Resource recommendation method and device, equipment and storage medium

Abstract

The embodiment of the invention provides a resource recommendation method and device, equipment and a storage medium, and is applied to the technical field of data personalized recommendation. According to the method, the local lightweight resource index pool and the local user behavior database are constructed, preference analysis and multi-dimensional scoring recommendation are directly performed on the terminal, the hysteresis of cloud computing is avoided, quick and accurate response can be realized, network dependence and server pressure are remarkably reduced, and the user experience is improved. Therefore, the problem of poor recommendation timeliness and accuracy in the existing user personalized data pushing scheme is solved.

Description

Technical Field

[0001] This disclosure relates to the field of personalized data recommendation technology, and in particular to a resource recommendation method, apparatus, device, and storage medium. Background Technology

[0002] Existing personalized data recommendation processes all involve analyzing user behavior, such as clicking on content blocks, browsing content, and playing time, and reporting these behaviors to the backend. The backend then analyzes this user behavior and combines it into different content recommendations for the user. This process leads to long waiting times for users, a poor user experience, increased service bandwidth costs, and network congestion, which in turn results in poor recommendation accuracy.

[0003] Therefore, existing personalized data push solutions suffer from poor timeliness and accuracy in recommendations. Summary of the Invention

[0004] This disclosure provides a resource recommendation method, apparatus, device, and storage medium.

[0005] According to a first aspect of this disclosure, a resource recommendation method is provided, the method comprising: Synchronize resource information under predefined content categories from the server and build a local resource index pool based on the synchronized resource information; Collect user behavior data within the ecosystem applications associated with the client and store the behavior data in a local database; Based on behavioral data in the local database, the user's preferred content categories are determined through a pre-set recommendation model; Based on the user's preferred content categories, a set of corresponding candidate resource identifiers is selected from the local resource index pool; Each resource identifier in the candidate resource identifier set is comprehensively scored and sorted to generate a recommended resource list; Render and display the list of recommended resources.

[0006] In some implementations of the first aspect, the aforementioned method, based on behavioral data in a local database, determines the user's preferred content category through a pre-defined recommendation model, including: Extract user interaction behavior sequences, application usage duration distribution, and context state data obtained through the client operating system interface from the local database. The context state data includes the type of currently connected peripheral device, system volume mode, and the number of simultaneously active user accounts detected. For the interaction behavior sequence, determine the weights that decay over time, generate corresponding scene features for the context state data, and determine the comprehensive behavior feature vector; The comprehensive behavioral feature vector is input into the preset recommendation model to generate the real-time probability distribution of user preferences for each predefined content category; Based on the real-time preference probability distribution, content categories with probability values ​​exceeding a dynamic threshold are selected and determined as the user's current preferred content categories. The dynamic threshold is adaptively adjusted based on the user's current active duration and historical interest stability.

[0007] In some implementations of the first aspect, the aforementioned determination of time-decreasing weights for the interaction sequence includes: For user interaction events, an adaptive decay strategy is used to calculate weights. This adaptive decay strategy dynamically adjusts the decay rate parameter based on the frequency of recent user behavior, satisfying the formula: decay rate λ t =λ0×[1+β×(f current -f avg ) / f avg ], Where λ0 is the reference decay rate, β is the adjustment coefficient, and f current f represents the frequency of user behavior during the current time period. avg This represents the average frequency of user behavior in history. Generate corresponding scene features based on context state data, including: processing device state change events and time period events through a scene decision model, using hierarchical judgment logic, outputting a confidence score for each identified scene, and generating soft-coded scene features.

[0008] In some implementations of the first aspect, the aforementioned determination of time-decreasing weights for the interaction behavior sequence, generation of corresponding scene features for context state data, and determination of a comprehensive behavior feature vector include: Based on the intersection of interactive behavior features and scene features on preset key dimensions, cross feature sub-vectors are formed; Based on the attention network, the importance weights of the original feature dimension and the cross feature are calculated according to the intermediate features of the current fusion. Based on the importance weights, the original features and the cross features are weighted and synthesized to obtain the initial comprehensive feature vector; The initial comprehensive feature vector is subjected to comparative regularization constraints. After optimization by constructing positive and negative sample pairs, the comprehensive behavioral feature vector is obtained.

[0009] In some implementations of the first aspect, the above-mentioned selection of content categories whose probability values ​​exceed a dynamic threshold based on the real-time preference probability distribution is used to determine the user's current preferred content category, including: Based on the real-time preference probability distribution and comprehensive behavioral feature vector, multiple dynamic factors such as the concentration of user interest distribution, behavioral stability, and current session context are determined. Based on multiple dynamic factors such as the concentration of user interest distribution, behavioral stability, and current session context, the dynamic threshold for adaptive preference selection is calculated in real time. The real-time preference probability distribution is smoothed over time, and the probability values ​​are dynamically adjusted based on real-time user interaction feedback. From the smoothed probability distribution, select categories that exceed the adaptive preference selection dynamic threshold, perform semantic deduplication and scene adaptation processing, and determine the user's current preferred content category.

