A data asset recommendation method and device based on multi-path recall and comprehensive scoring

Abstract

The application provides a data asset recommendation method and device based on multi-path recall and comprehensive scoring, wherein the method comprises the following steps: a first intelligent agent sends a first vector to a second intelligent agent; the second intelligent agent dynamically schedules a plurality of parallel recall channels according to the first vector to construct a candidate asset pool; a third intelligent agent scores each candidate asset in the candidate asset pool, and obtains a comprehensive score of each candidate asset by weighting and fusing the scores of each item; the first intelligent agent sorts the candidate assets in descending order according to the comprehensive score, selects a recommendation result list for output, and generates a recommendation reason according to the recall source and the item score of each candidate asset. The application realizes a dynamic recommendation mechanism based on multi-path parallel recall, six-dimensional weighted comprehensive scoring, automatic generation of recommendation reasons, and sandbox pre-performance gray deployment of scenario perception and user state information, and significantly improves the accuracy, diversity, explainability and system iteration safety of data asset recommendation.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and in particular to a data asset recommendation method and apparatus based on multi-path recall and comprehensive scoring. Background Technology

[0002] As enterprise data assets continue to expand, the efficiency of data asset discovery and the accuracy of recommendations have become key factors restricting the release of the value of data platforms. Multi-path recall and comprehensive scoring, as two core recommendation technologies, provide solutions from the dimensions of candidate coverage breadth and scoring ranking accuracy, respectively.

[0003] However, existing technologies suffer from the following drawbacks: First, they rely on a single recall channel, leading to severe homogenization of recommendation results and an inability to balance personalization and diversity. Second, the scoring mechanism is simplistic, typically relying only on access popularity or simple collaborative filtering, without integrating multi-dimensional features such as collections, browsing, departmental consistency, and freshness, making it difficult to achieve accurate comprehensive quality assessment. Third, the recommendation results lack interpretability, leaving users unable to understand the basis for recommendations, thus reducing trust and willingness to adopt them. Fourth, the strategy iteration carries high risk, with model updates being deployed directly without sandbox testing or gray-scale rollback mechanisms, easily leading to a regression in production metrics.

[0004] Therefore, there is an urgent need for a data asset recommendation method based on multi-path recall and comprehensive scoring to achieve multi-dimensional dynamic perception, explainable recommendation, and security strategy coordination. Summary of the Invention

[0005] This invention provides a data asset recommendation method based on multi-path recall and comprehensive scoring. This method can overcome the problems of single recall channels, one-sided scoring dimensions, lack of interpretability of recommendation results, and high risk of strategy iteration in existing technologies. It realizes a dynamic recommendation mechanism based on user status information and scene awareness, multi-path parallel recall, six-dimensional weighted comprehensive scoring, automatic generation of recommendation reasons, and sandbox pre-deployment, which significantly improves the accuracy, diversity, interpretability, and system iteration security of data asset recommendation.

[0006] Firstly, a data asset recommendation method based on multi-path recall and comprehensive scoring is provided, including: The first intelligent agent collects user status information and scene information, converts the user status information and scene information into a first vector, and sends it to the second intelligent agent. The second intelligent agent dynamically schedules multiple parallel recall channels based on the first vector, constructs a candidate asset pool, and sends the candidate asset pool to the third intelligent agent; the multiple recall channels include at least a tag similarity recall channel, a collection preference recall channel, a recently viewed recall channel, a popular recall channel within the same department, and a global popular recall channel; The third intelligent agent scores each candidate asset in the candidate asset pool based on the dimensions of collection hit rate, browsing hit rate, tag similarity, department consistency, popularity, and freshness. The scores of each sub-item are weighted and fused to obtain the comprehensive score of each candidate asset. The first intelligent agent sorts the candidate assets in descending order according to the comprehensive score, selects the top-ranked assets as a recommended result list, and generates recommendation reasons based on the recall source of each candidate asset and the sub-score.

[0007] In this way, the first intelligent agent collects user status and scenario information, while the second intelligent agent dynamically schedules multiple parallel recall channels to construct a diverse candidate asset pool, effectively overcoming the bias and cold start problems of a single recall strategy. The third intelligent agent performs comprehensive scoring and weighted fusion from six dimensions: collection hit rate, browsing hit rate, tag similarity, departmental consistency, popularity, and freshness, achieving multi-dimensional and accurate evaluation. The first intelligent agent outputs recommendation results based on the comprehensive score ranking and generates explainable recommendation reasons. This method significantly improves the accuracy, personalization, and diversity of data asset recommendations, while enhancing the explainability and user trust of the recommendation results, making it suitable for intelligent discovery scenarios in enterprise data asset management.

[0008] In some embodiments, the first agent collects user state information and scene information, converts the user state information and scene information into a first vector, and sends it to the second agent, including: The first intelligent agent acquires user identifier, user historical behavior data, user department information, and context information of the current recommendation scenario; wherein, the user historical behavior data includes the user's historical collection asset list, recently viewed asset list, and explicit feedback records of the user on historical recommendation results, and the context information includes the currently accessed page type, the currently viewed asset identifier, and the user's current task intent; The first intelligent agent performs statistical analysis on the user's historical behavior data and extracts behavior density features, which include at least the number of collected assets, recent browsing frequency, and active feedback activity. The first intelligent agent performs scene classification on the context information and determines the scene type label; The first intelligent agent concatenates and encodes the behavior density features and the scene type label to form the first vector.

[0009] In this way, the first intelligent agent collects multi-dimensional user states (historical behavior, department, context) and extracts behavioral density features. These features are then combined with scene classification to generate a unified vector, providing accurate personalized input for subsequent recall and scoring. This method effectively captures users' real-time intent and long-term preferences, improving the dynamic adaptability and accuracy of recommendations.

[0010] In some embodiments, the second agent dynamically schedules multiple parallel recall channels based on the first vector to construct a candidate asset pool, including: The second agent parses the first vector and extracts user behavior density features and scene type labels; Based on the scenario type tags and the user behavior density features, a subset of channels activated in this round is determined from the recall channel pool, and a recall quantity quota and initial weight coefficient are allocated to each activated channel; the recall channel pool includes tag similarity recall channels, collection preference recall channels, recently viewed recall channels, popular recall channels in the same department, and global popular recall channels; The activation channel is used to recall the data asset library in parallel, and the candidate asset list returned by each channel is received. Each candidate asset carries the original sorting position and matching score of its channel. The received candidate assets are deduplicated and merged using the asset identifier as the primary key. When the same asset is hit by multiple channels, the hit source markers are aggregated and the scores of each channel are merged to form an initial confidence level. The candidate assets after deduplication and merging are filtered to remove at least the assets that users have marked as uninteresting, and assets that meet the exposure fatigue criteria are down-weighted or removed. Output a structured candidate asset pool, in which each candidate asset includes an asset identifier, asset name, recall source list, initial confidence level, and source evidence chain.

[0011] In this way, by using a second intelligent agent to analyze the behavior density and scene labels in the first vector, multiple recall channels are dynamically scheduled and allocated quotas and weights to achieve scene-adaptive parallel recall. The results from multiple channels are deduplicated and merged, cross-channel scores are fused, and initial confidence is calculated, effectively avoiding duplicate recommendations and utilizing multi-source signals to improve candidate quality. Further filtering removes uninteresting assets and reduces the weight of exposure fatigue items, significantly improving user experience. The structured candidate asset pool constructed by this method retains the recall source and evidence chain, providing rich input for subsequent comprehensive scoring and explainable recommendations, thus improving the accuracy, diversity, and user satisfaction of recommendations.

