustrative of this specification. Optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A content recommendation method, characterized in that, include: Acquire target user's target topic behavior, wherein the target topic behavior includes at least one of click behavior, browsing behavior and interaction behavior; Based on the behavioral characteristics of the target topic behavior, the user preference tags of the target user are queried from the feature database, wherein the feature database is constructed based on the behavioral characteristics and user preference tags of historical users' historical topic behavior; Based on the user preference tags of the target user, retrieve preference-retrieved content from the content database; Based on the behavioral characteristics of the target topic, retrieve the topic-related content from the content database; The preference-recalled content and the topic-recalled content are sorted to determine the recommended content for the target user.

2. The method according to claim 1, characterized in that, Before querying the target user's user preference tags from the feature database based on the behavioral characteristics of the target topic behavior, the method further includes: Obtain the historical topic behavior and user preference tags of historical users, wherein the historical topic behavior includes at least one of click behavior, browsing behavior and interaction behavior; The behavioral characteristics of the historical topic behaviors and the user preference tags are subjected to feature standardization processing, encapsulated to obtain a feature database, and a feature query interface for the feature database is generated. The step of querying the target user's user preference tags from the feature database based on the behavioral characteristics of the target topic behavior includes: Based on the behavioral characteristics of the target topic behavior, the feature query interface is accessed to query the user preference tags of the target user from the feature database.

3. The method according to claim 2, characterized in that, Before performing feature standardization processing on the behavioral characteristics of the historical topic behavior and the user preference tags, and encapsulating them to obtain a feature database, the method further includes: Obtain historical users' geographic preference tags; The process of standardizing the behavioral characteristics of the historical topic behaviors and the user preference tags, encapsulating them to obtain a feature database, and generating a feature query interface for the feature database includes: The behavioral characteristics of the historical topic behavior, the geographical preference tags, and the user preference tags are subjected to feature standardization processing, encapsulated to obtain a feature database, and a feature query interface for the feature database is generated. Before accessing the feature query interface based on the behavioral characteristics of the target topic behavior to query the user preference tags of the target user from the feature database, the method further includes: Obtain the geographic preference tags of the target user; The step of accessing the feature query interface based on the behavioral characteristics of the target topic behavior and querying the user preference tags of the target user from the feature database includes: Based on the behavioral characteristics of the target topic behavior and the geographical preference tags, the feature query interface is accessed to query the user preference tags of the target user from the feature database.

4. The method according to claim 1, characterized in that, Before retrieving content from the content database based on the target user's user preference tags, the method further includes: Obtain user preference tags and historical preference content from historical users; Based on the user preference tags and historical preference content of the historical users, a preference recall model is trained; The step of retrieving preference-based content from the content database based on the target user's user preference tags includes: Using the aforementioned preference recall model, preference-recall content is retrieved from the content database based on the target user's user preference tags.

5. The method according to claim 4, characterized in that, The step of training a preference recall model based on the user preference tags and historical preference content of the historical users includes: Based on the user preference tags and historical preference content of the historical users, a preference recall model is trained according to a preset training strategy. The preset training strategy includes optimizing the model parameters of the preference recall model based on the association between the user preference tags and historical preference content of the historical users, and adjusting the strategy parameters of the preference recall model through offline evaluation.

6. The method according to claim 1, characterized in that, Before retrieving topic-related content from the content database based on the behavioral characteristics of the target topic behavior, the method further includes: Obtain the historical topic behaviors of historical users and the historical topic content corresponding to the historical topic behaviors; A topic recall model is trained based on the behavioral characteristics of the historical users' historical topic behaviors and the content of the historical topics. The step of retrieving topic-related content from the content database based on the behavioral characteristics of the target topic includes: Using the topic recall model, topic-related content is recalled from the content database based on the behavioral characteristics of the target topic behavior.

7. The method according to claim 6, characterized in that, The process of training a topic retrieval model based on the behavioral characteristics of the historical users' historical topic behaviors and the content of the historical topics includes: Based on the historical topic content corresponding to the historical topic behavior of the historical users, a topic similarity model is trained according to a preset training strategy. The preset training strategy includes optimizing the model parameters of the topic recall model based on the correlation between the behavioral features of the historical topic behavior and the historical topic content, and adjusting the strategy parameters of the topic recall model through offline evaluation.

8. The method according to claim 1, characterized in that, Before sorting the preference-recalled content and the topic-recalled content to determine the recommended content for the target user, the method further includes: Obtain the geographic preference tags of the target user; Based on the geographic preference tags, the preference recall content is filtered to obtain the filtered preference recall content.

9. The method according to claim 1, characterized in that, Before sorting the preference-recalled content and the topic-recalled content to determine the recommended content for the target user, the method further includes: Obtain the geographic topic tags of the target user; Based on the geographical topic tags, retrieve popular regional topic content from the content database; The topic recall content and the regional popular topic content are intersected to obtain the filtered topic recall content.

10. The method according to any one of claims 1-9, characterized in that, The process of sorting the preference-recalled content and the topic-recalled content to determine the recommended content for the target user includes: Using a coarse ranking model, the preferred recall content and the topic recall content are coarsely ranked based on preset coarse ranking indicators to obtain coarsely ranked recommended content. The preset coarse ranking indicators include at least one of estimated click-through rate, estimated browsing time and estimated interaction rate. Using a fine-ranking model, the coarse-ranked recommended content is fine-ranked based on preset fine-ranking indicators to obtain recommended content for the target user.

11. The method according to claim 10, characterized in that, Before using the fine-ranking model to fine-rank the coarse-ranked recommended content based on preset fine-ranking metrics to obtain recommended content for the target user, the method further includes: Based on the preset regional minimum quota, the preferred recall content and the topic recall content are filtered to obtain the filtered coarse-ranked recommended content, wherein the filtered coarse-ranked recommended content includes content that meets the regional minimum quota.

12. A content recommendation system, characterized in that, include: The acquisition unit is used to acquire the target user's target topic behavior, wherein the target topic behavior includes at least one of click behavior, browsing behavior and interaction behavior; The query unit is used to query the user preference tags of the target user from the feature database based on the behavioral characteristics of the target topic behavior, wherein the feature database is constructed based on the behavioral characteristics and user preference tags of historical users' historical topic behavior; The preference recall unit is used to recall preference recall content from the content database based on the user preference tags of the target user; The topic recall unit is used to recall topic recall content from the content database based on the behavioral characteristics of the target topic behavior. The sorting unit is used to sort the preference recall content and the topic recall content to determine the recommended content for the target user.

13. A computing device, characterized in that, include: Memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of the content recommendation method according to any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that, It stores a computer program / instruction that, when executed by a processor, implements the steps of the content recommendation method according to any one of claims 1 to 11.

15. A computer program product, characterized in that, Includes a computer program / instructions that, when executed by a processor, implement the steps of the content recommendation method according to any one of claims 1 to 11.

Images