[0010] In some implementations of the first aspect, the above method may also include: Monitor the display frequency of resources and user hit feedback for each content category. For content categories with a display frequency higher than a preset threshold, trigger incremental synchronization of the corresponding resource index entries from the server to supplement the local resource index pool. For resource index entries that are not hit within a preset time period, perform a local cleanup operation.

[0011] In some implementations of the first aspect, the above method also includes identifying the current scene type corresponding to the client; The recommended resource list is rendered and displayed, including: If the current scenario is a comprehensive information flow scenario, a hybrid strategy is adopted. From the recommended resource list, resource identifiers of one or more categories with the highest current user preference are selected as the first part of resource identifiers; from the recommended resource list, resource identifiers of other categories related to the current user preference are selected as the second part of resource identifiers; the first part of resource identifiers and the second part of resource identifiers are mixed according to a preset ratio to generate and display a hybrid recommendation list. If the current scenario type is a vertical content scenario, a focusing strategy is adopted to determine the specified vertical category corresponding to the vertical content scenario; all resource icons of the specified vertical category are filtered from the recommended resource list; and the filtered resource icons are arranged and displayed in a diverse manner according to the subcategory tags of the selected resource icons under the specified vertical category.

[0012] According to a second aspect of this disclosure, a resource recommendation apparatus is provided, the apparatus comprising: The building module is used to synchronize resource information under predefined content categories from the server and build a local resource index pool based on the synchronized resource information; The data collection module is used to collect user behavior data within the ecosystem applications associated with the client and store the behavior data in a local database. The analysis module is used to determine the user's preferred content categories based on behavioral data in the local database and through a preset recommendation model; The filtering module is used to filter the corresponding set of candidate resource identifiers from the local resource index pool based on the user's preferred content categories. The generation module is used to comprehensively score and sort each resource identifier in the candidate resource identifier set to generate a recommended resource list. The display module is used to render and display the list of recommended resources.

[0013] According to a third aspect of this disclosure, an electronic device is provided, comprising: a memory and a processor, wherein a computer program is stored in the memory, and the processor executes the program to implement the method described above.

[0014] According to a fourth aspect of this disclosure, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0015] The resource recommendation method, apparatus, device, and storage medium disclosed herein construct a local lightweight resource index pool and a local user behavior database, and perform preference analysis and multi-dimensional rating recommendations directly on the terminal. Without the lag of cloud computing, it can achieve fast and accurate response, significantly reduce network dependence and server pressure, and thus solve the problems of poor recommendation timeliness and accuracy in existing user personalized data push solutions.

[0016] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of this disclosure, nor is it intended to restrict the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0017] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of this disclosure. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein: Figure 1 A flowchart illustrating a resource recommendation method according to an embodiment of the present disclosure is shown; Figure 2 A flowchart illustrating another resource recommendation method according to an embodiment of the present disclosure is shown; Figure 3 A structural block diagram of a resource recommendation apparatus according to an embodiment of the present disclosure is shown; Figure 4 A block diagram of an exemplary electronic device capable of implementing embodiments of the present disclosure is shown. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0019] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0020] In order to solve the above-mentioned technical problems, this disclosure proposes a resource recommendation method, apparatus, device and storage medium. By constructing a local lightweight resource index pool and a local user behavior database, preference analysis and multi-dimensional rating recommendations are performed directly on the terminal. Without the lag of cloud computing, it can achieve fast and accurate response, significantly reduce network dependence and server pressure, and thus solve the problem of poor recommendation timeliness and accuracy in existing user personalized data push solutions.

[0021] Figure 1 The diagram illustrates a flowchart of a resource recommendation method according to an embodiment of this disclosure. This method can be applied to a client, such as... Figure 1 As shown, resource recommendation method 100 may include: S101, synchronize resource information under predefined content categories from the server, and build a local resource index pool based on the synchronized resource information; S102, collect user behavior data within the ecosystem applications associated with the client, and store the behavior data in a local database; S103, based on behavioral data in the local database, determines the user's preferred content category through a preset recommendation model; In some embodiments, the above-mentioned determination of user preferred content categories based on behavioral data in a local database and through a preset recommendation model includes: Extract user interaction behavior sequences, application usage duration distribution, and context state data obtained through the client operating system interface from the local database. The context state data includes the type of currently connected peripheral device, system volume mode, and the number of simultaneously active user accounts detected. For the interaction behavior sequence, determine the weight that decays over time, generate corresponding scene features for the context state data, and determine the comprehensive behavior feature vector. Specifically, for example, connecting a game controller may imply game preference, and detecting multiple accounts may trigger family hybrid recommendation logic. The comprehensive behavioral feature vector is input into the preset recommendation model to generate the real-time probability distribution of user preferences for each predefined content category; Based on the real-time preference probability distribution, content categories with probability values ​​exceeding a dynamic threshold are selected and determined as the user's current preferred content categories. The dynamic threshold is adaptively adjusted based on the user's current active duration and historical interest stability.

