A multi-source media asset configuration driving recommendation strategy generation and multi-level intervention method

Abstract

The application discloses a multi-source media asset configuration driving type recommendation strategy generation and multi-level intervention method, and belongs to the technical field of multimedia content recommendation. The method comprises the following steps: step S100, a multi-dimensional combined recommendation strategy configuration model is constructed, a recommendation strategy is defined through structured parameter combination, a double-mode effective mechanism of a'recommendation process effective' mode and a'recommendation complete effective' mode is introduced, and a candidate recommendation pool is dynamically generated; step S200, a double-layer recommendation library architecture of a default recommendation layer and a self-defined recommendation layer is constructed, and multi-source heterogeneous data is uniformly arranged into a standardized recommendation library; step S300, a multi-level intervention mechanism is constructed at a recommendation result level, multi-level fine intervention and whole life cycle tracking of the recommendation result are realized; and intervention tracking data is fed back to the step S100, so that a self-evolution operation closed loop is formed. According to the application, an operation personnel can complete strategy configuration and result intervention at a minute level without modifying codes, and the application is suitable for OTT / IPTV large-screen terminal media asset recommendation scenes.

Description

Technical Field

[0001] This invention relates to the field of multimedia content recommendation technology, specifically, to a method for generating multi-source media asset configuration-driven recommendation strategies and multi-level intervention. Background Technology

[0002] In large-screen terminal media asset recommendation scenarios such as OTT (Over The Top) and IPTV (Internet Protocol Television), recommendation systems typically employ automated recommendation methods based on algorithmic models. These systems generate recommendation results and display them to end users based on user profiles, content popularity, and other dimensions. However, existing recommendation systems suffer from the following problems in actual operation:

[0003] 1. Adjusting recommendation strategies relies on code modifications, resulting in long cycles and high costs. Existing recommendation systems typically hard-coded their recommendation strategies into the algorithm logic, making them fixed once deployed. When operations personnel need to adjust recommendation strategies—such as changing the source of recommended content, modifying content filtering conditions, or altering the target audience—the development team must modify the algorithm code and redeploy the system. This adjustment cycle usually takes several days or even weeks, making it difficult to meet rapidly changing operational needs.

[0004] Second, the recommendation results lack flexible operational intervention methods. Although existing recommendation systems can automatically generate recommendation results, it is difficult for operators to make fine-tuning adjustments to the recommendations without interfering with the algorithm's operation. For example, when it is necessary to promote temporary content, handle content compliance, or conduct special topic operations, operators either cannot intervene in the recommendation results or can only rely entirely on manual editing of the recommendation slots. The latter editing results lack traceability records and are difficult to maintain and manage.

[0005] Third, there is a lack of traceability and auditing capabilities for recommendation changes. Existing recommendation systems lack a robust change tracking mechanism for adjustments to recommendation strategies and interventions in recommendation results. Some systems rely on database triggers to record changes, but triggers are strongly coupled with the storage layer, making it difficult to capture changes at the business semantic level; other systems use full log recording, generating a large amount of redundant data, making it difficult for operations personnel to effectively identify key changes and intervention trends.

[0006] Fourth, the content filtering granularity is too limited and lacks flexibility. Existing recommendation systems typically have only a single granularity for content filtering—either globally blocking content from being displayed in all recommendation scenarios, or blocking it only in a single recommendation library. This "one-size-fits-all" filtering approach cannot meet the differentiated filtering needs of operators in different scenarios. For example, operators may want to block certain content in a specific recommendation slot, but still recommend that content normally in other recommendation slots.

[0007] Fifth, the timing of recommendation strategies is fixed and lacks flexibility. Existing recommendation systems typically embed strategy rules into a fixed stage of the recommendation pipeline, either performing filtering during the candidate pool construction phase or performing post-filtering during the result output phase. The timing of strategy implementation cannot be flexibly adjusted according to business scenarios.

[0008] In summary, existing technologies lack a unified approach that integrates strategy configuration, data orchestration, outcome intervention, and behavior tracking. This makes it difficult for recommendation systems to maintain the automated recommendation capabilities of algorithms while providing operators with flexible strategy configuration, outcome intervention, and traceability auditing capabilities. Summary of the Invention

[0009] To address the aforementioned technical problems, the present invention aims to provide a multi-source media asset configuration-driven recommendation strategy generation and multi-level intervention method. This method solves the problems of recommendation strategy adjustment relying on code modification, long cycles, and high costs; the lack of flexible operational intervention methods for recommendation results; the lack of traceability and auditing capabilities for recommendation changes; the single granularity of content filtering and lack of flexibility; and the fixed timing of recommendation strategy action and inflexible configuration.

[0010] The present invention solves the above problems through the following technical solution:

[0011] A method for generating and intervening in a multi-source media asset configuration-driven recommendation strategy includes the following steps:

[0012] Step S100: Configuration of multi-dimensional combined recommendation strategy and dual-mode activation mechanism;

[0013] A multi-dimensional combined recommendation strategy configuration model is constructed, defining recommendation strategies through structured parameter combinations. The configuration dimensions of the model include at least content source dimension, content category dimension, target audience dimension, and combination logic dimension. The system automatically generates query condition chains based on the configuration parameters, performs cross-dimensional cross-filtering on the content knowledge base, and dynamically generates a candidate recommendation pool. The configuration model includes a "dual-mode strategy activation mechanism," where each recommendation strategy independently selects its activation timing in the recommendation pipeline. The dual modes include an "activation during recommendation" mode and an "activation after recommendation completion" mode. The candidate recommendation pool is output to step S200.