[0012] In some embodiments, the third agent scores each candidate asset in the candidate asset pool based on four dimensions: collection hit rate, browsing hit rate, tag similarity, department consistency, popularity, and freshness. The scores from each dimension are then weighted and fused to obtain a comprehensive score for each candidate asset, including: The third intelligent agent calculates a collection hit score based on whether the candidate asset appears in the user's collection list; and calculates a browsing hit score based on whether the candidate asset is associated with the user's recently viewed assets. The third agent converts the labels, names, and descriptions of the candidate assets into semantic vectors, calculates the cosine similarity with the semantic vectors of the current context assets or the user interest profile vectors, and obtains the label similarity score. The third intelligent agent determines whether the department to which the candidate asset belongs is consistent with the department to which the user belongs, and obtains a department consistency score; The third intelligent agent normalizes the access frequency of the candidate assets across the entire platform or department to obtain a heat normalization score. The third intelligent agent calculates the freshness decay score using a time decay function based on the time difference between the release time of the candidate asset and the current time. The third intelligent agent calculates the comprehensive score of each candidate asset based on preset weight coefficients.

[0013] In this way, a third-party intelligent agent performs refined sub-scoring of candidate assets from six dimensions: collection hit rate, browsing hit rate, tag similarity, departmental consistency, popularity, and freshness. A weighted fusion is then used to obtain a comprehensive score. This method integrates user personalized behavior, social attributes, and global asset features, effectively overcoming the one-sidedness of single-dimensional scoring. Simultaneously, semantic vectors are used to calculate tag similarity, and a time decay function is employed to handle freshness, improving the accuracy and timeliness of the scoring. The weighted fusion mechanism can flexibly adjust weights according to the scenario, achieving dynamic optimization of the recommendation goal. This solution significantly improves the accuracy, personalization level, and user satisfaction of data asset recommendations, making it suitable for intelligent discovery scenarios of enterprise data assets.

[0014] In some embodiments, the first agent sorts the candidate assets in descending order according to the comprehensive score, and selects a preset number of top-ranked assets as the recommended result list for output, including: The first intelligent agent receives the comprehensive score of each candidate asset returned by the third intelligent agent; The first intelligent agent sorts the candidate assets from high to low according to the comprehensive score, and extracts the top N candidate assets to form a recommendation result list, where N is the preset number of recommendation displays; For each candidate asset in the recommendation result list, the first agent selects or splices a structured recommendation reason text from a preset recommendation reason template library based on its recall source channel and the top K sub-dimensions with the highest scores. The first intelligent agent generates a scoring radar chart for each candidate asset. The radar chart uses collection hit, browsing hit, tag similarity, department consistency, popularity, and freshness as axes, and the value of each axis is the normalized value of the corresponding sub-item score. The first intelligent agent outputs the list of recommended results, the text of the reasons for the recommendations, and the rating radar chart to the user interface for display.

[0015] In this way, the first intelligent agent sorts candidate assets in descending order of their comprehensive scores and extracts the top results, ensuring that the recommended content focuses on assets that users are most likely to be interested in. Based on this, structured recommendation reasons are generated from a template library according to the recall source and high-scoring dimensions, significantly improving the interpretability and user trust of the recommendation results. Simultaneously, a radar chart containing six-dimensional sub-scores is generated for each candidate asset, visually displaying the asset scoring structure to help users understand the recommendation basis and assist in decision-making. This method achieves accurate output, enhanced interpretability, and visual presentation of recommendation results, effectively improving the user interaction experience and the transparency of the recommendation system.

[0016] In some embodiments, the method further includes: In a non-production sandbox environment, the pre-rehearsal sandbox agent periodically obtains snapshots of the current policy model from the second agent and the third agent. By injecting preset perturbation patterns into the sandbox environment, the recommended performance under extreme scenarios is simulated, and the stability index of the strategy is evaluated. Once the evaluation is passed, the updated policy configuration will be pushed to the second and third agents in the production environment in a gray-scale manner, while retaining the previous version; If a production metric is detected to have regressed, the system will automatically roll back to the previous version.

[0017] In this way, by periodically capturing policy snapshots in a non-production environment through a pre-launch sandbox agent and injecting perturbation patterns to simulate extreme scenarios, policy stability can be assessed in advance without affecting online services, significantly reducing deployment risks. Employing a canary rollout and version retention mechanism, automatic rollback is supported when production metrics are rolled back, achieving safe iteration and rapid recovery capabilities for the recommendation system. This method improves the reliability, fault tolerance, and continuous delivery efficiency of recommendation strategies.

[0018] In some embodiments, when the method is deployed across multiple organizational nodes, the third agents within each organizational node collaborate through a federated learning framework, including: The third agent at each organizational node calculates the gradient or update parameters of the scoring model on local data, and uploads them to the federated aggregation server after encryption. The federated aggregation server securely aggregates the parameters uploaded by each organization node and generates a global parameter update. The federated aggregation server distributes the global parameter updates back to the third agents of each organization node, updating the local scoring model and knowledge graph embedding representation of each node.

[0019] This approach enables collaborative scoring models across multiple organizational nodes using a federated learning framework. Each node calculates gradients or parameter updates only on its local data, uploads them securely to an aggregation server after encryption, and then distributes them back to each node to update its local model. This method eliminates the need to share raw data, effectively protecting the data privacy and security of each organization, while leveraging cross-organizational data distribution to improve the generalization ability and accuracy of the scoring model. The introduction of joint updates using knowledge graph embedding representations further enhances the capture of asset association semantics. This solution addresses the data silo problem, achieves privacy-preserving collaborative modeling, and significantly improves the overall performance and scalability of the recommendation system.

[0020] Secondly, a data asset recommendation device based on multi-channel recall and comprehensive scoring is provided, including: The first processing module is used for the first intelligent agent to collect user status information and scene information, and to convert the user status information and scene information into a first vector and send it to the second intelligent agent. The second processing module is used by the second intelligent agent to dynamically schedule multiple parallel recall channels based on the first vector, construct a candidate asset pool, and send the candidate asset pool to the third intelligent agent; the multiple recall channels include at least a tag similarity recall channel, a collection preference recall channel, a recently viewed recall channel, a popular recall channel within the same department, and a global popular recall channel. The third processing module is used by the third intelligent agent to score each candidate asset in the candidate asset pool from the dimensions of collection hit, browsing hit, tag similarity, department consistency, popularity and freshness, and to weight and fuse the scores of each item to obtain the comprehensive score of each candidate asset. The fourth processing module is used by the first intelligent agent to sort the candidate assets in descending order according to the comprehensive score, select a preset number of assets with the highest ranking as the recommended result list, and generate recommendation reasons based on the recall source of each candidate asset and the sub-score.

[0021] In some possible implementations, the first processing module is used for the first intelligent agent to acquire user identifier, user historical behavior data, user department information, and context information of the current recommendation scenario; wherein, the user historical behavior data includes the user's historical collection asset list, recently viewed asset list, and explicit feedback records of the user on historical recommendation results, and the context information includes the currently accessed page type, the currently viewed asset identifier, and the user's current task intent; The first intelligent agent performs statistical analysis on the user's historical behavior data and extracts behavior density features, which include at least the number of collected assets, recent browsing frequency, and active feedback activity. The first intelligent agent performs scene classification on the context information and determines the scene type label; The first intelligent agent concatenates and encodes the behavior density features and the scene type label to form the first vector.

