A recommended method and apparatus, and a computer-readable storage medium
By constructing user profiles and using genetic algorithms to optimize recall parameters and dynamically adjusting the recall branch model, the problem of poor recall performance in existing recommendation systems is solved, achieving high efficiency and accuracy in personalized item recommendations.
Patent Information
- Application Number
- CN202210775609.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-01
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2042-07-01
AI Technical Summary
Existing recommendation systems struggle to achieve optimal effectiveness and efficiency when recalling users' actual needs, and recall schemes with uniform hyperparameters are ineffective and fail to meet the personalized needs of different user groups.
By constructing user profiles using multiple user information sources, determining the optimal recall parameters and recall strategies, and combining genetic algorithms to optimize the parameter space, the on/off state and hyperparameters of the recall branch model are dynamically adjusted to achieve personalized item recommendations.
This improves the success rate and accuracy of recommendation results, meets the personalized needs of different users, and enhances the efficiency and effectiveness of the recommendation system.
Smart Images

Figure CN117390252B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of recommendation application systems and algorithms, and more particularly to a recommendation method and apparatus, and a computer-readable storage medium. Background Technology
[0002] With the rapid development of technology and the ever-increasing volume of data, the goal of recommendation systems in e-commerce scenarios is to recommend products that match users' interests and preferences from massive datasets. Algorithmic recommendations aim to save users time, improve user satisfaction, and create more business value for companies. Broadly speaking, recommendation systems can be divided into two main subsystems: First, the recommendation recall system, based on certain user characteristics, quickly retrieves thousands of potentially interesting items from a massive database of hundreds of millions of products. This is then passed to the ranking stage, which can incorporate more features and use complex models to make accurate personalized recommendations.
[0003] However, when recommending what users actually need, on the one hand, activating all trigger branches for all requests is difficult to achieve optimal effect and efficiency; on the other hand, each branch model contains some hyperparameters, which are not suitable for all user groups, and the solution of using uniform hyperparameters for full branch recall also has the problem of poor effect. Summary of the Invention
[0004] This invention provides a recommendation method and apparatus, as well as a computer-readable storage medium, which can generate personalized strategies for each user and improve the success rate of user recommendation results.
[0005] The technical solution of this invention is implemented as follows:
[0006] This invention provides a recommended method, the method comprising:
[0007] Multiple user profiles were determined by using the collected user information.
[0008] Using the multiple user profiles, the model parameters corresponding to the multi-path recall branch model are combined to determine the optimal recall parameters for each of the multiple users; wherein at least one user corresponds to one optimal recall parameter.
[0009] Based on the optimal recall parameters and the multi-path recall branch model, at least one recall strategy is determined for each user.
[0010] Based on the at least one recall strategy and each user profile, item recommendations are made to obtain recommendation results.
[0011] In the above scheme, the model parameters include switching parameters and hyperparameters;
[0012] The step of combining the model parameters corresponding to the multi-path recall branch model using the multiple user profiles to determine the optimal recall parameters for each of the multiple user profiles includes:
[0013] The switching parameters and the hyperparameters are fused to determine multiple initial recall parameters corresponding to the multi-path recall branch model;
[0014] The multiple initial recall parameters are cross-crossed and mutated to determine the selectable parameter space;
[0015] Based on the multiple user profiles, the user characteristics corresponding to each user are determined; wherein, each user profile corresponds to one user.
[0016] Based on the selectable parameter space and the user characteristics corresponding to each of the multiple users, the optimal recall parameters corresponding to each of the multiple users are determined.
[0017] In the above scheme, the selectable parameter space includes multiple processed recall parameters;
[0018] The step of determining the optimal recall parameters for each of the multiple users based on the selectable parameter space and the user characteristics corresponding to each user includes:
[0019] Based on the multiple processed recall parameters and the user characteristics corresponding to each user, multiple current evaluation results are determined for each user;
[0020] Based on the multiple current evaluation results, the current recall parameters corresponding to each of the multiple processed recall parameters are determined;
[0021] Based on the current recall parameters corresponding to each of the multiple users, determine the optimal recall parameters for the current iteration for each of the multiple users;
[0022] Based on the current optimal recall parameters, the selectable parameter space is adjusted to obtain the next recall parameters based on the next selectable parameter space.
[0023] Based on the next recall parameters, the next optimal recall parameters for each of the multiple users are determined until the first preset iteration condition is met, at which point the P-th optimal recall parameters for each of the multiple users are obtained; where P is a positive integer greater than or equal to 2.