[0014] Step S200: Unified orchestration of the two-layer recommendation library architecture and multi-source heterogeneous data;

[0015] A two-layer recommendation library architecture is constructed, consisting of a default recommendation layer and a custom recommendation layer. The default recommendation layer includes non-personalized ranking recommendations based on an algorithm model and a personalized recommendation library based on user profiles. The custom recommendation layer autonomously creates a recommendation library through the recommendation strategy configuration model described in step S100, automatically pulling corresponding media asset content from external data sources and populating the recommendation library. The media asset content of the two recommendation libraries is uniformly mapped to a standardized recommendation basic information structure, providing a consistent data access protocol for the upper-layer business interface. The system uniformly arranges the candidate recommendation pool output in step S100 and the algorithm recommendation results of the default recommendation layer into a standardized recommendation library and outputs it to step S300.

[0016] Step S300: Multi-level refined intervention and full life cycle status tracking mechanism;

[0017] A multi-level intervention mechanism is constructed at the recommendation result level, including location-level content intervention, time-series intervention window management, intervention change difference tracking based on serialization comparison, and a two-layer granular blacklist filtering mechanism. Through intervention change difference tracking based on serialization comparison, the system automatically records the intervention change history and accumulates the number of interventions during each intervention operation, generating intervention tracking data. The intervention tracking data is fed back to step S100 as the basis for strategy configuration optimization.

[0018] As a further improvement of the present invention, in step S100, the "effective during recommendation" mode specifically means that when the system constructs the database query conditions of the content knowledge base, it dynamically concatenates the filtering parameters of each dimension in the strategy configuration into a database query condition chain, and determines the connection method between each dimension condition according to the combination logic parameters in the strategy configuration: if configured as "intersection" mode, the filtering conditions are connected by a logical AND, and the candidate content must simultaneously satisfy all the configured filtering conditions; if configured as "union" mode, a logical OR connection is used, and the candidate content only needs to satisfy any one of the configured filtering conditions; the query conditions generated thereby directly affect the data retrieval process of the content knowledge base, narrowing the candidate set from the input end.

[0019] As a further improvement of the present invention, in step S100, the "recommendation takes effect after completion" mode specifically means: the recommendation algorithm first completes the full recommendation calculation independently of the strategy configuration and outputs a complete recommendation result list without strategy constraints; then, the system reads the filtering parameters in the strategy configuration and performs a second filtering operation on the generated recommendation result list, retaining the content that meets the strategy conditions or supplementing the content that meets the conditions but was not hit by the algorithm from the candidate pool; finally, the recommendation result list after strategy adjustment is output.

[0020] As a further improvement of the present invention, in step S100, the selection of dual modes is achieved in the following way: In the visual management interface of the recommendation strategy configuration, the timing parameter of each strategy is specified through the "Effect Mode" option. After the system receives the configuration request, it persists the timing parameter of the strategy along with other dimension parameters of the strategy to the strategy configuration database. When the recommendation engine executes the recommendation task, it first reads the timing parameter of the strategy from the strategy configuration database and determines the execution branch of the current strategy in the recommendation pipeline based on its value.

[0021] As a further improvement of the present invention, in step S200, the default implementation of the recommendation layer is as follows: The system obtains recommendation results from two types of data sources respectively: For non-personalized ranking recommendations, the system performs an aggregation query on the content ranking data set and obtains the popularity ranking results under each category by grouping by content category; For personalized recommendation libraries, the system maintains a predefined enumeration structure, which pre-sets personalized recommendation libraries covering multiple content sources under each content category. Each personalized recommendation library corresponds to a data column in HBase. The system searches for the matching recommendation library from the enumeration structure according to the request parameters and reads the corresponding user profile recommendation data from HBase; After obtaining the two types of recommendation results respectively, the system uniformly maps them to a standardized recommendation basic information structure and superimposes the enabled / disabled status information of the recommendation library for final filtering.

[0022] As a further improvement of the present invention, the implementation of the custom recommendation layer in step S200 is as follows: After submitting a recommendation library configuration request through the configuration interface, the system first creates a recommendation library configuration record and marks its data generation status as "in progress," while deleting the existing old version generated content under this configuration; then, according to the content source identifier and content provider parameters specified in the configuration, the system asynchronously pulls matching media asset content from the external data source through the corresponding database access instance and populates it into the content set of the recommendation library, while recording the update timestamp of this configuration; when the external asynchronous process completes the content generation, the system automatically polls and checks all configurations in the "in progress" state each time it queries the recommendation library list. If the corresponding configuration already exists in the recommendation library and the configuration update timestamp recorded in the content record is consistent with the update timestamp of the current configuration, the status of the configuration is automatically updated to "completed."

[0023] As a further improvement of the present invention, in step S200, the unified strategy for the two types of recommendation results, namely the default recommendation layer and the custom recommendation layer, is as follows: When the business request specifies the default recommendation layer, the system determines the acquisition range according to the type parameter in the request; if no type is specified, the results of non-personalized list recommendations and personalized user profile recommendations are acquired simultaneously and output together; if a non-personalized type is specified, only the list recommendation results are acquired; if a personalized type is specified, only the user profile recommendation results are acquired; when the business request specifies the custom recommendation layer, the system directly queries the content from the custom recommendation library that has been generated with completed data according to the filtering conditions; and the output formats of the two paths are completely consistent.