[0022] In some possible implementations, a second processing module is used by the second agent to parse the first vector and extract user behavior density features and scene type labels; Based on the scenario type tags and the user behavior density features, a subset of channels activated in this round is determined from the recall channel pool, and a recall quantity quota and initial weight coefficient are allocated to each activated channel; the recall channel pool includes tag similarity recall channels, collection preference recall channels, recently viewed recall channels, popular recall channels in the same department, and global popular recall channels; The activation channel is used to recall the data asset library in parallel, and the candidate asset list returned by each channel is received. Each candidate asset carries the original sorting position and matching score of its channel. The received candidate assets are deduplicated and merged using the asset identifier as the primary key. When the same asset is hit by multiple channels, the hit source markers are aggregated and the scores of each channel are merged to form an initial confidence level. The candidate assets after deduplication and merging are filtered to remove at least the assets that users have marked as uninteresting, and assets that meet the exposure fatigue criteria are down-weighted or removed. Output a structured candidate asset pool, in which each candidate asset includes an asset identifier, asset name, recall source list, initial confidence level, and source evidence chain.

[0023] In some possible implementations, a third processing module is used by the third intelligent agent to calculate a collection hit score based on whether the candidate asset appears in the user's collection list; and to calculate a browsing hit score based on whether the candidate asset is associated with the user's recently viewed assets. The third agent converts the labels, names, and descriptions of the candidate assets into semantic vectors, calculates the cosine similarity with the semantic vectors of the current context assets or the user interest profile vectors, and obtains the label similarity score. The third intelligent agent determines whether the department to which the candidate asset belongs is consistent with the department to which the user belongs, and obtains a department consistency score; The third intelligent agent normalizes the access frequency of the candidate assets across the entire platform or department to obtain a heat normalization score. The third intelligent agent calculates the freshness decay score using a time decay function based on the time difference between the release time of the candidate asset and the current time. The third intelligent agent calculates the comprehensive score of each candidate asset based on preset weight coefficients.

[0024] In some possible implementations, a fourth processing module is used for the first agent to receive the comprehensive score of each candidate asset returned by the third agent; The first intelligent agent sorts the candidate assets from high to low according to the comprehensive score, and extracts the top N candidate assets to form a recommendation result list, where N is the preset number of recommendation displays; For each candidate asset in the recommendation result list, the first agent selects or splices a structured recommendation reason text from a preset recommendation reason template library based on its recall source channel and the top K sub-dimensions with the highest scores. The first intelligent agent generates a scoring radar chart for each candidate asset. The radar chart uses collection hit, browsing hit, tag similarity, department consistency, popularity, and freshness as axes, and the value of each axis is the normalized value of the corresponding sub-item score. The first intelligent agent outputs the list of recommended results, the text of the reasons for the recommendations, and the rating radar chart to the user interface for display.

[0025] In some possible implementations, the fourth processing module is used to pre-simulate the sandbox agent in a non-production sandbox environment and periodically obtain current policy model snapshots from the second agent and the third agent. By injecting preset perturbation patterns into the sandbox environment, the recommended performance under extreme scenarios is simulated, and the stability index of the strategy is evaluated. Once the evaluation is passed, the updated policy configuration will be pushed to the second and third agents in the production environment in a gray-scale manner, while retaining the previous version; If a production metric is detected to have regressed, the system will automatically roll back to the previous version.

[0026] In some possible implementations, a third processing module is used by the third agent of each organizational node to calculate the gradient or update parameters of the scoring model on local data, and then uploads them to the federated aggregation server after encryption. The federated aggregation server securely aggregates the parameters uploaded by each organization node and generates a global parameter update. The federated aggregation server distributes the global parameter updates back to the third agents of each organization node, updating the local scoring model and knowledge graph embedding representation of each node.

[0027] Thirdly, an electronic device is provided, comprising: one or more processors; one or more memories; and one or more programs, wherein the one or more programs are stored in the one or more memories, and the one or more programs include instructions that, when executed by the one or more processors, cause the configured device to perform the method as described in the first aspect.

[0028] Fourthly, a computer storage medium is provided, the computer storage medium storing instructions that, when executed by a computer, cause the computer to perform the method described in the first aspect or the second aspect.

[0029] Fifthly, a computer program product is provided, the computer program product storing instructions that, when executed by a computer, cause the computer to perform the method described in the first aspect or the second aspect.

[0030] The beneficial effects of the second to fifth aspects can be referred to the introduction of the beneficial effects of the first aspect above, and will not be repeated here. Attached Figure Description

[0031] Figure 1 This is the overall architecture for applying the data asset recommendation method based on multi-path recall and comprehensive scoring in this application; Figure 2 A flowchart illustrating the data asset recommendation method based on multi-path recall and comprehensive scoring provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the multi-path recall candidate construction process provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the comprehensive score calculation and ranking process provided in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the process of generating recommendation reasons for embodiments of this application; Figure 6 This is a schematic diagram of the recommendation reason generation process provided in an embodiment of the present invention; Figure 7 A schematic diagram of a data asset recommendation device based on multi-path recall and comprehensive scoring that can be used to implement the method of the present invention; Figure 8 This is a schematic diagram of an electronic device provided by the present invention. Detailed Implementation

[0032] The solutions provided by the embodiments of the present invention will now be described with reference to the accompanying drawings. In the embodiments of the present invention, "multiple" refers to two or more objects, and "various kinds" refers to two or more types. Terms such as "first," "second," etc., are only used to distinguish similar objects and are not necessarily used to describe a specific order or number of objects.

[0033] With the deepening of enterprise digital transformation, the number of data assets is exploding. In data asset portals, users (such as data analysts and business personnel) need to quickly and accurately discover the data tables, metrics, reports, and other assets they need. However, existing data asset recommendation methods have many shortcomings. First, the recommendation logic is simplistic, mostly relying solely on asset popularity (such as page views) or single tag similarity, lacking personalization and often resulting in recommendations that are disconnected from the user's true intent and contextual needs. Second, information utilization is insufficient; existing solutions struggle to effectively integrate various user behavioral data such as browsing, saving, and downloading, as well as heterogeneous information from multiple sources, such as user department affiliation and asset creation time, limiting the accuracy of recommendations. Third, when facing new users or users with sparse behavioral data (i.e., cold start scenarios), a single recommendation strategy often fails to generate effective recommendations or produces low-quality results. Finally, the recommendation process is often a "black box"—the system only provides results without explanation, leaving users unable to understand why a particular asset is recommended, reducing the credibility and user acceptance of the recommendations.

[0034] Given that existing data asset recommendation technologies suffer from problems such as single recommendation signals, insufficient coverage of candidate assets, poor performance in cold start scenarios, lack of interpretability of recommendation results, and difficulty in adapting to different business scenarios and strategy requirements, how to provide a personalized data asset recommendation method that can integrate multi-source heterogeneous information, dynamically adapt to different scenarios, and provide interpretable results is a technical problem that urgently needs to be solved by those skilled in the art.