[0022] In the above embodiments, by expanding the definition of "behavioral data" from traditional clicks and plays to richer device contexts and interaction scenarios (such as connecting a game controller may imply game preferences, and detecting multiple accounts may trigger family hybrid recommendation logic), and by introducing contextual state data and dynamic thresholds, the determination of preference classification is no longer a simple historical statistics, but rather combines real-time scene intelligent perception, which significantly improves the immediacy and accuracy of preference judgment.

[0023] In some embodiments, the above-mentioned selection of content categories whose probability values ​​exceed a dynamic threshold based on the real-time preference probability distribution, and determination of these categories as the user's current preferred content categories, includes: Based on the real-time preference probability distribution and comprehensive behavioral feature vector, multiple dynamic factors such as the concentration of user interest distribution, behavioral stability, and current session context are determined. Based on multiple dynamic factors such as the concentration of user interest distribution, behavioral stability, and current session context, the dynamic threshold for adaptive preference selection is calculated in real time. The real-time preference probability distribution is smoothed over time, and the probability values ​​are dynamically adjusted based on real-time user interaction feedback. From the smoothed probability distribution, select categories that exceed the adaptive preference selection dynamic threshold, perform semantic deduplication and scene adaptation processing, and determine the user's current preferred content category.

[0024] In the above embodiments, by combining and applying multiple methods such as dynamic thresholding corresponding to adaptive control theory, temporal smoothing corresponding to signal processing, feedback correction corresponding to online learning, and semantic deduplication to the specific recommendation system sub-problem from probability distribution to classification selection, the use of dynamic thresholding + semantic deduplication can bring more accurate recommendations, temporal smoothing can bring a more stable experience, real-time feedback correction can bring faster adaptation, and scene adaptation can bring stronger controllability.

[0025] In some embodiments, determining the time-decreasing weights for the interaction sequence may include: For user interaction events, an adaptive decay strategy is used to calculate weights. This adaptive decay strategy dynamically adjusts the decay rate parameter based on the frequency of recent user behavior, satisfying the formula: decay rate λ t =λ0×[1+β×(f current -f avg ) / f avg ], Where λ0 is the reference decay rate, β is the adjustment coefficient, and f current f represents the frequency of user behavior during the current time period. avg This represents the average frequency of user behavior in history. Generate corresponding scene features based on context state data, including: processing device state change events and time period events through a scene decision model, using hierarchical judgment logic, outputting a confidence score for each identified scene, and generating soft-coded scene features.

[0026] In the above embodiments, by incorporating user behavior patterns into the attenuation calculation, the timeliness judgment of features is made more intelligent, and the fuzzy recognition of the scene is achieved through confidence scores, thereby enhancing the ability to depict complex real-world scenes.

[0027] In some embodiments, the above-described determination of time-decreasing weights for the interaction behavior sequence, generation of corresponding scene features for the context state data, and determination of the comprehensive behavior feature vector may include: Based on the intersection of interactive behavior features and scene features on preset key dimensions, cross feature sub-vectors are formed; Based on the attention network, the importance weights of the original feature dimension and the cross feature are calculated according to the intermediate features of the current fusion. Based on the importance weights, the original features and the cross features are weighted and synthesized to obtain the initial comprehensive feature vector; The initial comprehensive feature vector is subjected to comparative regularization constraints. After optimization by constructing positive and negative sample pairs, the comprehensive behavioral feature vector is obtained.

[0028] In the above embodiments, by applying a contrastive regularization constraint to the initial comprehensive feature vector, this constraint constructs positive and negative sample pairs so that the feature vector maximizes the differentiation between different users while minimizing the feature differences of the same user in adjacent and similar contexts, thereby enhancing the robustness and generalization of the features.

[0029] S104. Based on the user's preferred content classification, filter the corresponding candidate resource identifier set from the local resource index pool; S105, Perform a comprehensive score and sort each resource identifier in the candidate resource identifier set to generate a recommended resource list; In the process of comprehensively scoring and ranking each resource identifier in the candidate resource identifier set, the multi-dimensional scoring rules used include user behavior dimension, content attribute dimension, control dimension and operation dimension; Among them, the data corresponding to the user behavior dimension includes preference intensity calculated based on the user's historical behavior, user activity time periods, and user profile tags; The data corresponding to the content attribute dimension includes the subcategory to which the resource belongs, the matching degree of the content tags, the popularity value of the resource itself, and the release time of the resource; The data corresponding to the control dimensions include the number of times the resource is displayed, the time of the most recent display, and whether it is a duplicate recommendation; The data corresponding to the operational dimension includes whether the resource is marked as operational promotion content, whether it belongs to a currently active targeted recommendation activity, and whether it meets the timeliness requirements.