[0024] As a further improvement of the present invention, the specific implementation of location-level content intervention in step S300 is as follows: Operators submit intervention requests through the intervention editing interface. The request contains multiple intervention records, each containing a location index, content identifier, content provider, intervention type, and effective time window. The system converts the intervention request into an intervention entity object and persists it to the database. When querying recommendation results, the system reads the currently effective intervention records and performs corresponding operations on the recommendation result list: For the "add" mode, the system inserts the content specified by the operator at the specified location index, and the original position and subsequent content are automatically shifted forward; for the "replace" mode, the system replaces the original algorithm-recommended content at the specified location index with the content specified by the operator, and the length of the recommendation list remains unchanged; and / or the specific implementation of time-sequential intervention window management is as follows: Each intervention record is configured with an independent effective time window, including start and end times; the system automatically determines the intervention status as "pending effectiveness," "effective," or "ineffective" based on the current time; this intervention status determination is calculated in real time each time recommendation results are queried, without the need for external scheduled tasks or manual start / stop operations.

[0025] As a further improvement of the present invention, the specific implementation of the intervention change difference tracking based on serialization comparison in step S300 is as follows: Each time an operator submits an intervention editing request, the system compares the new version of the intervention list with the old version of the intervention list stored in the database item by item; during the comparison process, the system serializes all fields of each intervention item into a structured string and then performs a string equality comparison; if an intervention item in a certain position in the new version does not exist in the old version, or although it exists, any field has changed, then the intervention item is included in the change set; the system only writes the items in the change set into the intervention history table, and accumulates the number of times the content has been intervened based on the combination of content identifier, content provider, and source type as a unique key.

[0026] As a further improvement of the present invention, in step S300, the two-layer granular blacklist filtering mechanism includes: "Specified Recommendation Library Deletion": When the content is blocked only in a specific recommendation library and other recommendation libraries are not affected; the system adds the identifier of the current recommendation library to the recommendation library identifier list in the deletion record, and marks the deletion type as "Specified Library Deletion". When querying the recommendation results of a certain recommendation library, the system checks whether the recommendation library identifier list of the deletion record contains the identifier of the current recommendation library. Only if it matches will the content be included in the exclusion scope; "Global Deletion": The content is blocked in all recommendation libraries and is not associated with any specific recommendation library; the system directly creates a record with the deletion type of "Global Deletion" and is not associated with any recommendation library identifier. When querying the recommendation results of any recommendation library, the content will be excluded; during the query, the system obtains the blacklist uniformly through the merge query logic: all records of the "Global Deletion" type and records of the "Specified Library Deletion" type that match the identifier of the current recommendation library are merged into a unified exclusion set, and then the content in the set is removed in batches from the recommendation results.

[0027] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0028] (1) With this invention, operators can independently create differentiated recommendation libraries by combining multi-dimensional parameters through a visual interface, reducing the traditional cycle (usually several days) of relying on the development team to modify recommendation algorithms to minutes of configuration completion. The dual-mode activation mechanism further enables operators to choose the timing of strategy implementation according to business scenarios, significantly improving the flexibility and iteration efficiency of recommendation strategies.

[0029] (2) The two-layer architecture of this invention takes into account both the system's automated recommendation capabilities and the operational customization needs. Multi-source heterogeneous data (data from different content platforms and different storage systems) are managed in a standardized manner through a unified orchestration model, which reduces the complexity of front-end access. At the same time, it supports operators to expand the recommendation library as needed without affecting the system's default recommendation capabilities.

[0030] (3) This invention realizes the "interventional, traceable, time-sensitive and granular" management capabilities of the recommendation results, enabling operators to finely adjust the algorithm recommendation results to meet business operation needs (such as temporary promotion, content compliance, special topic operation, etc.), while the complete intervention tracking record provides data support for operational decision-making.

[0031] (4) The multi-dimensional combined recommendation strategy configuration model in step S100 of this invention drives the system to automatically generate a candidate recommendation pool and outputs it to step S200; the two-layer architecture in step S200 unifies the algorithm recommendation results and operation configuration results into a standardized recommendation library and outputs it to step S300; the multi-level intervention mechanism in step S300 refines the unified recommendation results and generates intervention tracking data; the intervention tracking data, such as the number of times the content was intervened and the history of intervention changes, feeds back to step S100 to provide data support for operators to optimize strategy configuration. This closed-loop method of "configuration-driven generation, unified arrangement, refined intervention, and tracking feedback" enables the recommendation system to have self-evolution capabilities. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating a multi-source media asset configuration-driven recommendation strategy generation and multi-level intervention method according to the present invention. Detailed Implementation

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

[0034] Example:

[0035] Combined with appendix Figure 1 As shown, a multi-source media asset configuration-driven recommendation strategy generation and multi-level intervention method includes the following steps:

[0036] S100, multi-dimensional combined recommendation strategy configuration and dual-mode activation mechanism;

[0037] A multi-dimensional combined recommendation strategy configuration model is constructed, defining recommendation strategies through structured parameter combinations. The configuration dimensions of the model include at least content source dimension, content category dimension, target audience dimension, and combination logic dimension. The system automatically generates query condition chains based on the configuration parameters, performs cross-dimensional cross-filtering on the content knowledge base, and dynamically generates a candidate recommendation pool. The configuration model includes a "dual-mode strategy activation mechanism," where each recommendation strategy independently selects its activation timing in the recommendation pipeline. The dual modes include an "activation during recommendation" mode and an "activation after recommendation completion" mode. The candidate recommendation pool is output to step S200.