[0035] This invention introduces an intelligent agent system into a data asset recommendation method based on multi-path recall and comprehensive scoring. Through the division of labor and cooperation among functionally cohesive intelligent agents, a full-link intelligent closed loop is achieved, encompassing environmental perception, policy orchestration and recall management, multi-dimensional scoring and knowledge reasoning, user interaction and explainable recommendation, sandbox pre-testing, and federated evolution. After functional merging, the responsibilities and relationships of each intelligent agent are as follows: The first is the user interaction and explanation agent. This agent serves as the unified interaction interface between the system and the user, undertaking dual responsibilities: on the one hand, it is responsible for collecting user behavior data, contextual information, and explicit feedback, continuously maintaining dynamic user profiles, and transmitting the current state to the policy and candidate management agent; on the other hand, after the recommendation results are generated, it automatically analyzes the recall source and contribution dimensions of each candidate asset, generates structured recommendation reasons and a visual radar chart, and transparently presents the recommendation results and their explanations to the user. Simultaneously, this agent receives explicit user evaluations of the recommendation reasons and transmits these feedback signals to the scoring and knowledge agent, forming a closed loop from interaction to optimization.

[0036] Secondly, there is the strategy and candidate management agent, which undertakes the dual functions of strategy decision-making and candidate processing in the recommendation process. At the strategy decision-making level, it receives user status and scenario information from the user interaction and explanation agent, and dynamically determines the combination of recall channels to be used in this round of recommendation and the calling weight of each channel according to preset rules or reinforcement learning strategies. At the candidate processing level, it receives the asset lists returned by each recall channel, performs merging, deduplication, and filtering of exposed or user-uninterested assets, and marks the recall source and initial confidence level of each candidate asset, forming a unified candidate asset pool that is then passed to the scoring and knowledge agent.

[0037] Thirdly, there is the scoring and knowledge intelligence agent, which undertakes the dual functions of multi-dimensional scoring and knowledge graph maintenance. At the scoring level, it scores each asset in the candidate pool based on dimensions such as collection hit rate, browsing hit rate, tag similarity, department consistency, popularity, and freshness, and then weights and fuses these scores using learnable weight coefficients to obtain a comprehensive score S. Simultaneously, it uses feedback signals from user interaction and explanation agents as rewards, and continuously optimizes hyperparameters such as the weights of each dimension and the freshness decay coefficient online through reinforcement learning. At the knowledge graph level, it maintains a data asset knowledge graph, with nodes representing entities such as assets, tags, and departments, and edges representing relationships. During the recall phase, it provides the strategy and candidate management intelligence agent with the query capability of "knowledge graph adjacency recall." During the scoring phase, it calculates semantic similarity and relevance through graph relationships and continuously updates the graph structure based on new asset launches and user behavior data.

[0038] Fourthly, there is the pre-production sandbox agent. This agent is deployed in a non-production sandbox environment and periodically obtains snapshots of the current policy model from the policy and candidate management agent and the scoring and knowledge agent in the production environment. By injecting perturbations such as missing user behavior, label drift, and cold start, it simulates recommendation performance under various extreme scenarios, evaluates policy stability, and generates optimization suggestions. Policy updates that pass the sandbox validation are pushed to the production environment in a canary rollout manner, while retaining previous versions. If production metrics regress, the update is automatically rolled back.

[0039] Fifthly, there is the federated recommendation agent network. When this system is deployed across multiple organizational nodes, the scoring and knowledge agents within each organization collaborate through a federated learning framework. Each node calculates gradients or updates its model on local data and uploads the encrypted parameters to the federated aggregation server. The aggregation server securely aggregates the parameters from each node and distributes them back to each node, updating the local scoring model and knowledge graph embedding representation. This enables the co-evolution of cross-organizational recommendation capabilities without sharing the original data.

[0040] The system comprises three agent groups: a user interaction and explanation agent at the front end, which senses the user's state and provides recommendations and explanations; a policy and candidate management agent at the central level, which receives user status, schedules recall channels, processes candidate assets, and passes them to the scoring and knowledge agent; and a scoring and knowledge agent at the back end, which performs multi-dimensional scoring and ranking, and sends the scoring results and explanations back to the policy and candidate management agent, which is then presented by the user interaction and explanation agent. A pre-production sandbox agent serves as a bypass environment, periodically obtaining policy snapshots from the production agent for simulation verification, and then securely pushing the optimized configuration back to the production environment. The federated recommendation agent network is built across multiple organizations, coordinating the collaborative training of scoring and knowledge agents within each organization through a parameter server.

[0041] Figure 1 This is a flowchart illustrating a data asset recommendation method based on multi-path recall and comprehensive scoring according to the present invention. Figure 1 As shown, it includes the following steps: In step 110, user and context information are obtained.

[0042] In this embodiment of the application, the user interaction and explanation agent serves as a unified interaction interface with the user, obtaining the user identifier currently requesting recommendations, the user's historical behavior data, the user's department information, and the context information of the recommendation scenario.

[0043] The user's historical behavior data includes at least a list of assets the user has collected in the past, a list of assets recently viewed, and a record of the user's explicit feedback on historical recommendation results. The context information includes at least the type of page currently accessed, the asset information being viewed, or the user's current task intent. The agent continuously maintains a dynamic user profile and transmits user status and context information to the policy and candidate management agent.

[0044] In step 120, a candidate asset pool is constructed using a multi-path recall strategy.

[0045] Figure 3 This is a schematic diagram of the multi-path recall candidate construction process, such as... Figure 3 As shown in the embodiments of this application, the strategy and candidate management agent receives user state and scene information. As the strategy decision and candidate processing hub of the recommendation process, the agent obtains the user state vector and scene context from the user interaction and interpretation agent, and completes the construction of the candidate asset pool according to the two-stage pipeline of "strategy decision - fusion governance".

[0046] During the strategy decision-making phase, the strategy and candidate management agent first analyzes the input signals to complete scenario classification and user behavior density assessment. The scenario classifier maps the current request to a predefined scenario type, including but not limited to asset details page scenarios, portal homepage scenarios, cold start scenarios, and search results page scenarios. The behavior density assessment module then divides users into two states—"sufficient personalized signals" and "sparse behavior, i.e., cold start"—by statistically analyzing the user's historical collection base, recent browsing frequency, and active feedback activity.

[0047] Based on the above analysis results, the agent queries the internally maintained scenario-policy mapping library. This mapping library is initialized with preset expert rules and continuously optimized by a lightweight online reinforcement learning model. Specifically, the agent can use contextual bandit or multi-armed slot machine algorithms, using business metrics such as user click-through rate, collection rate, and dwell time in the final recommended list as reward functions, to iterate online strategies for channel combination schemes and recall quota allocation in various scenarios.

[0048] At the channel selection level, the agent dynamically activates the required subset of channels from the candidate recall channel pool for this round, and assigns a recall quantity cap and initial weight coefficient to each activated channel. The recall channel pool includes at least six categories: tag similarity recall, favorite preference recall, recently viewed recall, popular recall within the same department, globally popular recall, and adjacency recall based on knowledge graphs. When a cold start state is detected, the adaptive strategy automatically increases the quota and weight of popular recall within the same department and globally popular recall channels, while reducing or temporarily bypassing channels that rely on user history behavior, such as favorite preference recall and recently viewed recall, smoothly transferring the recall signal from individual behavior to collective intelligence, ensuring the usability of recommendations in cold start scenarios.

[0049] After completing the recall channel scheduling and executing the recall in parallel for each channel, the policy and candidate management agent enters the candidate asset fusion and governance phase. The candidate asset lists returned by each activated channel are aggregated into the agent, and each candidate asset carries its original ranking position, matching score, or similarity score in the corresponding channel.

