Method and device for adjusting multi-task multi-scene recommendation model
By adjusting the strategy model and media asset scoring information, and combining the matching degree weight matrix between scenarios and strategies, a strategy knowledge base is constructed. This solves the problem of accurate operation of existing recommendation models under multiple scenarios and tasks, and achieves more accurate media asset recommendations and improved user experience.
Patent Information
- Application Number
- CN202510064227.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-15
- Publication Date
- 2025-12-30
AI Technical Summary
Existing recommendation models cannot meet the needs of precise operation when dealing with multiple scenarios and tasks. They suffer from modeling bias, negative transfer, and seesaw effect caused by differences in data volume. They also cannot effectively integrate the importance of features in different scenarios, resulting in inaccurate recommendations.
By acquiring the recommendation strength and media asset score information of the strategy model, and combining the matching degree weight matrix of the scene and strategy, the ranking results of the media assets are adjusted, a strategy knowledge base is constructed and the model is trained. The weight matrix is dynamically adjusted to reflect the matching degree of the scene and strategy, thereby improving the accuracy of the recommendation.
It alleviates the negative transfer and recommendation seesaw effect in multi-scenario, multi-task models, improves recommendation accuracy and user experience, and meets the needs of precision marketing.
Smart Images
Figure CN121233797A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of communication technology, and in particular to a method and device for adjusting a multi-task, multi-scenario recommendation model. Background Technology
[0002] Current recommendation models, when handling multiple scenarios and tasks, often model multi-scenario recommendation as a multi-task learning problem. They frequently employ gated hybrid expert network frameworks as the basis for model improvement to learn the commonalities and characteristics between scenarios, or they hierarchically process information from multiple scenarios and tasks to model information sharing and collaboration between scenarios and tasks. However, the amount of data can vary significantly between different scenarios and tasks, causing model training to favor scenarios or tasks with larger datasets, thus impairing the modeling performance of scenarios or tasks with smaller datasets. While sharing the underlying structure can achieve knowledge transfer, the gradient directions of the shared layer parameters may differ across scenarios and tasks, leading to negative transfer and a seesaw effect. Simultaneously handling multiple scenarios and tasks increases system complexity and maintenance costs, requiring sophisticated model design and training strategies to balance performance and resources.
[0003] Addressing the differences and commonalities across multiple tasks and scenarios is a real-world requirement in production operations, designed to meet diverse personalized recommendation needs. The research objectives for multi-task, multi-scenario approaches are twofold: first, to use a single model to serve all recommendation scenarios, achieving better results by jointly modeling data generated by users after consumption in any scenario, and accurately modeling the underlying interest representations of users' true interests; second, to use a single model to serve all scenarios, effectively reducing machine and labor costs, improving R&D efficiency, and promoting higher levels of collaborative technology development.
[0004] While significant progress has been made in multi-task, multi-scenario learning within recommender systems, capturing user interests often suffers from a seesaw effect and is hampered by sparse labels and dependencies between multiple tasks. Mixing features and unifying model training ignores the differences between features, failing to align the importance of different features across various scenarios. Existing recommender models, after feature processing, often provide inaccurate recommendations for users whose particular features are not prominent in specific scenarios, failing to meet the needs of precise operational optimization. Summary of the Invention
[0005] The purpose of this application is to provide a method and device for adjusting a multi-task, multi-scenario recommendation model, so as to solve the problem that existing recommendation models cannot meet the needs of accurate operation.
[0006] To address the aforementioned issues, this application provides a method for adjusting a multi-task, multi-scenario recommendation model. The multi-task, multi-scenario recommendation model is used to generate target media asset information recommended to users. The method includes:
[0007] The recommendation strength of the strategy model for multiple media assets is obtained. The strategy model is trained by the user's strategy features and general features.
[0008] Based on the recommendation strength of the strategy model for multiple media assets, the ranking results of each media asset in the target media asset information are adjusted.
[0009] The method further includes:
[0010] Obtain the scoring information of media assets in different scenarios, as well as the matching degree weight matrix between scenarios and strategies;
[0011] The step of adjusting the ranking of each media asset in the target media asset information based on the recommendation strength of multiple media assets includes:
[0012] Based on the media asset scores in different scenarios, the strategy model adjusts the ranking of each media asset in the target media asset information according to the recommendation strength of multiple media assets and the matching degree weight matrix.
[0013] Specifically, based on the media asset score information in different scenarios, the recommendation strength of the strategy model for multiple media assets, and the matching degree weight matrix, the ranking results of each media asset in the target media asset information are adjusted, including:
[0014] The recommendation strength of the strategy model for multiple media assets is multiplied by the score information of the media assets in different scenarios to obtain the user's recommendation score for multiple media assets in different scenarios.
[0015] The recommendation scores of users for multiple media assets in different scenarios are multiplied by the matching degree weight matrix to obtain the comprehensive scores of users for multiple media assets in different scenarios.