[0038] The construction of a multi-dimensional combined recommendation strategy configuration model defines recommendation strategies through structured parameter combinations. This multi-dimensional combined recommendation strategy configuration model includes the following configuration dimensions:

[0039] • Content source dimension: Specify content resource providers (such as Tencent, iQiyi, Youku, Mango TV), supporting single or multiple source configurations;

[0040] • Content categorization dimensions: Set filtering criteria such as content type (TV series, movies, variety shows, animation, children's programs, documentaries, etc.), country / region of origin, year of release, and cast;

[0041] • Target audience dimension: Set age group groups and gender preference filtering conditions;

[0042] • Combinatorial logic dimension: The various filtering dimensions support switching between two combination modes, "intersection" and "union," to flexibly control the strictness of the recommendation pool.

[0043] The system automatically generates query condition chains based on configuration parameters, performs cross-dimensional cross-filtering on the content knowledge base, and dynamically generates a candidate recommendation pool, creating differentiated recommendation strategies without writing code or modifying algorithm logic.

[0044] Dual-mode strategy activation mechanism:

[0045] The aforementioned multi-dimensional combined recommendation strategy configuration model includes a key design that distinguishes it from existing technologies: a "dual-mode strategy activation mechanism." Each recommendation strategy can independently choose when to take effect in the recommendation pipeline.

[0046] "Effective during recommendation process" mode: The strategy performs filtering during the candidate recommendation pool construction phase. Only content that meets the strategy conditions will enter the subsequent ranking and recommendation calculations, narrowing down the candidate set from the source. In this mode, the strategy configuration directly affects the algorithm's input space, which is equivalent to the operations personnel "defining the algorithm's input boundaries" without touching the algorithm logic.

[0047] Specifically, when the strategy configuration is set to "effective during recommendation," the system dynamically concatenates the various filtering parameters (content source, content type, country / region, release date range, actor set, target audience, etc.) in the strategy configuration into a database query condition chain when constructing the database query conditions for the content knowledge base. The connection method between the various conditions is determined based on the combinational logic parameters in the strategy configuration. If configured as "intersection," the filtering conditions are joined using a logical AND, requiring candidate content to simultaneously satisfy all configured filtering conditions. If configured as "union," a logical OR is used, requiring candidate content to satisfy only one configured filtering condition. The resulting query conditions directly impact the data retrieval process of the content knowledge base, narrowing the candidate set from the input. Only content meeting the conditions is included in the subsequent sorting process. Essentially, this implementation directly transforms the strategy configuration parameters into database query semantics, integrating the strategy rules with the data retrieval process.

[0048] "Recommendation Takes Effect After Completion" Mode: The strategy performs a secondary filtering after the algorithm's recommendation ranking is complete. The complete recommendation results output by the algorithm are generated first, and then the strategy rules perform post-filtering or supplementation on the results. This mode retains the algorithm's complete recommendation capabilities while giving operators control over the final output.

[0049] Specifically, when the strategy configuration is specified as "take effect after recommendation completion," the recommendation algorithm first performs a full recommendation calculation independently of the strategy configuration, outputting a complete list of recommendation results without strategy constraints. Then, the system reads the filtering parameters in the strategy configuration and performs a second filtering operation on the generated recommendation result list: filtering the results based on dimensions such as content source, content category, and target audience configured in the strategy, retaining content that meets the strategy conditions, or supplementing the candidate pool with content that meets the conditions but was not selected by the algorithm. Finally, the system outputs a recommendation result list adjusted by the strategy. The essence of this implementation is to use the strategy configuration as a post-filter at the algorithm's output, ensuring that the algorithm's complete recommendation capability remains unaffected, allowing operators to control the final displayed content without interfering with the algorithm's operation.

[0050] The dual-mode selection is achieved as follows: In the visual management interface for recommendation strategy configuration, operations personnel specify the timing parameter for each strategy's activation ("Effective during recommendation" or "Effective after recommendation") via the "Effective Mode" option. Upon receiving the configuration request, the system persistently stores this parameter, along with other strategy dimension parameters (content source, content category, target audience, combination logic, etc.), in the strategy configuration database. When executing recommendation tasks, the recommendation engine first reads this parameter from the strategy configuration database and determines the execution branch of the current strategy in the recommendation pipeline based on its value—if it is "Effective during recommendation," it enters the filtering process of the query construction stage; if it is "Effective after recommendation," it enters the post-result filtering process. This process requires only one configuration from operations personnel; subsequently, the recommendation engine automatically selects the execution path based on the configuration parameters, eliminating the need for manual intervention.

[0051] The difference between this design and existing technologies is that existing recommendation systems typically embed policy rules into a fixed part of the recommendation pipeline, while this invention treats the timing of policy action as a configurable parameter, enabling the same policy configuration model to play a role at different stages of the recommendation pipeline, thus decoupling policy configuration from algorithm execution.

[0052] Technical benefits: Operations personnel can independently create differentiated recommendation libraries by combining multi-dimensional parameters through a visual interface, reducing the traditional cycle (usually several days) of relying on the development team to modify recommendation algorithms to minutes of configuration completion; the dual-mode activation mechanism further enables operations personnel to choose when to apply strategies according to business scenarios, significantly improving the flexibility and iteration efficiency of recommendation strategies.

[0053] S200, a two-layer recommendation library architecture, and unified orchestration of multi-source heterogeneous data;

[0054] A two-layer recommendation library architecture is constructed, consisting of a default recommendation layer and a custom recommendation layer. The default recommendation layer includes non-personalized ranking recommendations based on an algorithm model and a personalized recommendation library based on user profiles. The custom recommendation layer autonomously creates a recommendation library through the recommendation strategy configuration model described in step S100, automatically pulling corresponding media asset content from external data sources and populating the recommendation library. The media asset content of the two recommendation libraries is uniformly mapped to a standardized recommendation basic information structure, providing a consistent data access protocol for the upper-layer business interface. The system uniformly arranges the candidate recommendation pool output in step S100 and the algorithm recommendation results of the default recommendation layer into a standardized recommendation library and outputs it to step S300.