[0050] The intelligent agent performs deduplication and merging operations using the asset's unique identifier as the primary key. When the same asset is hit by multiple recall channels, the system processes it according to a multi-source signal fusion strategy: retaining and aggregating all hit source tags to form a source tag set; normalizing the original scores assigned to each channel and taking the highest value or weighted fusion value as the initial confidence level of the asset; merging the recall evidence descriptions from each channel to construct a structured source evidence chain. The fact of multiple cross-hit events itself serves as a positive signal of the asset's credibility and can be used for weighted enhancement of subsequent scoring.

[0051] After deduplication and merging are completed, the agent calls upon the user feedback archive maintained by the user interaction and interpretation agent to perform filtering and governance operations. The filtering rules include: explicit negative feedback filtering, which removes assets that users have explicitly marked as uninteresting; exposure fatigue control, which downgrades or eliminates assets that have been exposed multiple times recently but for which users have not generated positive interactions; and self-owned asset filtering, which excludes assets created by the user or for which the user has management authority, to avoid ineffective recommendations.

[0052] After the above processing, the structured candidate asset pool for this round of recommendations is completed. Each candidate asset in the pool carries a standardized metadata structure, including asset identifier, name and description, recall source list and corresponding channel-level score, initial confidence after fusion, and a complete source evidence chain. The policy and candidate management agent uses this candidate asset pool as an intermediate representation and passes it to the scoring and knowledge agent for subsequent multi-dimensional comprehensive scoring.

[0053] This two-stage design integrates strategy orchestration and data governance within the same intelligent body, forming a complete closed loop of "scene perception → strategy decision-making → multi-path parallel recall → fusion deduplication → feedback filtering → candidate pool output". This provides a high-quality, diverse, and fully traceable data foundation for downstream scoring and recommendation presentation.

[0054] In step 130, the candidate assets are evaluated using a multi-dimensional comprehensive scoring system and enhanced with knowledge.

[0055] In the embodiments of this application, Figure 4 This is a schematic diagram of the comprehensive scoring calculation and ranking process, such as... Figure 4 As shown, the scoring and knowledge agent undertakes the dual functions of comprehensive scoring and knowledge graph maintenance, and performs the following for each candidate asset in the candidate asset pool: Multi-dimensional scoring: Scores Sf, Sb, St, Sd, Sh, and Sn are calculated for each dimension, including collection hit rate, browsing hit rate, tag similarity, department consistency, popularity, and freshness. The tag similarity score St can be calculated using vector semantic similarity or knowledge graph relationship strength.

[0056] The weighted fusion calculation of the comprehensive score S is as follows: Based on the learnable weight coefficients λ1~λ6, S = λ1Sf + λ2Sb + λ3St + λ4Sd + λ5Sh + λ6Sn is calculated, and the freshness decay function Sn = exp(-Δt / τ) can be introduced as needed.

[0057] Online weight optimization: User feedback (clicks, favorites, dwell time, recommendation reason evaluation, etc.) collected by the user interaction and explanation agent is used as a reward signal. The weights of each dimension and hyperparameter τ are adjusted online through reinforcement learning (such as policy gradient or Q-learning) to continuously evolve the scoring model.

[0058] Knowledge graph maintenance and query: Maintain the knowledge graph of data assets, with nodes containing entities such as assets, tags, and departments, and edges representing relationships; provide knowledge graph adjacency recall capabilities to the policy and candidate management agents during the recall phase, and enhance semantic relevance calculation based on the graph structure during the scoring phase; and continuously update the graph based on the launch of new assets and user behavior feedback.

[0059] For example, the comprehensive scoring formula can be defined as S=λ1Sf+λ2Sb+λ3St+λ4Sd+λ5Sh+λ6Sn, where Sf represents the collection hit score, Sb represents the browsing hit score, St represents the tag similarity score, Sd represents the department consistency score, Sh represents the popularity normalization score, Sn represents the freshness decay score, λ1 to λ6 are the corresponding configuration weights, and λ1+λ2+λ3+λ4+λ5+λ6=1.

[0060] Figure 5 This is a flowchart illustrating the process of generating recommendation reasons, such as... Figure 5 As shown, in some implementations, the tag similarity score can be calculated according to the normalization function of the tag intersection size and the candidate tag set size to avoid excessive amplification of similarity by high-frequency general tags; the freshness score can be calculated according to the time decay function between the asset creation time and the current time, for example Sn=exp(-Δt / τ), where Δt represents the difference between the asset release time and the current time, and τ represents the freshness decay coefficient.

[0061] For example, for a candidate asset, if its tags highly overlap with the tags currently followed by users, and it is also accessed by a large number of users in the same department, then the St and Sd scores will be higher; if the asset is recently created and its popularity is rising rapidly, then the Sn and Sh scores will also increase accordingly. By superimposing multiple signals, recommendation bias can be avoided due to anomalies in only one indicator.

[0062] When a user lacks sufficient collection and browsing history, the system automatically reduces the weights of Sf and Sb, and increases the weights of Sd and Sh. At the same time, it relaxes the tag recall threshold to ensure that a sufficient number of high-quality candidates can still be output even when behavioral data is insufficient.

[0063] In step 140, sorted output and interpretable recommendations are presented.

[0064] In the embodiments of this application, Figure 6This is a flowchart illustrating the process of generating recommendation reasons, such as... Figure 6 As shown, the scoring and knowledge agent sends candidate assets and their comprehensive scores back to the strategy and candidate management agent, which then passes them on to the user interaction and explanation agent. The user interaction and explanation agent sorts all candidate assets in descending order based on the comprehensive score S, selects the top N assets as the final recommendation list, and then executes the following: For each data asset in the recommendation list, a structured recommendation reason is automatically generated based on its recall source information and the sub-dimensions with higher scores.

[0065] Generate a rating radar chart, using normalized scores across six dimensions—collection, browsing, tags, department, popularity, and freshness—as axes to visually demonstrate the strength or weakness of assets in each dimension.

[0066] It receives explicit feedback from users, such as likes, questions, or corrections, on the reasons for recommendations, and sends these feedback signals back to the scoring and knowledge intelligence agent for online optimization of the scoring model.

[0067] In step 150, the pre-implementation sandbox agent performs policy pre-verification and security push.

[0068] In this embodiment, the pre-launch sandbox agent is deployed in a non-production sandbox environment. It periodically obtains snapshots of the current policy model from the policy and candidate management agent and the scoring and knowledge agent. By injecting perturbations such as missing user behavior, label drift, and cold start, it simulates recommendation performance under extreme scenarios to evaluate policy stability and quality metrics. New policies that pass verification are pushed to the production agent in a canary release manner, while previous versions are retained. If production metrics regress, the system automatically rolls back, ensuring the recommendation system continues to evolve with zero risk.

[0069] In step 160, the federated recommends that the agent network achieve cross-organizational collaborative evolution.

[0070] In this embodiment, when the system is deployed across multiple organizational nodes, the scoring and knowledge agents within each organization collaborate through a federated learning framework. Each node calculates model gradients or updates parameters on its local data, and uploads them to the federated aggregation server after encryption. The aggregation server securely aggregates parameters from each node and then distributes the global update back to each node, updating the local scoring model and knowledge graph embedding representation. This process achieves a joint improvement in recommendation capabilities across organizations without sharing raw user behavior data or asset details.