[0030] In the above embodiments, by creatively integrating four goals—long-term user interests, real-time scenarios, system regulation, and business operations—which are traditionally handled by different teams at different system levels, into a unified real-time scoring model that runs on the client side, the model addresses the four major challenges faced by edge recommendations: personalization accuracy, content freshness, system ecosystem health, and achievement of business goals. This achieves a balance between recommendation accuracy and diversity, seamless optimization of user experience, and self-sustaining health of the system ecosystem.

[0031] S106 renders and displays the recommended resource list.

[0032] In some embodiments, Figure 1 The methods may also include: monitoring the display frequency of resources for each content category and user hit feedback; for content categories with a display frequency higher than a preset threshold, triggering incremental synchronization of the corresponding resource index entries from the server to supplement the local resource index pool. For resource index entries that are not hit within a preset time period, a local cleanup operation is performed to maintain a dynamic balance between the activity and capacity of the local resource index pool.

[0033] In the above embodiments, "display frequency" is monitored as a proxy indicator of demand popularity. High-frequency display categories indicate that users' current interests are concentrated there, and the likelihood of future clicks is high. Triggering incremental synchronization at this time essentially predicts future demand based on local behavior, proactively preheating resources. This represents a paradigm shift from "passive retrieval" to "proactive predictive caching," significantly improving the accuracy and efficiency of resource synchronization. Furthermore, the cleanup criterion is "not being hit within a preset time period." This directly measures the economic value of resources in the local environment; even if a resource has high global popularity, if it remains untouched by a particular user for a long time, its local value is zero and it should be cleaned up. This achieves a balance between "global popularity" and "personal interest," ensuring that every storage space in the local resource pool is used to store content with the highest potential value to the user, maximizing storage utilization. The synchronization behavior in the above embodiments is highly accurate, avoiding the transmission of large amounts of invalid data, and the local database always retains the "hot" resources most likely to be consumed by users, while cold data is promptly eliminated. Moreover, content that users are interested in is more likely to be pre-loaded locally, resulting in faster recommendation responses and more accurate content.

[0034] In some embodiments, Figure 1 The method may also include identifying the current scene type corresponding to the client; The recommended resource list is rendered and displayed, including: If the current scenario is a comprehensive information flow scenario, a hybrid strategy is adopted. From the recommended resource list, resource identifiers of one or more categories with the highest current user preference are selected as the first part of resource identifiers; from the recommended resource list, resource identifiers of other categories related to the current user preference are selected as the second part of resource identifiers; the first part of resource identifiers and the second part of resource identifiers are mixed according to a preset ratio to generate and display a hybrid recommendation list. If the current scenario type is a vertical content scenario, a focusing strategy is adopted to determine the specified vertical category corresponding to the vertical content scenario; all resource icons of the specified vertical category are filtered from the recommended resource list; and the filtered resource icons are arranged and displayed in a diverse manner according to the subcategory tags of the selected resource icons under the specified vertical category.

[0035] In the above embodiments, scenario type is explicitly introduced as the primary criterion for strategy selection in the client-side recommendation. It recognizes that different entry points (such as "Homepage" and "Children's Channel") carry different user mindsets and business goals, thus requiring completely different recommendation logics. This is a deepening of "personalized recommendations," upgrading to "personalized scenarios for each user." Furthermore, by introducing sub-category diversity constraints, it ensures that the content distribution within vertical categories is rich and balanced, improving the browsing experience and avoiding the problem of "seemingly abundant but actually repetitive content." In summary, the above embodiments can achieve the most expected content organization in different scenarios, offering surprises in comprehensive scenarios and sufficient depth in vertical scenarios. Moreover, the hybrid strategy helps improve user stickiness and content discovery, while the focused strategy helps improve the depth of content consumption and user satisfaction in vertical channels.

[0036] In the S101-S106 process, by building a local lightweight resource index pool and a local user behavior database, preference analysis and multi-dimensional rating recommendations are performed directly on the terminal. Without the lag of cloud computing, it can achieve fast and accurate response, significantly reducing network dependence and server pressure, thereby solving the problem of poor recommendation timeliness and accuracy in existing user personalized data push solutions.