[0055] The two-layer recommendation library architecture consisting of a "default recommendation layer and a custom recommendation layer" is described below:

[0056] Default recommendation layer: includes two types of recommendation sources pre-set by the system - (a) non-personalized list recommendation based on algorithm model (automatically generated according to the popularity ranking of content categories), and (b) personalized recommendation library based on user profile (pre-set personalized recommendation results covering multiple content sources under each category, and user profile data is stored in HBase columnar database to support real-time read and write).

[0057] The default recommendation layer is implemented as follows: The system obtains recommendation results from two types of data sources. For non-personalized ranking recommendations, the system performs an aggregation query on the content ranking dataset, grouping the results by content category (TV series, movies, variety shows, animation, children's content, documentaries, etc.) to obtain the popularity ranking results for each category. For personalized recommendation libraries, the system maintains a predefined enumeration structure. This enumeration structure pre-sets personalized recommendation libraries covering multiple content sources (such as Tencent, iQiyi, Youku, and Mango TV) under various content categories. Each personalized recommendation library corresponds to a data column in HBase. The system searches for matching recommendation libraries from the enumeration structure based on request parameters and reads the corresponding user profile recommendation data from HBase. After obtaining the two types of recommendation results, the system maps them uniformly to a standardized recommendation basic information structure and overlays the enabled / disabled status information of the recommendation library (determined by records in the status table; those not recorded are assumed to be enabled) for final filtering.

[0058] Custom Recommendation Layer: Operations personnel independently create a recommendation library using the multi-dimensional combined recommendation measurement configuration model described in step S100. The system automatically pulls corresponding media asset content from external data sources and populates the recommendation library. The data source is heterogeneous with the default recommendation layer, but the output format is unified. The data generation process of the custom recommendation library is managed using an asynchronous state machine ("In Progress" → "Completed"). The generation status is automatically determined by comparing the configuration update timestamp with the content generation timestamp, avoiding manual monitoring of data preparation progress.

[0059] The specific implementation process of the custom recommendation layer is as follows: After the operations personnel submit a recommendation library configuration request through the configuration interface, the system first creates a recommendation library configuration record and marks its data generation status as "in progress," while deleting any existing old version generated content under this configuration. Subsequently, based on the content source identifier and content provider parameters specified in the configuration, the system asynchronously pulls matching media asset content from external data sources through the corresponding database access instance (the system maintains an independent database access instance for each content source to support isolated access to multi-source data) and populates it into the content set of the recommendation library, while recording the update timestamp of this configuration. When the external asynchronous process completes content generation, the system automatically polls and checks all configurations in the "in progress" state each time it queries the recommendation library list: if the corresponding configuration already exists in the recommendation library and the configuration update timestamp recorded in the content record is consistent with the update timestamp of the current configuration, it indicates that the content has been generated with the latest configuration version, and the system automatically updates the status of the configuration to "completed." This asynchronous state machine mechanism based on timestamp version matching eliminates the need for operations personnel to manually monitor the data preparation progress, as the system automatically completes the state transition.

[0060] The media assets of the two-layer recommendation library are uniformly mapped to a standardized recommendation basic information structure (including fields such as recommendation identifier, title, type, tags, source, update time, and status), providing a consistent data access protocol for the upper-layer business interface. At the same time, the recommendation library and the business display scenario are linked together with a single click, and the recommendation results are automatically sent to the corresponding front-end scenario (such as the homepage waterfall layout and channel page recommendation positions) through a real-time content injection interface.

[0061] The unified strategy for both types of recommendation results is as follows: When the business request specifies the default recommendation layer, the system determines the scope of retrieval based on the type parameter in the request—if no type is specified, both non-personalized ranking recommendations and personalized user profile recommendations are retrieved and merged for output; if a non-personalized type is specified, only ranking recommendation results are retrieved; if a personalized type is specified, only user profile recommendation results are retrieved. When the business request specifies a custom recommendation layer, the system directly queries content from the pre-generated custom recommendation library according to the filtering conditions. The output format of both paths is completely consistent, transparent to the upper-layer business interface, and the front-end does not need to be aware of the specific source of the recommendation data.

[0062] Technical benefits: The two-tier architecture balances the system's automated recommendation capabilities with operational customization needs. Multi-source heterogeneous data (data from different content platforms and storage systems) is managed in a standardized manner through a unified orchestration model, reducing the complexity of front-end access. At the same time, it allows operators to expand the recommendation library as needed without affecting the system's default recommendation capabilities.

[0063] S300, multi-level refined intervention and full life cycle status tracking mechanism;

[0064] A multi-level intervention mechanism is constructed at the recommendation result level, including location-level content intervention, time-series intervention window management, intervention change difference tracking based on serialization comparison, and a two-layer granular blacklist filtering mechanism. Through intervention change difference tracking based on serialization comparison, the system automatically records the intervention change history and accumulates the number of interventions during each intervention operation, generating intervention tracking data. The intervention tracking data is fed back to step S100 as the basis for strategy configuration optimization.