[0016] Based on the user's comprehensive score for multiple media assets in different scenarios, the sorting results of each media asset in the target media asset information are adjusted.
[0017] The acquisition of the matching weight matrix between the scenario and the strategy includes:
[0018] Based on the degree of influence of scene features on the strategy, a first weight is assigned to the scene features;
[0019] Based on the applicability of the strategy features in different scenarios, a second weight is assigned to the strategy features;
[0020] The first weights of multiple scene features are weighted and summed to obtain the comprehensive feature vector of the scene;
[0021] The second weights of multiple policy features are weighted and summed to obtain the comprehensive feature vector of the policy.
[0022] Calculate the similarity value between the scene and the strategy based on the comprehensive feature vector of the scene and the comprehensive feature vector of the strategy;
[0023] The similarity value between the scenario and the strategy is used as an element in the matching weight matrix; the larger the similarity value, the greater the matching degree between the scenario and the strategy.
[0024] The method further includes:
[0025] The matching degree weight matrix is adjusted based on at least one of the following: user feedback, changes in scene characteristics, and strategy adjustments.
[0026] Specifically, adjusting the matching degree weight matrix based on at least one of user feedback, changes in scene characteristics, and strategy adjustments includes:
[0027] Based on at least one of user feedback, changes in scene features, and policy adjustment information, adjust the first weight of the scene feature and / or the second weight of the policy feature;
[0028] The adjusted matching degree weight matrix is determined based on the adjusted first and second weights.
[0029] The method further includes:
[0030] Obtain user strategy characteristics and general characteristics;
[0031] Construct a strategy knowledge base based on the strategy features and the general features;
[0032] The strategy model is obtained by training the model based on the strategy knowledge base.
[0033] The step of training the model based on the strategy knowledge base to obtain the strategy model includes:
[0034] The requirements of the strategy features are vectorized to obtain a first vector;
[0035] The labels of the general features are vectorized to obtain the second vector;
[0036] The descriptive information of the general features is vectorized to obtain a third vector;
[0037] Calculate the attention weights of the first vector and the second vector, and multiply the attention weights by the third vector to obtain the target matrix;
[0038] The summation and average of the target matrix are calculated using a truncation function, and the general features with higher weights are truncated as the output of the attention mechanism.
[0039] The policy features and the general features output by the attention mechanism are used as inputs to train the model and determine the policy model;
[0040] The strategy model is used to output the recommendation strength for media assets.
[0041] This application embodiment also provides an adjustment device for a multi-task, multi-scenario recommendation model. The multi-task, multi-scenario recommendation model is used to generate target media asset information recommended to users. The device includes a processor and a transceiver. The transceiver receives and sends data under the control of the processor. The processor is used to perform the following operations:
[0042] The recommendation strength of the strategy model for multiple media assets is obtained. The strategy model is trained by the user's strategy features and general features.
[0043] Based on the recommendation strength of the strategy model for multiple media assets, the ranking results of each media asset in the target media asset information are adjusted.
[0044] The processor is also used to perform the following operations:
[0045] Obtain the scoring information of media assets in different scenarios, as well as the matching degree weight matrix between scenarios and strategies;
[0046] Based on the media asset scores in different scenarios, the strategy model adjusts the ranking of each media asset in the target media asset information according to the recommendation strength of multiple media assets and the matching degree weight matrix.
[0047] The processor is also used to perform the following operations:
[0048] The recommendation strength of the strategy model for multiple media assets is multiplied by the score information of the media assets in different scenarios to obtain the user's recommendation score for multiple media assets in different scenarios.
[0049] The recommendation scores of users for multiple media assets in different scenarios are multiplied by the matching degree weight matrix to obtain the comprehensive scores of users for multiple media assets in different scenarios.
[0050] Based on the user's comprehensive score for multiple media assets in different scenarios, the sorting results of each media asset in the target media asset information are adjusted.
[0051] The processor is also used to perform the following operations:
[0052] Based on the degree of influence of scene features on the strategy, a first weight is assigned to the scene features;
[0053] Based on the applicability of the strategy features in different scenarios, a second weight is assigned to the strategy features;
[0054] The first weights of multiple scene features are weighted and summed to obtain the comprehensive feature vector of the scene;
[0055] The second weights of multiple policy features are weighted and summed to obtain the comprehensive feature vector of the policy.
[0056] Calculate the similarity value between the scene and the strategy based on the comprehensive feature vector of the scene and the comprehensive feature vector of the strategy;
[0057] The similarity value between the scenario and the strategy is used as an element in the matching weight matrix; the larger the similarity value, the greater the matching degree between the scenario and the strategy.
[0058] The processor is also used to perform the following operations:
[0059] The matching degree weight matrix is adjusted based on at least one of the following: user feedback, changes in scene characteristics, and strategy adjustments.