[0024] The optimal recall parameters for the Pth iteration corresponding to each of the multiple users are determined as the optimal recall parameters for each of the multiple users.
[0025] In the above scheme, determining the current recall parameters corresponding to each of the multiple users from the multiple processed recall parameters based on the multiple current evaluation results includes:
[0026] Compare multiple current evaluation results for each user to determine the current optimal evaluation result for each user;
[0027] Among the multiple processed recall parameters, the recall parameter corresponding to the current best evaluation result of each user is determined as the current recall parameter corresponding to each of the multiple users.
[0028] In the above scheme, determining the optimal recall parameters for the current iteration for each of the multiple users based on their respective current recall parameters includes:
[0029] Based on the user characteristics corresponding to each of the multiple users, the similarity between the multiple users is determined;
[0030] Based on the similarity among the multiple users, the multiple users are clustered to obtain multiple clusters; each cluster includes at least one user;
[0031] The current recall parameters corresponding to at least one user in each cluster are compared to determine the current optimal recall parameters for at least one user in each cluster, thereby obtaining the current optimal recall parameters for each of the multiple users belonging to the multiple clusters.
[0032] In the above scheme, the optimal model parameters include the item recall quantity;
[0033] The step of determining at least one recall strategy for each user based on the optimal model parameters and the multi-path recall branch model includes:
[0034] Based on the optimal model parameters corresponding to each user, determine the target recall branch model corresponding to each user and the item recall quantity corresponding to the target recall branch model;
[0035] If each user has a target recall branch model, then a recall strategy for each user is determined based on the target recall branch model and the corresponding item recall volume.
[0036] If each user corresponds to multiple target recall branch models, then the multiple target recall branch models are arranged and combined to determine multiple sets corresponding to each user; wherein, each set includes at least one target recall branch model and the item recall volume corresponding to the at least one target recall branch model;
[0037] Based on the multiple sets corresponding to each user, multiple recall strategies are determined for each user; wherein, one set corresponds to one recall strategy.
[0038] In the above scheme, the step of recommending items based on the at least one recall strategy and each user profile to obtain recommendation results includes:
[0039] Obtain at least one recall strategy corresponding to each user;
[0040] Using the recall branch model corresponding to the at least one recall strategy, analyze each user profile, make item recommendations, and determine the initial recommendation result for each user.
[0041] The initial recommendation results are sorted to obtain the final recommendation result.
[0042] This embodiment provides a recommendation device, which includes a recommendation unit and a determination unit; wherein...
[0043] The determining unit is used to determine multiple user profiles using the collected user information.
[0044] The determining unit is further configured to combine the model parameters corresponding to the multi-path recall branch model using the multiple user profiles to determine the optimal recall parameters corresponding to each of the multiple users; wherein at least one user corresponds to one optimal recall parameter.
[0045] The determining unit is further configured to determine at least one recall strategy for each user based on the optimal recall parameters and the multi-path recall branch model.
[0046] The recommendation unit is also used to recommend items based on the at least one recall strategy and each user profile to obtain recommendation results.
[0047] This embodiment also provides a recommendation device, the recommendation device comprising:
[0048] Memory, used to store executable instructions;
[0049] When a processor executes executable instructions stored in the memory, the processor performs the recommended method when the executable instructions are executed.
[0050] This invention provides a storage medium storing executable instructions, which, when executed, cause a first processor to perform the recommended method as described in this invention.
[0051] This invention provides a recommendation method and apparatus, and a computer-readable storage medium. The method includes: determining multiple user profiles using collected user information; combining model parameters corresponding to a multi-path recall branch model using the multiple user profiles to determine optimal recall parameters for each of the multiple users; wherein at least one user corresponds to one optimal recall parameter; determining at least one recall strategy for each user based on the optimal recall parameters and the multi-path recall branch model; and recommending items based on the at least one recall strategy and each user profile to obtain recommendation results. By combining model parameters corresponding to a multi-path recall branch model using multiple user profiles to determine optimal model parameters for each user, and selecting a recall strategy and recall branch model to recommend items to users based on the optimal model parameters, personalized strategy generation can be performed for each user, improving the success rate of recommendation results. Attached Figure Description
[0052] Figure 1 A flowchart illustrating a recommended method provided in an embodiment of the present invention. Figure 1 ;
[0053] Figure 2 This is a schematic diagram illustrating the effect of the dynamic routing recall framework provided in an embodiment of the present invention;
[0054] Figure 3 This is a schematic diagram illustrating the effect of the triggering strategy controller provided in an embodiment of the present invention;
[0055] Figure 4 A flowchart illustrating a recommended method provided in an embodiment of the present invention. Figure 2 ;
[0056] Figure 5 A schematic diagram of the structure of a recommended device provided in an embodiment of the present invention;
[0057] Figure 6 A schematic diagram of another recommended device provided in an embodiment of the present invention. Detailed Implementation
[0058] 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 protection scope of the present invention.