[0065] The multi-level intervention mechanism constructed at the recommendation result level specifically includes the following four sub-mechanisms:

[0066] (a) Location-level content intervention;

[0067] At a specified location in the recommendation results list, two intervention operation modes are supported: "Add" mode (insert specified content at a specified location) and "Replace" mode (replace the algorithm recommendation results with specified content). The intervention content includes the specified content ID, source, and intervention time window.

[0068] The specific implementation is as follows: Operations personnel submit intervention requests through the intervention editing interface. Each request contains multiple intervention records, each including a location index (specifying the location of the intervention), a content identifier (specifying the media asset content), the content provider, the intervention type, and an effective time window. The system converts the intervention request into an intervention entity object and persists it to the database. When querying recommendation results, the system reads the currently effective intervention record (determining whether it is within the intervention time window based on the current time) and performs corresponding operations on the recommendation result list: For the "Add" mode, the system inserts the content specified by the operations personnel at the specified location index, automatically shifting the original position and subsequent content forward, and increasing the recommendation list length by one; for the "Replace" mode, the system directly replaces the original algorithm-recommended content at the specified location index with the content specified by the operations personnel, keeping the recommendation list length unchanged. Each intervention record contains an independent effective time window (start and end time). The system calculates the intervention status (pending effectiveness, in effect, expired) in real time during each query, achieving automated lifecycle management of the intervention.

[0069] (ii) Time-sequential intervention window management;

[0070] The specific implementation method is as follows: Each intervention record is configured with an independent effective time window, including the start and end times. The system automatically determines the intervention status as "pending effectiveness," "effective," or "ineffective" based on the current time. This intervention status determination is calculated in real time each time recommendation results are queried, without the need for external scheduled tasks or manual start / stop operations, thus achieving automated lifecycle management of interventions.

[0071] (iii) Tracking of intervention changes based on serialization alignment;

[0072] By tracking intervention change differences based on serialization comparison, the system automatically performs a structured comparison between the newly submitted intervention list and the historical version for each intervention operation. Only intervention items that have undergone substantial changes are recorded in the intervention history table, and the number of times that content has been intervened is accumulated. This allows the intervention tracking data to be fed back to step S100, providing data support for strategy configuration optimization.

[0073] The feedback mechanism is implemented as follows: Each time an operations staff member submits an intervention editing request, the system compares the new intervention list with the old intervention list stored in the database, item by item. During the comparison, the system serializes all fields of each intervention entry (including location index, content identifier, content provider, intervention type, remarks, time window, etc.) into a structured string and then performs a string equality comparison. If an intervention entry in a certain position in the new version does not exist in the old version (i.e., a newly added intervention), or although it exists, any field has changed (i.e., a modified intervention), then the entry is included in the change set. The system only writes the entries in the change set into the intervention history table, and increments the "number of times" counter for that content based on the combination of content identifier, content provider, and source type as a unique key (if the content has no history, a new record is created and initialized to one). The operations team can query the intervention history table to obtain the cumulative number of times each media asset has been intervened, the specific changes made during each intervention, and the timeline. This allows them to identify which content is frequently intervened, the specific reasons for the intervention, and the trends in change. This provides data support for subsequent optimization of recommendation strategy configuration. For example, if a certain type of content is frequently replaced, it indicates that the algorithm's recommendation effect for that type of content is not good, and the corresponding filtering conditions in the strategy configuration can be adjusted. If certain content has not been intervened for a long time, it indicates that the algorithm's recommendation effect is good and can be used as a positive reference. This forms a complete feedback loop from "intervention behavior records" to "strategy optimization decisions."

[0074] The difference between this design and existing technologies lies in the fact that change tracking in existing recommendation systems typically relies on database triggers or full log recording. The former is tightly coupled with the storage layer and struggles to capture changes at the business semantic level, while the latter generates a large amount of redundant data. This invention employs an application-layer serialization comparison scheme, treating the intervention list as a whole for structured comparison. This enables precise identification of "what changed in which piece of content at which location," while avoiding the problems of storage layer coupling and log redundancy. The cumulative number of interventions also provides the operations team with a quantitative indicator of the level of attention given to the content.

[0075] (iv) Two-layer granular blacklist filtering mechanism;

[0076] For content that needs to be removed from the recommendation results, two levels of deletion granularity are supported:

[0077] "Delete from Specific Recommendation Library": This option only blocks the content in a specific recommendation library; other recommendation libraries remain unaffected. In practice, when an operator selects the "Delete from Specific Recommendation Library" mode, the system adds the identifier of the current recommendation library to the recommendation library identifier list in the deletion record (using an append method, allowing for the addition of multiple recommendation libraries), and marks the deletion type as "Delete from Specific Library." When querying recommendation results for a specific recommendation library, the system checks whether the recommendation library identifier list in the deletion record contains the identifier of the current recommendation library; only if a match is found will the content be excluded.

[0078] "Global Deletion": This content is hidden across all recommendation libraries and is not associated with any specific recommendation library. In practice, the system directly creates a record with the deletion type "Global Deletion," without associating it with any recommendation library identifier. This content will be excluded when querying recommendation results from any recommendation library.

[0079] During a query, the system uses a unified blacklist obtained through a merge query logic of "delete from a specified library" or "delete globally": all records of the "delete globally" type are merged with records of the "delete from a specified library" type that match the current recommendation library identifier into a unified exclusion set, and then the content in this set is removed in batches from the recommendation results. The non-obviousness of this design lies in the fact that existing recommendation systems typically only have a single granularity for content filtering (either global blocking or single-library blocking). This invention stores the two granularity filtering rules independently and dynamically combines them through a merge query during the query process, allowing operators to flexibly choose the filtering granularity in different scenarios and avoid the side effects of a "one-size-fits-all" content blocking.