[0037] In a specific example, the client-oriented resource recommendation method described above is used to accurately recommend personalized content based on user behavior under limited local storage conditions. Specifically, the client periodically (e.g., daily or each time a user is active) synchronizes resource information for each content category from the server. Each category (e.g., movies, children's content, education, sports, etc.) initially retrieves approximately 1000 resource data entries, mainly including resource IDs and their brief information (e.g., titles, covers, tags, etc.), and stores them in a local database. This "shallow data" synchronization mechanism effectively alleviates the client's storage pressure while providing an indexing foundation for subsequent rapid recommendations.

[0038] When users engage with ecosystem applications (such as LeTV Video and children's apps), their behavioral data (including click records, dwell time, usage frequency, installed applications, and favorites) is recorded in a local database. The system analyzes this user behavior data using a lightweight model (such as a content interest profile model or a collaborative filtering recommendation algorithm) to extract current user preference categories.

[0039] Based on the identified user preferences, the client will prioritize filtering within the corresponding local resource library. This filtering method not only considers user preferences but also incorporates multiple factors to improve the accuracy and diversity of recommendations. These factors include, but are not limited to: 1. User dimension: user history preferences, behavioral activity, usage frequency, time period characteristics, basic attributes (such as age group); 2. Content Dimensions: Main and subcategories of resources, content tags (such as "animation", "education", "tools"), popularity, and update time; 3. Control dimensions: whether the resource has been recommended, whether it was a hit (whether the user interacted with it), display frequency, and deduplication control; 4. Operational Dimension: Whether it is an operational promotion resource, whether it is under a targeted recommendation strategy, whether it is a time-sensitive resource (such as holiday content), etc.

[0040] This method constructs a comprehensive scoring mechanism based on the aforementioned dimensions, quickly extracting the resource ID that best matches the current user's preferences and context from the local database, and then using it to populate content in multiple display modules within the ecosystem (such as the desktop recommendation area and channel recommendation pages). For example, in the hybrid recommendation module, the system can recommend 20 pieces of content, 12 of which are from the user's preferred category, and the remaining 8 are from other highly related categories; while in specific scenarios (such as the children's zone), the recommended resources will come entirely from that category, further subdivided into subcategories such as educational, animation, and English learning categories to meet vertical needs.

[0041] To ensure the timeliness of recommended content and the sufficiency of the resource pool, the client continuously and incrementally retrieves new content for frequently displayed categories, further enriching the local resource pool. Meanwhile, resources that have not been clicked or interacted with by users for an extended period are considered "missed" and removed from the local database through a data cleaning mechanism to maintain resource activity and dynamic balance.

[0042] Furthermore, this system supports a flexible operational configuration mechanism. The filtering rules, weight parameters, and recommendation category priorities used in the recommendation strategy can all be configured by operations personnel in the backend management system (Content Management System, CMS) and pushed to clients via the configuration distribution system. The client takes effect immediately upon receiving the strategy, thus achieving an organic balance between accurate recommendations and operability, satisfying personalized recommendation effects while maintaining a controllable operational window.

[0043] Figure 2 This is a flowchart illustrating a resource recommendation method provided in an embodiment of this disclosure, combined with... Figure 2 The flowchart corresponding to the method in the text is as follows: Figure 1 The resource recommendation method shown will be described in further detail.

[0044] This method is implemented based on multiple applications running on client devices (such as smart TVs and set-top boxes), remote servers, and a data storage center. All parties collaborate to complete the entire process from data synchronization to localized personalized recommendations. The specific steps are as follows: Step 1: Server-side data and model preparation Process 1.1: Behavioral Data Collection and Processing The server receives anonymized behavioral data reported by clients with user authorization, and then cleans, aggregates, and stores it.

[0045] Process 1.2: Global Model Training and Resource Index Construction The server uses the collected behavioral data to periodically train or update the global recommendation model within the "data collection / training / storage stack". Simultaneously, the server "extracts data" from the "data storage center" (such as a content library) and performs "data extraction / filtering" based on strategies such as global popularity and content freshness, generating lightweight resource index entries (containing resource IDs and key metadata) to form a distributable recommendation data package.

[0046] Step 2: The server synchronizes data and models with the client. Process 2.1: Client synchronizes backend recommendation data The server, either proactively or in response to a client request, synchronizes the resource index entry package generated in step 1.2 and organized by content category to the client via the downlink channel. This forms the basis for the client to build its local resource index pool.

[0047] Process 2.2: Client-side synchronous background training model The server synchronizes a lightweight recommendation model (such as a pruned neural network or an interpretable set of decision rules) to the client via a downlink channel. This model is the core algorithm for the client to perform local preference analysis.

[0048] Step 3: Local Client Applications and Data Processing Process 3.1: Local Data Storage and Index Pool Construction The client receives and stores resource index entries and recommendation models from the server. The resource index entries are categorized and stored to build a local resource index pool.