[0060] The processor is also used to perform the following operations:
[0061] Based on at least one of user feedback, changes in scene features, and policy adjustment information, adjust the first weight of the scene feature and / or the second weight of the policy feature;
[0062] The adjusted matching degree weight matrix is determined based on the adjusted first and second weights.
[0063] The processor is also used to perform the following operations:
[0064] Obtain user strategy characteristics and general characteristics;
[0065] Construct a strategy knowledge base based on the strategy features and the general features;
[0066] The strategy model is obtained by training the model based on the strategy knowledge base.
[0067] The processor is also used to perform the following operations:
[0068] The requirements of the strategy features are vectorized to obtain a first vector;
[0069] The labels of the general features are vectorized to obtain the second vector;
[0070] The descriptive information of the general features is vectorized to obtain a third vector;
[0071] Calculate the attention weights of the first vector and the second vector, and multiply the attention weights by the third vector to obtain the target matrix;
[0072] The summation and average of the target matrix are calculated using a truncation function, and the general features with higher weights are truncated as the output of the attention mechanism.
[0073] The policy features and the general features output by the attention mechanism are used as inputs to train the model and determine the policy model;
[0074] The strategy model is used to output the recommendation strength for media assets.
[0075] This application embodiment also provides an adjustment device, including a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the adjustment method of the multi-task multi-scenario recommendation model as described above.
[0076] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in the adjustment method of the multi-task, multi-scenario recommendation model as described above.
[0077] This application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the adjustment method for the multi-task, multi-scenario recommendation model as described above.
[0078] The above-mentioned technical solution of this application has at least the following beneficial effects:
[0079] In the adjustment method and device of the multi-task multi-scenario recommendation model in this application embodiment, the recommendation strength of the strategy model for media poses is used as a feature to adjust the ranking results of each media asset in the target media asset information generated by the recommendation model. This makes the recommended media assets finally output by the recommendation model meet the requirements of accurate recommendation by the strategy, alleviates the negative transfer and recommendation seesaw situation of large multi-scenario multi-task models, improves the effect of accurate marketing, and thus improves the user experience. Attached Figure Description
[0080] Figure 1 This is a flowchart illustrating the steps of the adjustment method for the multi-task, multi-scenario recommendation model provided in the embodiments of this application;
[0081] Figure 2 This is a schematic diagram illustrating the principle of the adjustment method for the multi-task, multi-scenario recommendation model provided in the embodiments of this application;
[0082] Figure 3 This is a schematic diagram illustrating the structure of the adjustment device for the multi-task, multi-scenario recommendation model provided in this application embodiment. Detailed Implementation
[0083] To make the technical problems, technical solutions and advantages of this application clearer, a detailed description will be provided below in conjunction with the accompanying drawings and specific embodiments.
[0084] The multi-task, multi-scenario recommendation model can be called a large-scale recommendation model or a general-purpose recommendation model, without specific limitations here. This recommendation model generates target media asset information recommended to users based on at least one task and at least one scenario. The target media asset information includes multiple media assets, which are ranked according to the recommendation strength of the multiple media assets (which can also be understood as the user's preference for the media assets). The task refers to the indicator task that the target media asset information must meet, and the scenario refers to the recommendation scenario corresponding to the target media asset information, without specific limitations here.
[0085] like Figure 1 As shown in the figure, this application embodiment provides a method for adjusting a multi-task, multi-scenario recommendation model. The multi-task, multi-scenario recommendation model is used to generate target media asset information recommended to users. The method includes:
[0086] Step 101: Obtain the recommendation strength of the strategy model for multiple media assets. The strategy model is trained by the user's strategy features and general features.
[0087] Optionally, user strategy features are used to characterize user preference features under specific strategies, such as time-sharing preferences, IP preferences, subscription preferences, and family composition preferences. Strategy features are the most salient features generated by the current strategy requirements. General features are existing features required for training the recommendation model, such as content features.
[0088] Step 102: Adjust the ranking of each media asset in the target media asset information according to the recommendation strength of the strategy model for multiple media assets.
[0089] Optionally, the recommendation strength of the strategy model for multiple media assets can be represented as User Prefer, which is a vector related to the media asset.
[0090] This application embodiment uses the recommendation strength of the strategy model for media poses as a feature to adjust the ranking results of each media asset in the target media asset information generated by the recommendation model, so that the recommended media assets finally output by the recommendation model meet the requirements of accurate recommendation of the strategy, alleviate the negative transfer and recommendation seesaw situation of large models in multiple scenarios and tasks, improve the effect of accurate marketing, and thus improve the user experience.
[0091] In at least one embodiment of this application, the method further includes:
[0092] Obtain the scoring information of media assets in different scenarios (which can be represented as Res), and the matching weight matrix between scenarios and strategies (which can be represented as Scene Weight);
[0093] Accordingly, step 102 includes:
[0094] Based on the media asset scores in different scenarios, the strategy model adjusts the ranking of each media asset in the target media asset information according to the recommendation strength of multiple media assets and the matching degree weight matrix.