[0080] Technical effects: It has achieved the management capabilities of "intervention, traceability, timeliness, and granularity" of recommendation results, enabling operators to fine-tune the algorithm recommendation results to meet business operation needs (such as temporary promotion, content compliance, special topic operation, etc.), while the complete intervention tracking record provides data support for operational decision-making.

[0081] This invention, through the above three steps, forms a complete closed loop of data flow and feedback: Step S100's multi-dimensional combined recommendation strategy configuration model drives the system to automatically generate a candidate recommendation pool, which is output to Step S200; Step S200's two-layer architecture unifies the algorithm recommendation results and operational configuration results into a standardized recommendation library, which is output to Step S300; Step S300's multi-level intervention mechanism refines the unified recommendation results and generates intervention tracking data; Intervention tracking data, such as the number of times content was intervened and the history of intervention changes, feeds back to Step S100, providing data support for operators to optimize strategy configuration. This closed-loop method of "configuration-driven generation, unified arrangement, refined intervention, and tracking feedback" enables the recommendation system to have self-evolution capabilities, a technical effect that existing recommendation technologies, whether purely algorithmic or purely manual, do not possess.

[0082] Although the present invention has been described herein with reference to illustrative embodiments, the above embodiments are merely preferred embodiments of the present invention, and the implementation of the present invention is not limited to the above embodiments. It should be understood that those skilled in the art can devise many other modifications and implementations, which will fall within the scope and spirit of the principles disclosed in this application.

Claims

1. A method for generating and intervening in multi-level media asset allocation-driven recommendation strategies, characterized in that, Includes the following steps: Step S100: Configuration of multi-dimensional combined recommendation strategy and dual-mode activation mechanism; A multi-dimensional combined recommendation strategy configuration model is constructed, defining recommendation strategies through structured parameter combinations. The configuration dimensions of the model include at least content source dimension, content category dimension, target audience dimension, and combination logic dimension. The system automatically generates query condition chains based on the configuration parameters, performs cross-dimensional cross-filtering on the content knowledge base, and dynamically generates a candidate recommendation pool. The configuration model includes a "dual-mode strategy activation mechanism," where each recommendation strategy independently selects its activation timing in the recommendation pipeline. The dual modes include an "activation during recommendation" mode and an "activation after recommendation completion" mode. The candidate recommendation pool is output to step S200. Step S200: Unified orchestration of the two-layer recommendation library architecture and multi-source heterogeneous data; A two-layer recommendation library architecture is constructed, consisting of a default recommendation layer and a custom recommendation layer; the default recommendation layer includes non-personalized leaderboard recommendations based on algorithm models and a personalized recommendation library based on user profiles; The custom recommendation layer autonomously creates a recommendation library through the recommendation strategy configuration model described in step S100, and automatically pulls corresponding media asset content from external data sources and populates it into the recommendation library; The media asset content of the two-layer recommendation library is uniformly mapped to a standardized recommendation basic information structure, providing a consistent data access protocol for the upper-layer business interface; the system will uniformly arrange the candidate recommendation pool output in step S100 and the algorithm recommendation results of the default recommendation layer into a standardized recommendation library and output it to step S300. Step S300: Multi-level refined intervention and full life cycle status tracking mechanism; A multi-level intervention mechanism is constructed at the recommendation result level, including location-level content intervention, time-series intervention window management, intervention change difference tracking based on serialization comparison, and a two-layer granular blacklist filtering mechanism. Through intervention change difference tracking based on serialization comparison, the system automatically records the intervention change history and accumulates the number of interventions during each intervention operation, generating intervention tracking data. The intervention tracking data is fed back to step S100 as the basis for strategy configuration optimization.

2. The method for generating and intervening in a multi-source media asset configuration-driven recommendation strategy according to claim 1, characterized in that, In step S100, the "effective during recommendation" mode specifically means: When constructing database query conditions for the content knowledge base, the system dynamically concatenates the filtering parameters of each dimension in the strategy configuration into a chain of database query conditions, and determines the connection method between the conditions of each dimension according to the combinational logic parameters in the strategy configuration: if configured as "intersection" mode, the filtering conditions are connected by a logical AND, and the candidate content must satisfy all the configured filtering conditions at the same time; if configured as "union" mode, a logical OR connection is used, and the candidate content only needs to satisfy any one of the configured filtering conditions; the query conditions generated in this way directly affect the data retrieval process of the content knowledge base, narrowing the candidate set from the input end.

3. The method for generating and intervening in a multi-source media asset allocation-driven recommendation strategy according to claim 2, characterized in that, In step S100, the "recommendation takes effect after completion" mode specifically means: The recommendation algorithm first performs full recommendation calculations independently of the policy configuration, outputting a complete list of recommendation results without policy constraints. Then, the system reads the filtering parameters in the policy configuration and performs a second filtering operation on the generated list of recommendation results, retaining content that meets the policy conditions or supplementing content that meets the conditions but was not selected by the algorithm from the candidate pool. The final output is a list of recommended results after strategy adjustments.

4. The method for generating and intervening in a multi-source media asset configuration-driven recommendation strategy according to claim 3, characterized in that, In step S100, the selection of dual modes is achieved in the following way: In the visual management interface for recommendation strategy configuration, the timing parameter for each strategy is specified through the "Effect Mode" option. After receiving the configuration request, the system persists the timing parameter along with other dimension parameters of the strategy to the strategy configuration database. When the recommendation engine executes a recommendation task, it first reads the timing parameter from the strategy configuration database and determines the execution branch of the current strategy in the recommendation pipeline based on its value.