[0049] Process 3.2: Local User Behavior Collection The ecosystem applications running on the client, such as Application 1, Application 2, Application 3, and Application 4, collect user behavior data (such as clicks, plays, and dwell times) locally in real time during user use and store it in the client's local database. This process emphasizes that the behavior data is temporarily stored locally and is not reported in real time.

[0050] Process 3.3: Localized Model Inference and Recommendation Generation The client invokes a locally stored recommendation model, combines it with user behavior history from the local database, performs real-time inference, and analyzes user preferences. Subsequently, based on these preferences, candidate resources are selected from the local resource index pool, sorted through multi-dimensional scoring, and finally a personalized list of recommended resources is generated.

[0051] Process 3.4: Recommended Content Display and Interaction The client renders and displays the generated recommendation list in the corresponding user interface (such as desktop feed or in-app recommendation slot). User interactions with the recommended content (such as clicks) are then collected by applications 1 to 4, forming a closed-loop feedback.

[0052] Step 4: Client Data Reporting and System Evolution Process 4.1: Periodic reporting of anonymized data The client periodically (or under specific triggering conditions) reports the de-identified and aggregated local behavioral data summary, as well as model feedback data (such as gradient updates and performance metrics), to the server.

[0053] Process 4.2: Server-side model and strategy iteration The server uses data aggregated from a massive number of clients to further optimize the global model and data filtering strategy in the "data collection / training / storage stack". In the next cycle, it pushes better models and data to the clients through step two, thus achieving continuous evolution of the whole.

[0054] It can be seen that, Figure 2 A two-tier architecture of cloud collaboration and terminal intelligence is disclosed: the server side is responsible for knowledge mining, condensation, and distribution, namely training large models and filtering high-quality global resources. The client side is responsible for local personalized perception, decision-making, and execution, namely collecting local behavior, running lightweight models, and achieving millisecond-level recommendations. The two are coupled through periodic data / model synchronization and periodic summary reporting, which ensures both the real-time nature and privacy of recommendations, and enables continuous updates to the system's knowledge.

[0055] The above is an introduction to the method embodiments. The following describes the present disclosure further through device embodiments.

[0056] Figure 3 A block diagram of a resource recommendation apparatus according to an embodiment of the present disclosure is shown.

[0057] like Figure 3 As shown, the resource recommendation device 300 may include: Module 301 is used to synchronize resource information under predefined content categories from the server and build a local resource index pool based on the synchronized resource information. The data collection module 302 is used to collect user behavior data within the ecosystem applications associated with the client and store the behavior data in a local database. Analysis module 303 is used to determine the user's preferred content categories based on behavioral data in the local database and through a preset recommendation model; The filtering module 304 is used to filter the corresponding set of candidate resource identifiers from the local resource index pool according to the user's preferred content classification. The generation module 305 is used to comprehensively score and sort each resource identifier in the candidate resource identifier set to generate a recommended resource list. The display module 306 is used to render and display the list of recommended resources.

[0058] In some embodiments, the above-mentioned determination of user preferred content categories based on behavioral data in a local database and through a preset recommendation model includes: Extract user interaction behavior sequences, application usage duration distribution, and context state data obtained through the client operating system interface from the local database. The context state data includes the type of currently connected peripheral device, system volume mode, and the number of simultaneously active user accounts detected. For the interaction behavior sequence, determine the weights that decay over time, generate corresponding scene features for the context state data, and determine the comprehensive behavior feature vector; The comprehensive behavioral feature vector is input into the preset recommendation model to generate the real-time probability distribution of user preferences for each predefined content category; Based on the real-time preference probability distribution, content categories with probability values ​​exceeding a dynamic threshold are selected and determined as the user's current preferred content categories. The dynamic threshold is adaptively adjusted based on the user's current active duration and historical interest stability.

[0059] In some embodiments, determining the time-decaying weights for the interaction sequence includes: For user interaction events, an adaptive decay strategy is used to calculate weights. This adaptive decay strategy dynamically adjusts the decay rate parameter based on the frequency of recent user behavior, satisfying the formula: decay rate λ t =λ0×[1+β×(f current -f avg ) / f avg ], Where λ0 is the reference decay rate, β is the adjustment coefficient, and f currentf represents the frequency of user behavior during the current time period. avg This represents the average frequency of user behavior in history. Generate corresponding scene features based on context state data, including: processing device state change events and time period events through a scene decision model, using hierarchical judgment logic, outputting a confidence score for each identified scene, and generating soft-coded scene features.