[0059] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Figure 1A recommended method flowchart provided for embodiments of the present invention Figure 1 , will combine Figure 1 The steps shown are explained.
[0060] S101. Using the collected user information, determine multiple user profiles.
[0061] In embodiments of the present invention, information from multiple users is collected. This user information may include attribute information, behavioral information, user preference information, and user lifestyle information, etc. The collected user information is used to construct corresponding user profiles. User profiles can be used to analyze users, and then recommendations can be provided based on this analysis.
[0062] In embodiments of the present invention, the acquisition and use of user information are carried out with the knowledge and consent of the relevant parties, and the acquisition and use of user information comply with relevant laws and public order and good morals.
[0063] In embodiments of this invention, a user profile refers to a tagged user model abstracted based on user attributes, preferences, lifestyle habits, and behaviors. Simply put, it's about tagging users, and tags are highly refined feature identifiers derived from user information analysis. Tagging allows for the description of users using highly generalized and easily understood features, making it easier for humans to understand users and facilitating computer processing. User profile data may include the following: interest characteristics: browsed content, saved content, reading information, purchase preferences, etc.; consumption characteristics: consumption-related characteristics; device attributes: terminal characteristics used, etc.; behavioral data: user behavior log data on the website, such as access time and browsing path; and social data: user social-related data. User profiles are frequently used in personalized recommendations, advertising systems, event marketing, content recommendations, and interest preference scenarios.
[0064] S102. Using multiple user profiles, combine the model parameters corresponding to the multi-path recall branch model to determine the optimal recall parameters for each user; wherein, at least one user corresponds to one optimal recall parameter.
[0065] In an embodiment of the present invention, after constructing each user profile, the recall parameters corresponding to the multi-path recall branch model are combined according to each user profile to determine the optimal recall parameters corresponding to each user.
[0066] In embodiments of the present invention, the recall parameters corresponding to each branch model include switch parameters, hyperparameters, and item recall quantities, etc., and the switch parameters are composed of the ordinal numbers of each branch model. A trigger strategy controller (not limited by the present invention) can be used to model the user's profile, behavior, and the relationship between the switch and hyperparameters of the trigger branches. If there are m recall branch models, the switch parameters take values of (1, 2, ..., m). After processing by the controller, the optimal recall parameters corresponding to the user are obtained, and corresponding items are recommended to the corresponding user based on the parameter values included in the optimal recall.
[0067] S103. Based on the optimal recall parameters and the multi-path recall branch model, determine at least one recall strategy for each user.
[0068] In embodiments of the present invention, at least one recall strategy is determined for each user based on the determined optimal recall parameters and the multi-path recall branch model.
[0069] S104. Based on at least one recall strategy and each user profile, recommend items and obtain recommendation results.
[0070] In an embodiment of the present invention, after obtaining at least one recall strategy corresponding to a user, item recommendations are made based on the user profile corresponding to the user according to the recall strategy to obtain recommendation results.
[0071] In some embodiments of the present invention, the overall multi-path recall framework is as follows: Figure 2 As shown. In Figure 2 In this approach, multiple users, their corresponding behavior profiles, and the model parameters (HP) of each branch model (M1, M2, ..., Mm) are used as inputs to the routing controller. This controller models the relationships between user profiles, behaviors, trigger branch on / off states, and hyperparameters, ultimately outputting the optimal recall parameters for each user. Based on these optimal recall parameters, each user determines their corresponding recall strategy. Based on the recall strategy, the corresponding recall branch model is determined. This determined recall branch model is then used to analyze the user's profile and recommend items. Figure 2 In this process, the user profiles and parameters of the multi-branch model are processed using a triggering policy controller to obtain at least one recall strategy corresponding to each user. For user 1, model 2 is an invalid path, meaning that model 2 is prohibited from being used to recommend items to the user. Models 1, 3, and m are the item recall models corresponding to user 1. Arbitrary combinations of the corresponding item recall models will form multiple sets, such as:
[0072] A set of models (Model 1, Model 3, ..., Model m), (Model 1, Model 3, ..., Model m-1), and (Model 3, Model 4, ..., Model m), etc., corresponds to a recall strategy. The corresponding recall model is selected for item recall based on the chosen strategy. During item recall, each recall model recalls items based on its corresponding recall volume. The recall volumes of each model are merged to form a candidate set. Finally, the candidate set is ranked using a ranking model to obtain the final recommended items for user 1. For user 2, Model 1 is banned. For user n, Model 1 and Model m are banned. The same principle is used to recommend items to users.