[0071] In summary, this application utilizes a first intelligent agent to collect user status and scenario information, and a second intelligent agent to dynamically schedule multiple parallel recall channels, constructing a diverse candidate asset pool to effectively overcome the bias and cold start problems of a single recall strategy. A third intelligent agent performs comprehensive scoring and weighted fusion from six dimensions: collection hit rate, browsing hit rate, tag similarity, departmental consistency, popularity, and freshness, achieving multi-dimensional and accurate evaluation. The first intelligent agent outputs recommendation results based on the comprehensive score ranking and generates explainable recommendation reasons. This method significantly improves the accuracy, personalization, and diversity of data asset recommendations, while enhancing the explainability and user trust of the recommendation results, making it suitable for intelligent discovery scenarios in enterprise data asset management.

[0072] Figure 2 A flowchart illustrating a data asset recommendation method based on multi-path recall and comprehensive scoring is provided, such as... Figure 2 As shown, it includes: 210: The first intelligent agent collects user status information and scene information, converts the user status information and scene information into a first vector, and sends it to the second intelligent agent; 220: The second intelligent agent dynamically schedules multiple parallel recall channels based on the first vector, constructs a candidate asset pool, and sends the candidate asset pool to the third intelligent agent; The multiple recall channels include at least the tag similarity recall channel, collection preference recall channel, recently viewed recall channel, popular recall channel within the same department, and global popular recall channel; 230: The third intelligent agent scores each candidate asset in the candidate asset pool from the dimensions of collection hit, browsing hit, tag similarity, department consistency, popularity and freshness, and weights and fuses the scores of each item to obtain the comprehensive score of each candidate asset. 240: The first intelligent agent sorts the candidate assets in descending order according to the comprehensive score, selects the top-ranked preset number of assets as the recommended result list, and generates recommendation reasons based on the recall source of each candidate asset and the sub-item score.

[0073] In some embodiments, step 210 includes: The first intelligent agent acquires user identifier, user historical behavior data, user department information, and context information of the current recommendation scenario; wherein, the user historical behavior data includes the user's historical collection asset list, recently viewed asset list, and explicit feedback records of the user on historical recommendation results, and the context information includes the currently accessed page type, the currently viewed asset identifier, and the user's current task intent; The first intelligent agent performs statistical analysis on the user's historical behavior data and extracts behavior density features, which include at least the number of collected assets, recent browsing frequency, and active feedback activity. The first intelligent agent performs scene classification on the context information and determines the scene type label; The first intelligent agent concatenates and encodes the behavior density features and the scene type label to form the first vector.

[0074] In some embodiments, step 220 includes: The second agent parses the first vector and extracts user behavior density features and scene type labels; Based on the scenario type tags and the user behavior density features, a subset of channels activated in this round is determined from the recall channel pool, and a recall quantity quota and initial weight coefficient are allocated to each activated channel; the recall channel pool includes tag similarity recall channels, collection preference recall channels, recently viewed recall channels, popular recall channels in the same department, and global popular recall channels; The activation channel is used to recall the data asset library in parallel, and the candidate asset list returned by each channel is received. Each candidate asset carries the original sorting position and matching score of its channel. The received candidate assets are deduplicated and merged using the asset identifier as the primary key. When the same asset is hit by multiple channels, the hit source markers are aggregated and the scores of each channel are merged to form an initial confidence level. The candidate assets after deduplication and merging are filtered to remove at least the assets that users have marked as uninteresting, and assets that meet the exposure fatigue criteria are down-weighted or removed. Output a structured candidate asset pool, in which each candidate asset includes an asset identifier, asset name, recall source list, initial confidence level, and source evidence chain.

[0075] In some embodiments, step 230 includes: The third intelligent agent calculates a collection hit score based on whether the candidate asset appears in the user's collection list; and calculates a browsing hit score based on whether the candidate asset is associated with the user's recently viewed assets. The third agent converts the labels, names, and descriptions of the candidate assets into semantic vectors, calculates the cosine similarity with the semantic vectors of the current context assets or the user interest profile vectors, and obtains the label similarity score. The third intelligent agent determines whether the department to which the candidate asset belongs is consistent with the department to which the user belongs, and obtains a department consistency score; The third intelligent agent normalizes the access frequency of the candidate assets across the entire platform or department to obtain a heat normalization score. The third intelligent agent calculates the freshness decay score using a time decay function based on the time difference between the release time of the candidate asset and the current time. The third intelligent agent calculates the comprehensive score of each candidate asset based on preset weight coefficients.

[0076] In some embodiments, step 240 includes: The first intelligent agent receives the comprehensive score of each candidate asset returned by the third intelligent agent; The first intelligent agent sorts the candidate assets from high to low according to the comprehensive score, and extracts the top N candidate assets to form a recommendation result list, where N is the preset number of recommendation displays; For each candidate asset in the recommendation result list, the first agent selects or splices a structured recommendation reason text from a preset recommendation reason template library based on its recall source channel and the top K sub-dimensions with the highest scores. The first intelligent agent generates a scoring radar chart for each candidate asset. The radar chart uses collection hit, browsing hit, tag similarity, department consistency, popularity, and freshness as axes, and the value of each axis is the normalized value of the corresponding sub-item score. The first intelligent agent outputs the list of recommended results, the text of the reasons for the recommendations, and the rating radar chart to the user interface for display.

[0077] In some embodiments, it also includes: In a non-production sandbox environment, the pre-rehearsal sandbox agent periodically obtains snapshots of the current policy model from the second agent and the third agent. By injecting preset perturbation patterns into the sandbox environment, the recommended performance under extreme scenarios is simulated, and the stability index of the strategy is evaluated. Once the evaluation is passed, the updated policy configuration will be pushed to the second and third agents in the production environment in a gray-scale manner, while retaining the previous version; If a production metric is detected to have regressed, the system will automatically roll back to the previous version.

[0078] By using a pre-launch sandbox agent to periodically capture policy snapshots in a non-production environment and injecting perturbation patterns to simulate extreme scenarios, the stability of the strategy can be assessed in advance without affecting online services, significantly reducing deployment risks. Employing a canary rollout and version retention mechanism, it supports automatic rollback when production metrics are rolled back, achieving safe iteration and rapid recovery capabilities for the recommendation system. This method improves the reliability, fault tolerance, and continuous delivery efficiency of recommendation strategies.

[0079] In some embodiments, when the method is deployed across multiple organizational nodes, the third agents within each organizational node collaborate through a federated learning framework, including: The third agent at each organizational node calculates the gradient or update parameters of the scoring model on local data, and uploads them to the federated aggregation server after encryption. The federated aggregation server securely aggregates the parameters uploaded by each organization node and generates a global parameter update. The federated aggregation server distributes the global parameter updates back to the third agents of each organization node, updating the local scoring model and knowledge graph embedding representation of each node.

[0080] This approach employs a federated learning framework to facilitate collaborative scoring models across multiple organizational nodes. Each node computes gradients or parameter updates only on its local data, which are then encrypted and uploaded to an aggregation server for secure aggregation before being distributed back to each node to update its local model. This method eliminates the need to share raw data, effectively protecting the data privacy and security of each organization, while leveraging cross-organizational data distribution to enhance the generalization ability and accuracy of the scoring model. Furthermore, the introduction of joint updates using knowledge graph embeddings further strengthens the capture of asset association semantics. This solution addresses the data silo problem, enables privacy-preserving collaborative modeling, and significantly improves the overall performance and scalability of the recommender system.