[0060] In some embodiments, the above-described determination of time-decaying weights for the interaction behavior sequence, generation of corresponding scene features for the context state data, and determination of the comprehensive behavior feature vector include: Based on the intersection of interactive behavior features and scene features on preset key dimensions, cross feature sub-vectors are formed; Based on the attention network, the importance weights of the original feature dimension and the cross feature are calculated according to the intermediate features of the current fusion. Based on the importance weights, the original features and the cross features are weighted and synthesized to obtain the initial comprehensive feature vector; The initial comprehensive feature vector is subjected to comparative regularization constraints. After optimization by constructing positive and negative sample pairs, the comprehensive behavioral feature vector is obtained.

[0061] In some embodiments, the above-mentioned selection of content categories whose probability values ​​exceed a dynamic threshold based on the real-time preference probability distribution, and determination of these categories as the user's current preferred content categories, includes: Based on the real-time preference probability distribution and comprehensive behavioral feature vector, multiple dynamic factors such as the concentration of user interest distribution, behavioral stability, and current session context are determined. Based on multiple dynamic factors such as the concentration of user interest distribution, behavioral stability, and current session context, the dynamic threshold for adaptive preference selection is calculated in real time. The real-time preference probability distribution is smoothed over time, and the probability values ​​are dynamically adjusted based on real-time user interaction feedback. From the smoothed probability distribution, select categories that exceed the adaptive preference selection dynamic threshold, perform semantic deduplication and scene adaptation processing, and determine the user's current preferred content category.

[0062] In some embodiments, the above-mentioned device can also monitor the resource display frequency and user hit feedback of each content category. For content categories with a display frequency higher than a preset threshold, it can trigger incremental synchronization of the corresponding resource index entries from the server to supplement the local resource index pool. For resource index entries that are not hit within a preset time period, perform a local cleanup operation.

[0063] In some embodiments, the above-described apparatus can also be used to identify the current scene type corresponding to the client; The rendering and display of the recommended resource list as described above may include: If the current scenario is a comprehensive information flow scenario, a hybrid strategy is adopted. From the recommended resource list, resource identifiers of one or more categories with the highest current user preference are selected as the first part of resource identifiers; from the recommended resource list, resource identifiers of other categories related to the current user preference are selected as the second part of resource identifiers; the first part of resource identifiers and the second part of resource identifiers are mixed according to a preset ratio to generate and display a hybrid recommendation list. If the current scenario type is a vertical content scenario, a focusing strategy is adopted to determine the specified vertical category corresponding to the vertical content scenario; all resource icons of the specified vertical category are filtered from the recommended resource list; and the filtered resource icons are arranged and displayed in a diverse manner according to the subcategory tags of the selected resource icons under the specified vertical category.

[0064] 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 this disclosure is not limited to the described order of actions, because according to this disclosure, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this disclosure.

[0065] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0066] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0067] Figure 4 A block diagram of an exemplary electronic device capable of implementing embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0068] Device 400 includes a computing unit 401, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 402 or a computer program loaded into random access memory (RAM) 403 from storage unit 408. The RAM 403 may also store various programs and data required for the operation of device 400. The computing unit 401, ROM 402, and RAM 403 are interconnected via bus 404. An input / output (I / O) interface 405 is also connected to bus 404.

[0069] Multiple components in device 400 are connected to I / O interface 405, including: input unit 406, such as keyboard, mouse, etc.; output unit 407, such as various types of monitors, speakers, etc.; storage unit 408, such as disk, optical disk, etc.; and communication unit 409, such as network card, modem, wireless transceiver, etc. Communication unit 409 allows device 400 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0070] The computing unit 401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 401 performs the various methods and processes described above, such as method 100. For example, in some embodiments, method 100 may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 408. In some embodiments, part or all of the computer program may be loaded and / or installed on device 400 via ROM 402 and / or communication unit 409. When the computer program is loaded into RAM 403 and executed by the computing unit 401, one or more steps of method 100 described above may be performed.

[0071] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0072] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0073] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0074] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0075] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0076] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0077] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this disclosure can be achieved, and this is not limited herein.

[0078] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A resource recommendation method, characterized in that, Applied to a client, the method includes: Synchronize resource information under predefined content categories from the server and build a local resource index pool based on the synchronized resource information; Collect user behavior data within the ecosystem applications associated with the client, and store the behavior data in a local database; Based on the behavioral data in the local database, the user's preferred content categories are determined through a preset recommendation model; Based on the user's preferred content category, a set of corresponding candidate resource identifiers is selected from the local resource index pool; Each resource identifier in the candidate resource identifier set is comprehensively scored and sorted to generate a recommended resource list; The recommended resource list is rendered and displayed.