[0095] In one implementation, based on the media asset score information in different scenarios, the recommendation strength of the strategy model for multiple media assets, and the matching degree weight matrix, the ranking result of each media asset in the target media asset information is adjusted, including:
[0096] The recommendation strength of the strategy model for multiple media assets is multiplied by the score information of the media assets in different scenarios to obtain the user's recommendation score for multiple media assets in different scenarios.
[0097] The recommendation scores of users for multiple media assets in different scenarios are multiplied by the matching degree weight matrix to obtain the comprehensive scores of users for multiple media assets in different scenarios.
[0098] Based on the user's comprehensive score for multiple media assets in different scenarios, the sorting results of each media asset in the target media asset information are adjusted.
[0099] For example, considering the combined score information Res of media assets in different scenarios, the strategy model's recommendation strength User Prefer for multiple media assets, and the matching degree weight matrix Scene Weight, the ranking formula for each media asset in the target media asset information is as follows:
[0100] FT Prefer=Sort(Sum(User prefer*Res)*Scene Weight)
[0101] Here, Res represents the score matrix of media asset sequences in different scenarios, reflecting the score of each media asset in a specific scenario. The score can be calculated based on various factors, such as the popularity of the media asset, user reviews, and type relevance.
[0102] User Prefer can be understood as the recommendation strength of the strategy model for different media assets; it is a media asset-related vector. By multiplying User Prefer by Res, the recommendation of the strategy model can be combined with the media asset score in the scenario to obtain a recommendation score.
[0103] Scene Weight is a weight matrix representing the matching degree between scenes and strategies. Its purpose is to adjust the ranking results based on the degree of matching between different scenes and strategies. Scenes with a high degree of matching will have a larger weight, thus giving media assets in those scenes a greater advantage in the ranking.
[0104] In at least one embodiment of this application, obtaining the matching degree weight matrix between the scene and the strategy includes:
[0105] Based on the degree of influence of scene features on the strategy, a first weight is assigned to the scene features;
[0106] Based on the applicability of the strategy features in different scenarios, a second weight is assigned to the strategy features;
[0107] The first weights of multiple scene features are weighted and summed to obtain the comprehensive feature vector of the scene;
[0108] The second weights of multiple policy features are weighted and summed to obtain the comprehensive feature vector of the policy.
[0109] Calculate the similarity value between the scene and the strategy based on the comprehensive feature vector of the scene and the comprehensive feature vector of the strategy;
[0110] The similarity value between the scenario and the strategy is used as an element in the matching weight matrix; the larger the similarity value, the greater the matching degree between the scenario and the strategy.
[0111] For example, feature extraction can be performed on different scenarios. Scenario features can include, but are not limited to, the following: scenario type (e.g., movie, TV series, short video, etc.), user group characteristics (e.g., age, gender, region, etc.), time factors (e.g., daytime, evening, weekend, etc.), and device type (e.g., mobile phone, tablet, computer, etc.). For each scenario feature, different initial weights can be assigned based on its influence on the strategy. For example, in the short video scenario, if the user group is mainly young people, then some popular and fashionable strategies may have a higher matching degree, so features related to the young user group can be given higher weights.
[0112] For example, feature extraction is also performed on the strategy model. For each strategy feature, different secondary weights can be assigned based on its applicability in different scenarios. For instance, the time-segmented and group-based recommendation strategy may have a higher matching degree in scenarios where user group characteristics differ greatly (such as a family user combination of men, women, children, and four generations living together under the same set-top box), so features related to personalized recommendations can be given higher weights.
[0113] Once we have the comprehensive feature vectors of the scene and the comprehensive feature vectors of the strategy, we can use various methods to calculate their similarity. Commonly used methods include cosine similarity and Euclidean distance.
[0114] Taking cosine similarity as an example, let the scene feature vector be S = (s1, s2, ..., s...). n The strategy feature vector is T = (t1, t2, ..., t3).n If ), then the cosine similarity between them is:
[0115]
[0116] The calculated similarity value can be used as an element in the matching weight matrix. The higher the similarity, the greater the corresponding weight, indicating a higher degree of matching between the scenario and the strategy.
[0117] As time progresses and user behavior changes, the matching degree between scenarios and strategies may also change. Therefore, it is necessary to update and adjust the matching degree weight matrix regularly. Accordingly, the method also includes:
[0118] The matching degree weight matrix is adjusted based on at least one of the following: user feedback, changes in scene characteristics, and strategy adjustments.
[0119] For example, the weight matrix can be dynamically adjusted by monitoring user feedback (such as click-through rate, playback duration, ratings, etc.), changes in scene characteristics (such as updates to the user base, changes in time factors, etc.), and strategy adjustment information. For instance, if users provide positive feedback on a particular strategy in a certain scene, the matching weight between that scene and that strategy can be appropriately increased.