[0081] Figure 7 A schematic diagram of a data asset recommendation device based on multi-path recall and comprehensive scoring is provided, such as... Figure 7 As shown, it includes: The first processing module is used for the first intelligent agent to collect user status information and scene information, and to convert the user status information and scene information into a first vector and send it to the second intelligent agent. The second processing module is used by the second intelligent agent to dynamically schedule multiple parallel recall channels based on the first vector, construct a candidate asset pool, and send the candidate asset pool to the third intelligent agent; the multiple recall channels include at least a tag similarity recall channel, a collection preference recall channel, a recently viewed recall channel, a popular recall channel within the same department, and a global popular recall channel. The third processing module is used by the third intelligent agent to score each candidate asset in the candidate asset pool from the dimensions of collection hit, browsing hit, tag similarity, department consistency, popularity and freshness, and to weight and fuse the scores of each item to obtain the comprehensive score of each candidate asset. The fourth processing module is used by the first intelligent agent to sort the candidate assets in descending order according to the comprehensive score, select a preset number of assets with the highest ranking as the recommended result list, and generate recommendation reasons based on the recall source of each candidate asset and the sub-score.

[0082] In some possible implementations, the first processing module is used for the first intelligent agent to acquire user identifier, user historical behavior data, user department information, and context information of the current recommendation scenario; wherein, the user historical behavior data includes the user's historical collection asset list, recently viewed asset list, and explicit feedback records of the user on historical recommendation results, and the context information includes the currently accessed page type, the currently viewed asset identifier, and the user's current task intent; The first intelligent agent performs statistical analysis on the user's historical behavior data and extracts behavior density features, which include at least the number of collected assets, recent browsing frequency, and active feedback activity. The first intelligent agent performs scene classification on the context information and determines the scene type label; The first intelligent agent concatenates and encodes the behavior density features and the scene type label to form the first vector.

[0083] In some possible implementations, a second processing module is used by the second agent to parse the first vector and extract user behavior density features and scene type labels; Based on the scenario type tags and the user behavior density features, a subset of channels activated in this round is determined from the recall channel pool, and a recall quantity quota and initial weight coefficient are allocated to each activated channel; the recall channel pool includes tag similarity recall channels, collection preference recall channels, recently viewed recall channels, popular recall channels in the same department, and global popular recall channels; The activation channel is used to recall the data asset library in parallel, and the candidate asset list returned by each channel is received. Each candidate asset carries the original sorting position and matching score of its channel. The received candidate assets are deduplicated and merged using the asset identifier as the primary key. When the same asset is hit by multiple channels, the hit source markers are aggregated and the scores of each channel are merged to form an initial confidence level. The candidate assets after deduplication and merging are filtered to remove at least the assets that users have marked as uninteresting, and assets that meet the exposure fatigue criteria are down-weighted or removed. Output a structured candidate asset pool, in which each candidate asset includes an asset identifier, asset name, recall source list, initial confidence level, and source evidence chain.

[0084] In some possible implementations, a third processing module is used by the third intelligent agent to calculate a collection hit score based on whether the candidate asset appears in the user's collection list; and to calculate a browsing hit score based on whether the candidate asset is associated with the user's recently viewed assets. The third agent converts the labels, names, and descriptions of the candidate assets into semantic vectors, calculates the cosine similarity with the semantic vectors of the current context assets or the user interest profile vectors, and obtains the label similarity score. The third intelligent agent determines whether the department to which the candidate asset belongs is consistent with the department to which the user belongs, and obtains a department consistency score; The third intelligent agent normalizes the access frequency of the candidate assets across the entire platform or department to obtain a heat normalization score. The third intelligent agent calculates the freshness decay score using a time decay function based on the time difference between the release time of the candidate asset and the current time. The third intelligent agent calculates the comprehensive score of each candidate asset based on preset weight coefficients.

[0085] In some possible implementations, a fourth processing module is used for the first agent to receive the comprehensive score of each candidate asset returned by the third agent; The first intelligent agent sorts the candidate assets from high to low according to the comprehensive score, and extracts the top N candidate assets to form a recommendation result list, where N is the preset number of recommendation displays; For each candidate asset in the recommendation result list, the first agent selects or splices a structured recommendation reason text from a preset recommendation reason template library based on its recall source channel and the top K sub-dimensions with the highest scores. The first intelligent agent generates a scoring radar chart for each candidate asset. The radar chart uses collection hit, browsing hit, tag similarity, department consistency, popularity, and freshness as axes, and the value of each axis is the normalized value of the corresponding sub-item score. The first intelligent agent outputs the list of recommended results, the text of the reasons for the recommendations, and the rating radar chart to the user interface for display.

[0086] In some possible implementations, the fourth processing module is used to pre-simulate the sandbox agent in a non-production sandbox environment and periodically obtain current policy model snapshots from the second agent and the third agent. By injecting preset perturbation patterns into the sandbox environment, the recommended performance under extreme scenarios is simulated, and the stability index of the strategy is evaluated. Once the evaluation is passed, the updated policy configuration will be pushed to the second and third agents in the production environment in a gray-scale manner, while retaining the previous version; If a production metric is detected to have regressed, the system will automatically roll back to the previous version.

[0087] In some possible implementations, a third processing module is used by the third agent of each organizational node to calculate the gradient or update parameters of the scoring model on local data, and then uploads them to the federated aggregation server after encryption. The federated aggregation server securely aggregates the parameters uploaded by each organization node and generates a global parameter update. The federated aggregation server distributes the global parameter updates back to the third agents of each organization node, updating the local scoring model and knowledge graph embedding representation of each node.

[0088] Those skilled in the art will readily recognize that, based on the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein, the present invention can be implemented in hardware or a combination of hardware and computer software. Whether a function is implemented in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the present invention.

[0089] It should be noted that, Figure 7 The division of modules / units is illustrative and represents only one logical functional division; in actual implementation, other division methods are possible. For example, two or more functions can be integrated into a single data acquisition module. The integrated modules described above can be implemented either in hardware or as software functional modules.

[0090] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-mentioned data asset recommendation methods based on multi-path recall and comprehensive scoring. That is, an electronic device according to an embodiment of the present invention may include, but is not limited to: a processor and a memory; the memory is used to store the computer program; the processor is used to execute the data asset recommendation method based on multi-path recall and comprehensive scoring shown in any embodiment of the present invention by calling the computer program.

[0091] In one alternative embodiment, an electronic device is provided, such as Figure 8 As shown, Figure 8The illustrated electronic device 800 includes a processor 801 and a memory 803. The processor 801 and the memory 803 are connected, for example, via a bus 802. Optionally, the electronic device 800 may further include a transceiver 804, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 804 is not limited to one type, and the structure of the electronic device 800 does not constitute a limitation on the embodiments of the present invention.

[0092] Among them, electronic devices can also be terminal devices, which can be any device that can install applications, including at least one of smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, smart TVs, and smart in-vehicle devices.

[0093] It should be noted that, Figure 8 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0094] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned data asset recommendation methods based on multi-path recall and comprehensive scoring.

[0095] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0096] It should be noted that the terms "first," "second," etc., used in the specification and claims of this invention are used to distinguish similar objects and represent a limitation on a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of the invention described herein can be implemented in an order other than that shown or described.

[0097] Those skilled in the art will recognize that this invention can be implemented as a system, method, or computer program product. Therefore, this invention can be specifically implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, this invention can also be implemented as a computer program product contained in one or more computer-readable media, which includes computer-readable program code.