2. The method according to claim 1, characterized in that, The process of determining user-preferred content categories based on behavioral data in the local database using a preset recommendation model includes: From the local database, extract the user's interaction behavior sequence in the current session and historical sessions, the application usage duration distribution, and the context state data obtained through the client operating system interface. The context state data includes the type of currently connected peripheral device, the system volume mode, and the number of simultaneously active user accounts detected. For the interaction behavior sequence, a weight that decays over time is determined, and corresponding scene features are generated for the context state data to determine a comprehensive behavior feature vector; The comprehensive behavioral feature vector is input into the preset recommendation model to generate a real-time probability distribution of user preferences for each predefined content category; Based on the real-time preference probability distribution, content categories with probability values ​​exceeding a dynamic threshold are selected and determined as the user's current preferred content categories. The dynamic threshold is adaptively adjusted based on the user's current active duration and historical interest stability.

3. The method according to claim 2, characterized in that, Determining the time-decaying weights for the interaction behavior sequence includes: For user interaction events, an adaptive decay strategy is used to calculate weights. This adaptive decay strategy dynamically adjusts the decay rate parameter based on the frequency of recent user behavior, satisfying the formula: decay rate λ t =λ0×[1+β×(f current -f avg ) / f avg ], Where λ0 is the reference decay rate, β is the adjustment coefficient, and f current f represents the frequency of user behavior during the current time period. avg This represents the average frequency of user behavior in history. The step of generating corresponding scene features for the context state data includes: processing device state change events and time period events for the context state data through a scene decision model, using hierarchical judgment logic, outputting a confidence score for each identified scene, and generating soft-coded scene features.

4. The method according to claim 3, characterized in that, The process of determining time-decreasing weights for the interaction behavior sequence, generating corresponding scene features for the context state data, and determining a comprehensive behavior feature vector includes: Based on the intersection of interactive behavior features and scene features on preset key dimensions, cross feature sub-vectors are formed; Based on the attention network, the importance weights of the original feature dimension and the cross feature are calculated according to the intermediate features of the current fusion. Based on the aforementioned importance weights, the original features and cross features are weighted and synthesized to obtain an initial comprehensive feature vector; The initial comprehensive feature vector is subjected to comparative regularization constraints, and after optimization by constructing positive and negative sample pairs, a comprehensive behavioral feature vector is obtained.

5. The method according to claim 2, characterized in that, The step of selecting content categories whose probability values ​​exceed a dynamic threshold based on the real-time preference probability distribution and determining them as the user's current preferred content categories includes: Based on the real-time preference probability distribution and comprehensive behavioral feature vector, multiple dynamic factors such as user interest distribution concentration, behavioral stability, and current session context are determined. Based on multiple dynamic factors such as the concentration of user interest distribution, behavioral stability, and current session context, the dynamic threshold for adaptive preference selection is calculated in real time. The real-time preference probability distribution is subjected to time-series smoothing, and the probability values ​​are dynamically corrected based on real-time user interaction feedback. From the smoothed probability distribution, select categories that exceed the adaptive preference selection dynamic threshold, perform semantic deduplication and scene adaptation processing, and determine the user's current preferred content category.

6. The method according to claim 1, characterized in that, The method further includes: Monitor the display frequency of resources and user hit feedback for each content category. For content categories with a display frequency higher than a preset threshold, trigger incremental synchronization of the corresponding resource index entries from the server to supplement the local resource index pool. For resource index entries that are not hit within a preset time period, perform a local cleanup operation.

7. The method according to claim 1, characterized in that, The method also includes identifying the current scene type corresponding to the client; The rendering and displaying of the recommended resource list includes: If the current scenario is a comprehensive information flow scenario, a hybrid strategy is adopted. From the recommended resource list, one or more categories with the highest user preference are selected as the first part of the resource identifiers; from the recommended resource list, resource identifiers of other categories associated with the user's current preference are selected as the second part of the resource identifiers; the first part of the resource identifiers and the second part of the resource identifiers are mixed in a preset ratio to generate and display a hybrid recommendation list. If the current scenario type is a vertical content scenario, a focusing strategy is adopted to determine the specified vertical category corresponding to the vertical content scenario; all resource identifiers of the specified vertical category are filtered from the recommended resource list; and the filtered resource identifiers are arranged and displayed in a diverse manner according to the subcategory tags under the specified vertical category.

8. A resource recommendation device, characterized in that, The device includes: The building module is used to synchronize resource information under predefined content categories from the server and build a local resource index pool based on the synchronized resource information; The data collection module is used to collect user behavior data within the ecosystem applications associated with the client and store the behavior data in a local database. The analysis module is used to determine the user's preferred content categories based on behavioral data in the local database and through a preset recommendation model; The filtering module is used to filter the corresponding set of candidate resource identifiers from the local resource index pool according to the user's preferred content classification. The generation module is used to comprehensively score and sort each resource identifier in the candidate resource identifier set to generate a recommended resource list; The display module is used to render and display the recommended resource list.

9. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; The memory is characterized in that it stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to perform the method according to any one of claims 1-7.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-7.

Images