[0120] In one implementation, the matching degree weight matrix is adjusted based on at least one of user feedback, changes in scene features, and strategy adjustment information, including:
[0121] Based on at least one of user feedback, changes in scene features, and policy adjustment information, adjust the first weight of the scene feature and / or the second weight of the policy feature;
[0122] The adjusted matching degree weight matrix is determined based on the adjusted first and second weights.
[0123] By dynamically adjusting the matching degree weight matrix using the above methods, the matching degree of scenarios and strategies can be reflected more accurately, thereby achieving more personalized and efficient media asset sorting and improving user experience.
[0124] In at least one embodiment of this application, the method further includes:
[0125] Obtain user strategy characteristics and general characteristics;
[0126] Construct a strategy knowledge base based on the strategy features and the general features;
[0127] The strategy model is obtained by training the model based on the strategy knowledge base.
[0128] Optionally, personalized recommendation systems that cater to each individual user have higher requirements for application performance. To ensure that online metrics meet expectations, different strategies are adopted. These strategies vary depending on the business and scenario. Strategy tags are the salient features of the current strategy requirements, and the corresponding business and scenario knowledge is called the strategy knowledge base.
[0129] like Figure 2 As shown, in order to improve the performance of recommendation algorithms in multiple scenarios and tasks, reduce recommendations that are irrelevant to user interests, and meet the needs of precise operation, this embodiment first generates precise user strategy tags (i.e., strategy features), then constructs a user strategy knowledge base, performs precise modeling of user strategies (i.e., strategy model modeling), and finally adjusts the results of the multi-scenario, multi-task recommendation model through the features output by the strategy model, so that the recommendation results are more in line with operational expectations and improve the user experience of the large-screen recommendation system.
[0130] For example, strategy features are user preference features generated under specific strategies based on operational needs, where discrete features are represented as [U id ,{T i :S i}]; the association feature is represented as [U id ,{T xi ,T yi ,S xyi}]. Where U id For user identification, T i S represents the policy characteristics. i Represents the policy feature score, {T xi ,T yi ,S xyi The subscript indicates the relationship between tags.
[0131] For example, there are strategy features such as time-sharing preference, IP preference, subscription preference, and family composition preference; among them, the time-sharing preference feature is represented as [U id hour i ,channel i score i User IP preference characteristics are represented as [U] id ,channel i ,{ip1,...,ip i ,...,p n The time / channel or channel / IP characteristics are related and thus considered related characteristics; users subscribe to media assets [U id ,{tag1:s1,..,tag i :s i ,...,tag n :s n}], family composition preference characteristics are represented as [U id [{kids:score1,old:score2,young:score3,middle:score4}] contains multiple features that are not related to each other and are considered discrete features. The goal is to expand or merge these features to make their meaning unique, merging related features and splitting discrete features, using the scores as feature values, and completing the data cleaning and transformation process.
[0132] Strategy features are the most prominent characteristics of current strategy requirement production. General features already have fixed production patterns and are rich in types. Obtaining relevant features from general features can both expand current strategy features and effectively utilize existing features to reduce the pressure of feature production. Therefore, a strategy knowledge base is constructed based on a general feature library (including multiple general features) and strategy features. For example, the strategy feature label is denoted as Feature, and the detailed description of the strategy feature is denoted as Des. Based on these two types of information, the strategy requirements or problems are constructed. For example, in time-based preferences, it can be expressed as "I want to create a user channel preference strategy based on time periods".
[0133] Optionally, the policy model is obtained by training a model based on the policy knowledge base, including:
[0134] The requirements of the strategy features are vectorized to obtain the first vector Q;
[0135] The labels of the general features are vectorized to obtain the second vector K;
[0136] The descriptive information of the general features is vectorized to obtain the third vector D;
[0137] Calculate the attention weights of the first vector and the second vector, and multiply the attention weights by the third vector to obtain the target matrix;
[0138] The summation and average of the target matrix are calculated using a truncation function, and the general features with higher weights are truncated as the output of the attention mechanism.
[0139] The policy features and the general features output by the attention mechanism are used as inputs to train the model and determine the policy model;
[0140] The strategy model is used to output the recommendation strength for media assets.
[0141] Suppose there are n policy inputs, denoted as X = {x1, x2, ..., xn}. n}, each x i It is a d-dimensional word vector representation. Therefore, the dimension of the input matrix X is n×d.
[0142] The similarity calculation uses three types of data vectorization: strategy requirement vectorization denoted as Q, general feature label vectorization denoted as K, and detailed description of the general feature vectorization denoted as D. The dimensions of Q, K, and D are as follows:
[0143] Q∈R n*dk ;
[0144] K∈R n*dk ;
[0145] D∈R n*dk .