[0098] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A data asset recommendation method based on multi-path recall and comprehensive scoring, characterized in that, include: The first intelligent agent collects user status information and scene information, converts the user status information and scene information into a first vector, and sends it to the second intelligent agent. The second intelligent agent dynamically schedules multiple parallel recall channels based on the first vector, constructs a candidate asset pool, and sends the candidate asset pool to the third intelligent agent; the multiple recall channels include at least a tag similarity recall channel, a collection preference recall channel, a recently viewed recall channel, a popular recall channel within the same department, and a global popular recall channel; The third intelligent agent scores each candidate asset in the candidate asset pool based on the dimensions of collection hit rate, browsing hit rate, tag similarity, department consistency, popularity, and freshness. The scores of each sub-item are weighted and fused to obtain the comprehensive score of each candidate asset. The first intelligent agent sorts the candidate assets in descending order according to the comprehensive score, selects the top-ranked assets as a recommended result list, and generates recommendation reasons based on the recall source of each candidate asset and the sub-score.

2. The method according to claim 1, characterized in that, The first intelligent agent collects user state information and scene information, converts the user state information and scene information into a first vector, and sends it to the second intelligent agent, including: The first intelligent agent acquires user identifier, user historical behavior data, user department information, and context information of the current recommendation scenario; wherein, the user historical behavior data includes the user's historical collection asset list, recently viewed asset list, and explicit feedback records of the user on historical recommendation results, and the context information includes the currently accessed page type, the currently viewed asset identifier, and the user's current task intent; The first intelligent agent performs statistical analysis on the user's historical behavior data and extracts behavior density features, which include at least the number of collected assets, recent browsing frequency, and active feedback activity. The first intelligent agent performs scene classification on the context information and determines the scene type label; The first intelligent agent concatenates and encodes the behavior density features and the scene type label to form the first vector.

3. The method according to claim 1, characterized in that, The second agent, based on the first vector, dynamically schedules multiple parallel recall channels to construct a candidate asset pool, including: The second agent parses the first vector and extracts user behavior density features and scene type labels; Based on the scenario type tags and the user behavior density features, a subset of channels activated in this round is determined from the recall channel pool, and a recall quantity quota and initial weight coefficient are allocated to each activated channel; the recall channel pool includes tag similarity recall channels, collection preference recall channels, recently viewed recall channels, popular recall channels in the same department, and global popular recall channels; The activation channel is used to recall the data asset library in parallel, and the candidate asset list returned by each channel is received. Each candidate asset carries the original sorting position and matching score of its channel. The received candidate assets are deduplicated and merged using the asset identifier as the primary key. When the same asset is hit by multiple channels, the hit source markers are aggregated and the scores of each channel are merged to form an initial confidence level. The candidate assets after deduplication and merging are filtered to remove at least the assets that users have marked as uninteresting, and assets that meet the exposure fatigue criteria are down-weighted or removed. Output a structured candidate asset pool, in which each candidate asset includes an asset identifier, asset name, recall source list, initial confidence level, and source evidence chain.

4. The method according to claim 1, characterized in that, The third intelligent agent scores each candidate asset in the candidate asset pool based on four dimensions: collection hit rate, browsing hit rate, tag similarity, department consistency, popularity, and freshness. The scores from each dimension are then weighted and fused to obtain a comprehensive score for each candidate asset, including: The third intelligent agent calculates a collection hit score based on whether the candidate asset appears in the user's collection list; and calculates a browsing hit score based on whether the candidate asset is associated with the user's recently viewed assets. The third agent converts the labels, names, and descriptions of the candidate assets into semantic vectors, calculates the cosine similarity with the semantic vectors of the current context assets or the user interest profile vectors, and obtains the label similarity score. The third intelligent agent determines whether the department to which the candidate asset belongs is consistent with the department to which the user belongs, and obtains a department consistency score; The third intelligent agent normalizes the access frequency of the candidate assets across the entire platform or department to obtain a heat normalization score. The third intelligent agent calculates the freshness decay score using a time decay function based on the time difference between the release time of the candidate asset and the current time. The third intelligent agent calculates the comprehensive score of each candidate asset based on preset weight coefficients.

5. The method according to claim 1, characterized in that, The first intelligent agent sorts the candidate assets in descending order based on the comprehensive score, and selects a preset number of top-ranked assets as the recommended result list for output, including: The first intelligent agent receives the comprehensive score of each candidate asset returned by the third intelligent agent; The first intelligent agent sorts the candidate assets from high to low according to the comprehensive score, and extracts the top N candidate assets to form a recommendation result list, where N is the preset number of recommendation displays; For each candidate asset in the recommendation result list, the first agent selects or splices a structured recommendation reason text from a preset recommendation reason template library based on its recall source channel and the top K sub-dimensions with the highest scores. The first intelligent agent generates a scoring radar chart for each candidate asset. The radar chart uses collection hit, browsing hit, tag similarity, department consistency, popularity, and freshness as axes, and the value of each axis is the normalized value of the corresponding sub-item score. The first intelligent agent outputs the list of recommended results, the text of the reasons for the recommendations, and the rating radar chart to the user interface for display.

6. The method according to claim 1, characterized in that, The method further includes: In a non-production sandbox environment, the pre-rehearsal sandbox agent periodically obtains snapshots of the current policy model from the second agent and the third agent. By injecting preset perturbation patterns into the sandbox environment, the recommended performance under extreme scenarios is simulated, and the stability index of the strategy is evaluated. Once the evaluation is passed, the updated policy configuration will be pushed to the second and third agents in the production environment in a gray-scale manner, while retaining the previous version; If a production metric is detected to have regressed, the system will automatically roll back to the previous version.

7. The method according to claim 1, characterized in that, When the method is deployed across multiple organizational nodes, the third agents within each organizational node collaborate through a federated learning framework, including: The third agent at each organizational node calculates the gradient or update parameters of the scoring model on local data, and uploads them to the federated aggregation server after encryption. The federated aggregation server securely aggregates the parameters uploaded by each organization node and generates a global parameter update. The federated aggregation server distributes the global parameter updates back to the third agents of each organization node, updating the local scoring model and knowledge graph embedding representation of each node.

8. A data asset recommendation device based on multi-path recall and comprehensive scoring, characterized in that, include: The first processing module is used for the first intelligent agent to collect user status information and scene information, and to convert the user status information and scene information into a first vector and send it to the second intelligent agent. The second processing module is used by the second intelligent agent to dynamically schedule multiple parallel recall channels based on the first vector, construct a candidate asset pool, and send the candidate asset pool to the third intelligent agent; the multiple recall channels include at least a tag similarity recall channel, a collection preference recall channel, a recently viewed recall channel, a popular recall channel within the same department, and a global popular recall channel. The third processing module is used by the third intelligent agent to score each candidate asset in the candidate asset pool from the dimensions of collection hit, browsing hit, tag similarity, department consistency, popularity and freshness, and to weight and fuse the scores of each item to obtain the comprehensive score of each candidate asset. The fourth processing module is used by the first intelligent agent to sort the candidate assets in descending order according to the comprehensive score, select a preset number of assets with the highest ranking as the recommended result list, and generate recommendation reasons based on the recall source of each candidate asset and the sub-score.

9. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the data asset recommendation method based on multi-path recall and comprehensive scoring as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which is loaded and executed by a processor to enable the computer to implement the data asset recommendation method based on multiple-way recall and comprehensive scoring as described in any one of claims 1 to 7.

Images