5. The method for generating and intervening in a multi-source media asset configuration-driven recommendation strategy according to claim 1, characterized in that, In step S200, the default implementation of the recommendation layer is as follows: The system retrieves recommendation results from two types of data sources: For non-personalized ranking recommendations, the system performs an aggregation query on the content ranking dataset, grouping the results by content category to obtain the popularity ranking results for each category; for personalized recommendation libraries, the system maintains a predefined enumeration structure, which pre-sets personalized recommendation libraries covering multiple content sources under each content category. Each personalized recommendation library corresponds to a data column in HBase. The system searches for matching recommendation libraries from the enumeration structure based on request parameters and reads the corresponding user profile recommendation data from HBase. After retrieving the two types of recommendation results, the system maps them uniformly to a standardized recommendation basic information structure and overlays the enabled / disabled status information of the recommendation library for final filtering.

6. The method for generating and intervening in a multi-source media asset configuration-driven recommendation strategy according to claim 5, characterized in that, In step S200, the custom recommendation layer is implemented as follows: After submitting a recommendation library configuration request through the configuration interface, the system first creates a recommendation library configuration record and marks its data generation status as "in progress," while deleting the existing old version of the generated content under this configuration. Subsequently, the system asynchronously retrieves matching media asset content from the external data source through the corresponding database access instance based on the content source identifier and content provider parameters specified in the configuration, and populates the content set of the recommendation library, while recording the update timestamp of this configuration. When the external asynchronous process completes the content generation, the system automatically polls and checks all configurations in the "in progress" state each time it queries the recommendation library list. If the corresponding configuration already exists in the recommendation library and the configuration update timestamp recorded in the content record is consistent with the update timestamp of the current configuration, the status of the configuration is automatically updated to "completed".

7. The method for generating and intervening in a multi-source media asset configuration-driven recommendation strategy according to claim 6, characterized in that, In step S200, the unified strategy for the two types of recommendation results, the default recommendation layer and the custom recommendation layer, is as follows: When the business request specifies the default recommendation layer, the system determines the scope of retrieval based on the type parameter in the request; if no type is specified, the system retrieves the results of both non-personalized ranking recommendations and personalized user profile recommendations and merges them for output; if a non-personalized type is specified, only the ranking recommendation results are retrieved. If specified as a personalized type, only user profile recommendation results will be obtained; When the business request specifies a custom recommendation layer, the system directly queries the content from the custom recommendation library that has been generated with completed data according to the filtering conditions. Furthermore, the output formats of the two paths are completely identical.

8. The method for generating and intervening in a multi-source media asset configuration-driven recommendation strategy according to claim 1, characterized in that, In step S300, the specific implementation method of location-level content intervention is as follows: Operations personnel submit intervention requests through the intervention editing interface. Each request contains multiple intervention records, including a location index, content identifier, content provider, intervention type, and effective time window. The system converts the intervention request into an intervention entity object and persists it to the database. When querying recommendation results, the system reads the currently effective intervention records and performs corresponding operations on the recommendation result list: For the "add" mode, the system inserts the content specified by the operations personnel at the specified location index, and the original content at the specified location and subsequent content are automatically shifted to the right; for the "replace" mode, the system replaces the original algorithm-recommended content at the specified location index with the content specified by the operations personnel, while the length of the recommendation list remains unchanged. The specific implementation methods of and / or time-sequential intervention window management are as follows: Each intervention record is configured with an independent effective time window, including the start and end times. The system automatically determines the intervention status as "pending effectiveness", "effective", or "ineffective" based on the current time. This intervention status determination is calculated in real time each time the recommendation results are queried, without the need for external scheduled tasks or manual start / stop operations.

9. The method for generating and intervening in a multi-source media asset configuration-driven recommendation strategy according to claim 8, characterized in that, In step S300, the specific implementation method of intervention change difference tracking based on serialization alignment is as follows: Each time an operations staff member submits an intervention editing request, the system compares the new version of the intervention list with the old version stored in the database, item by item. During the comparison, the system serializes all fields of each intervention entry into a structured string and then performs a string equality comparison. If an intervention entry in the new version does not exist in the old version, or if it exists but any field has changed, then the intervention entry is included in the change set. The system only writes the entries in the change set into the intervention history table, and accumulates the number of times the content has been intervened based on the combination of content identifier, content provider, and source type as a unique key.

10. The method for generating and intervening in a multi-source media asset configuration-driven recommendation strategy according to claim 9, characterized in that, In step S300, the two-layer granular blacklist filtering mechanism includes: "Delete from a specific recommendation library": This option only blocks the content in a specific recommendation library, while other recommendation libraries remain unaffected. The system adds the identifier of the current recommendation library to the recommendation library identifier list in the deletion record and marks the deletion type as "Delete from a specific library". When querying the recommendation results for a certain recommendation library, the system checks whether the recommendation library identifier list in the deletion record contains the identifier of the current recommendation library. Only if a match is found will the content be excluded. "Global Deletion": This content is hidden in all recommendation libraries and is not associated with any specific recommendation library; the system directly creates a record with the deletion type "Global Deletion", which is not associated with any recommendation library identifier. When querying recommendation results in any recommendation library, this content will be excluded. During the query, the system obtains the blacklist in a unified manner through the merge query logic: all records of the "global deletion" type and records of the "delete from specified library" type that match the current recommendation library identifier are merged into a unified exclusion set, and then the contents of this set are removed in batches from the recommendation results.

Images