[0146] Calculate a series of weights for Q and the general feature K, then normalize them using softmax to obtain the attention weights. Multiply these attention weights by D. G represents the truncation function, which calculates the sum and average of the matrices to truncate the general features with higher weights. The final attention output is represented as:
[0147]
[0148] Using the aforementioned strategy features and the general features corresponding to the final attention output as model inputs, the model is trained to obtain the recommendation strength of the strategy model for media assets (which can also be referred to as the user's preference for media assets), expressed as follows:
[0149] User Prefer = [U id {album i :score i}];
[0150] Among them, User Prefer refers to users' preferences for media assets; U id User identifier; album i For media resources i; score i This represents the degree of user preference for media assets.
[0151] This application embodiment constructs a strategy knowledge base to enhance strategy features, automatically calculates important features under the current strategy and assigns weights; based on the strategy knowledge base, strategy modeling is performed, and the results of the recommendation model are fine-tuned using a collaborative approach of large and small models (strategy model + recommendation model) to meet the requirements of accurate strategy recommendation, alleviate the negative transfer of large models in multiple scenarios and tasks and the seesaw effect of recommendation, and improve the effect of precision marketing.
[0152] like Figure 3 As shown in the figure, this application embodiment provides an adjustment device for a multi-task, multi-scenario recommendation model. The multi-task, multi-scenario recommendation model is used to generate target media asset information recommended to users. The device includes a processor 300 and a transceiver 310. The transceiver 310 receives and sends data under the control of the processor 300. The processor 300 is used to perform the following operations:
[0153] The recommendation strength of the strategy model for multiple media assets is obtained. The strategy model is trained by the user's strategy features and general features.
[0154] Based on the recommendation strength of the strategy model for multiple media assets, the ranking results of each media asset in the target media asset information are adjusted.
[0155] As an optional embodiment, the processor is also configured to perform the following operations:
[0156] Obtain the scoring information of media assets in different scenarios, as well as the matching degree weight matrix between scenarios and strategies;
[0157] Based on the media asset scores in different scenarios, the strategy model adjusts the ranking of each media asset in the target media asset information according to the recommendation strength of multiple media assets and the matching degree weight matrix.
[0158] As an optional embodiment, the processor is also configured to perform the following operations:
[0159] The recommendation strength of the strategy model for multiple media assets is multiplied by the score information of the media assets in different scenarios to obtain the user's recommendation score for multiple media assets in different scenarios.
[0160] The recommendation scores of users for multiple media assets in different scenarios are multiplied by the matching degree weight matrix to obtain the comprehensive scores of users for multiple media assets in different scenarios.
[0161] Based on the user's comprehensive score for multiple media assets in different scenarios, the sorting results of each media asset in the target media asset information are adjusted.
[0162] As an optional embodiment, the processor is also configured to perform the following operations:
[0163] Based on the degree of influence of scene features on the strategy, a first weight is assigned to the scene features;
[0164] Based on the applicability of the strategy features in different scenarios, a second weight is assigned to the strategy features;
[0165] The first weights of multiple scene features are weighted and summed to obtain the comprehensive feature vector of the scene;
[0166] The second weights of multiple policy features are weighted and summed to obtain the comprehensive feature vector of the policy.
[0167] Calculate the similarity value between the scene and the strategy based on the comprehensive feature vector of the scene and the comprehensive feature vector of the strategy;
[0168] The similarity value between the scenario and the strategy is used as an element in the matching weight matrix; the larger the similarity value, the greater the matching degree between the scenario and the strategy.
[0169] As an optional embodiment, the processor is also configured to perform the following operations:
[0170] The matching degree weight matrix is adjusted based on at least one of the following: user feedback, changes in scene characteristics, and strategy adjustments.
[0171] As an optional embodiment, the processor is also configured to perform the following operations:
[0172] Based on at least one of user feedback, changes in scene features, and policy adjustment information, adjust the first weight of the scene feature and / or the second weight of the policy feature;
[0173] The adjusted matching degree weight matrix is determined based on the adjusted first and second weights.
[0174] As an optional embodiment, the processor is also configured to perform the following operations:
[0175] Obtain user strategy characteristics and general characteristics;
[0176] Construct a strategy knowledge base based on the strategy features and the general features;
[0177] The strategy model is obtained by training the model based on the strategy knowledge base.
[0178] As an optional embodiment, the processor is also configured to perform the following operations:
[0179] The requirements of the strategy features are vectorized to obtain a first vector;
[0180] The labels of the general features are vectorized to obtain the second vector;
[0181] The descriptive information of the general features is vectorized to obtain a third vector;
[0182] Calculate the attention weights of the first vector and the second vector, and multiply the attention weights by the third vector to obtain the target matrix;
[0183] The summation and average of the target matrix are calculated using a truncation function, and the general features with higher weights are truncated as the output of the attention mechanism.
[0184] The policy features and the general features output by the attention mechanism are used as inputs to train the model and determine the policy model;
[0185] The strategy model is used to output the recommendation strength for media assets.
[0186] In this embodiment, the recommendation strength of the strategy model for media poses is used as a feature to adjust the ranking of each media asset in the target media asset information generated by the recommendation model. This ensures that the recommended media assets finally output by the recommendation model meet the requirements of accurate recommendation by the strategy, alleviates the negative transfer and recommendation seesaw situation of large models in multiple scenarios and tasks, improves the effect of accurate marketing, and thus improves the user experience.
[0187] In other words, the embodiments of this application adopt a strategy model + recommendation big model scheme, construct a strategy knowledge base and design a reusable general feature calculation method, and use the strategy model to correct the recommendation results in multiple scenarios so that the final presented content is more in line with operational needs.
[0188] It should be noted that the adjustment device for the multi-task multi-scenario recommendation model provided in this application embodiment is a device capable of executing the above-described adjustment method for the multi-task multi-scenario recommendation model. Therefore, all embodiments of the above-described adjustment method for the multi-task multi-scenario recommendation model are applicable to this device and can achieve the same or similar beneficial effects, which will not be repeated here.
[0189] This application also provides an adjustment device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the various processes in the above-described multi-task multi-scenario recommendation model adjustment method embodiment and achieves the same technical effect. To avoid repetition, it will not be described again here.
[0190] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, this program implements the various processes described above in the embodiment of the adjustment method for the multi-task, multi-scenario recommendation model, and achieves the same technical effect. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.
[0191] This application also provides a computer program product, including computer instructions. When executed by a processor, these computer instructions implement the various processes of the above-described multi-task, multi-scenario recommendation model adjustment method embodiment and achieve the same technical effect. To avoid repetition, they will not be described again here.
[0192] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.
[0193] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 A device for one or more processes and / or the functions specified in one or more boxes.
[0194] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce a paper article including an instruction means, the instruction means being implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0195] These computer program instructions can also be loaded onto a computer or other programmable data processing equipment, causing the computer or other programmable equipment to perform a series of operational steps to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0196] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.
Claims
1. An adjustment method of a multi-task multi-scene recommendation model, the multi-task multi-scene recommendation model being used to generate target media information recommended to a user, characterized in that, The method comprises: obtaining the recommendation strength of the strategy model on the multiple media resources, the strategy model being trained by the user's strategy features and general features; adjusting the ranking results of each media resource in the target media resource information according to the recommendation strength of the strategy model on the multiple media resources.
2. The method of claim 1, wherein, The method further comprises: obtaining the score information of media resources under different scenarios, and the matching degree weight matrix of scenarios and strategies; The method further comprises: obtaining the score information of media resources under different scenarios, and the matching degree weight matrix of scenarios and strategies; 3. The method of claim 2, wherein, The method further comprises: multiplying the recommendation strength of the strategy model on the multiple media resources by the score information of media resources under different scenarios to obtain the recommendation scores of the multiple media resources under different scenarios; multiplying the recommendation scores of the multiple media resources under different scenarios by the matching degree weight matrix to obtain the comprehensive scores of the multiple media resources under different scenarios; adjusting the ranking results of each media resource in the target media resource information according to the comprehensive scores of the multiple media resources under different scenarios.
4. The method of claim 2, wherein, The method further comprises: assigning a first weight to the scenario features according to the influence degree of the scenario features on the strategies; assigning a second weight to the strategy features according to the applicability of the strategy features under different scenarios; weighting and summing the first weights of the multiple scenario features to obtain a comprehensive feature vector of the scenario; weighting and summing the second weights of the multiple strategy features to obtain a comprehensive feature vector of the strategy; calculating the similarity value of the scenario and the strategy according to the comprehensive feature vector of the scenario and the comprehensive feature vector of the strategy; taking the similarity value of the scenario and the strategy as an element in the matching degree weight matrix; the greater the similarity value, the greater the matching degree of the scenario and the strategy.
5. The method of claim 4, wherein, The method further comprises: adjusting the matching degree weight matrix according to at least one of the feedback information of the user, the change information of the scenario features, and the adjustment information of the strategy.
6. The method of claim 5, wherein, The method further comprises: adjusting the matching degree weight matrix according to at least one of the feedback information of the user, the change information of the scenario features, and the adjustment information of the strategy. The method further comprises:
7. The method according to any one of claims 1 to 6, characterized in that, adjusting the first weight of the scenario features and / or the second weight of the strategy features according to at least one of the feedback information of the user, the change information of the scenario features, and the adjustment information of the strategy; determining the adjusted matching degree weight matrix according to the adjusted first weight and second weight. The method further comprises: obtaining the strategy features and general features of the user; 8. The method of claim 7, wherein, constructing a strategy knowledge base according to the strategy features and the general features; training a model based on the strategy knowledge base to obtain the strategy model. The method further comprises: training a model based on the strategy knowledge base to obtain the strategy model. vectorize a demand of the strategy feature to obtain a first vector; vectorize a label of the general feature to obtain a second vector; vectorize description information of the general feature to obtain a third vector; calculate an attention weight of the first vector and the second vector, and multiply the attention weight by the third vector to obtain a target matrix; use a clipping function to obtain a sum average of the target matrix and clip a general feature with a higher weight as an output of an attention mechanism; use the strategy feature and the general feature output by the attention mechanism as inputs to train a model to determine the strategy model; wherein the strategy model is configured to output a recommendation strength of the media asset. 9.A device for adjusting a multi-task multi-scene recommendation model, the multi-task multi-scene recommendation model being used to generate target media information recommended to a user, the device comprising a processor and a transceiver, the transceiver receiving and sending data under the control of the processor, characterized in that, The processor is configured to perform the following operations: obtain a recommendation strength of a plurality of media assets by a strategy model, wherein the strategy model is trained by a strategy feature and a general feature of a user; adjust a ranking result of each media asset in the target media asset information according to the recommendation strength of the plurality of media assets by the strategy model.
10. The apparatus of claim 9, wherein, The processor is further configured to perform the following operations: obtain score information of media assets in different scenarios and a matching degree weight matrix of scenarios and strategies; adjust a ranking result of each media asset in the target media asset information according to the score information of the media assets in the different scenarios, the recommendation strength of the plurality of media assets by the strategy model, and the matching degree weight matrix.
11. The apparatus of claim 10, wherein, The processor is further configured to perform the following operations: multiply the recommendation strength of the plurality of media assets by the strategy model by the score information of the media assets in the different scenarios to obtain a recommendation score of the plurality of media assets by the user in the different scenarios; multiply the recommendation score of the plurality of media assets by the user in the different scenarios by the matching degree weight matrix to obtain a comprehensive score of the plurality of media assets by the user in the different scenarios; adjust a ranking result of each media asset in the target media asset information according to the comprehensive score of the plurality of media assets by the user in the different scenarios.
12. The apparatus of claim 10, wherein, The processor is further configured to perform the following operations: assign a first weight to a scenario feature according to an influence degree of the scenario feature on a strategy; assign a second weight to a strategy feature according to applicability of the strategy feature in different scenarios; perform weighted summation on the first weights of a plurality of scenario features to obtain a comprehensive feature vector of the scenario; perform weighted summation on the second weights of a plurality of strategy features to obtain a comprehensive feature vector of the strategy; calculate a similarity value of the scenario and the strategy according to the comprehensive feature vector of the scenario and the comprehensive feature vector of the strategy; use the similarity value of the scenario and the strategy as an element in the matching degree weight matrix; the greater the similarity value, the greater the matching degree of the scenario and the strategy.
13. The apparatus of claim 12, wherein, The processor is further configured to perform the following operations: adjust the matching degree weight matrix according to at least one of feedback information of the user, change information of the scenario feature, and adjustment information of the strategy.
14. The apparatus of claim 13, wherein, The processor is further configured to perform the following operations: adjust the first weight of the scenario feature and / or the second weight of the strategy feature according to at least one of the feedback information of the user, the change information of the scenario feature, and the adjustment information of the strategy; determine an adjusted matching degree weight matrix according to the adjusted first weight and the second weight.
15. The apparatus of any one of claims 9-14, wherein, The processor is further configured to perform the following operations: Obtaining a policy feature and a general feature of a user; According to the policy feature and the general feature, a policy knowledge base is constructed; Based on the policy knowledge base, model training is performed to obtain the policy model.
16. The apparatus of claim 15, wherein, The processor is further configured to perform the following operations: Vectorize a requirement vector of the policy feature to obtain a first vector; Vectorize a label of the general feature to obtain a second vector; Vectorize description information of the general feature to obtain a third vector; Calculate an attention weight of the first vector and the second vector, and multiply the attention weight by the third vector to obtain a target matrix; Use a truncation function to obtain a sum average of the target matrix and cut off general features with higher weights as an output of an attention mechanism; Use the policy feature and the general feature output by the attention mechanism as inputs to perform model training to determine the policy model; The policy model is used to output a recommendation strength for media resources.
17. An adjustment device comprising a memory, a processor, and a program stored on the memory and executable on the processor; characterized in that, The processor executes the program to implement the adjustment method of the multi-task multi-scenario recommendation model according to any one of claims 1 to 9.
18. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps in the adjustment method of the multi-task multi-scenario recommendation model according to any one of claims 1 to 9.
19. A computer program product, characterised in that, The program includes computer instructions that are executed by the processor to implement the steps in the adjustment method of the multi-task multi-scenario recommendation model according to any one of claims 1 to 9.