[0073] In some embodiments of the present invention, the main working principle of the triggering policy controller is as follows: Figure 3 As shown, the controller is iteratively updated based on the idea of reinforcement learning, which mainly includes three modules: strategy layer, state layer, and reward layer. The trigger strategy controller mainly controls the on / off state of trigger branches and the setting of hyperparameters for activation branches. The on / off state of trigger branches and the setting of hyperparameters for each branch are determined by a series of parameters. Multiple branches and multiple parameters result in a large search space when searching for the best strategy for users, leading to parameter space explosion. Therefore, the main role of the state layer is to solve the performance problem caused by parameter space explosion using a genetic algorithm. With a large parameter space and a large number of users, it is impractical to obtain the personalized best recall parameters for users through brute force search. Parameter space explosion is difficult to solve. Therefore, this invention proposes to search for locally optimal recall parameters (actions) for users based on a genetic algorithm (GA) to solve the performance problem caused by parameter space explosion. First, the parameter space is constructed using the on / off parameters and hyperparameters of each recall branch model. The on / off parameters and hyperparameters are combined to generate a parameter set (HPs), which contains multiple actions. The initial action space includes multiple initial actions, from action1 to action... mThis invention utilizes a genetic algorithm for crossover and mutation processing to generate a computable selected action space. Based on this space, the recall parameters for each user are determined. Given the randomness of a user's search action, a weighted voting method is proposed to select the optimal recall parameters for the target user by considering the optimal recall parameters of N similar users. Since user behavior is assumed to measure user similarity, it is natural to incorporate more information to improve the accuracy of similarity measurement. Therefore, this invention proposes constructing user features using user-side features and contextual features. Measuring similarity through multi-dimensional construction of user features improves accuracy. The similarity between two users is evaluated by the similarity of the products they click on and purchase. Here, this invention proposes constructing user features based on multiple dimensions of information, including historical behavior, attributes, and contextual information. In the strategy layer, users are clustered based on their characteristics, resulting in multiple action clusters. The recall parameters of users within each cluster are compared, and the optimal recall parameter is selected. This optimal recall parameter is then used as the recall parameter for each user in that cluster. After one round, each cluster and its users will have an optimal recall parameter. For example, cluster 1 (cl-uster1) corresponds to optimal recall parameter 1 (action1), cluster 2 (cluster2) corresponds to optimal recall parameter 2 (action1), and cluster m (clusterm) corresponds to optimal recall parameter m (actionm), and so on. The feedback layer feeds back the multiple optimal recall parameters obtained after one round to the state layer for a new round of cross-mutation processing until a preset number of iterations is reached, ultimately yielding the optimal recall parameter for each user.
[0074] It is understood that, in the embodiments of the present invention, by utilizing the multiple user profiles, the model parameters corresponding to the multi-path recall branch model are combined to determine the optimal recall parameters for each user; based on the optimal recall parameters and the multi-path recall branch model, at least one recall strategy is determined for each user; based on at least one recall strategy and each user profile, item recommendation is performed to obtain the recommendation result. This method enables the construction of personalized fusion strategies through user profiles, allowing for the generation of personalized strategies for each user, thereby improving the accuracy of recommending items to users.
[0075] In some embodiments of the present invention Figure 4 A recommended method flowchart provided for embodiments of the present invention Figure 2S102 can be implemented through S1021 to S1024, which will be explained in conjunction with the following steps.
[0076] S1021. Integrate the switching parameters and hyperparameters to determine multiple initial recall parameters corresponding to the multi-path recall branch model.
[0077] In some embodiments of the present invention, the switching parameters and hyperparameters are fused to form multiple initial recall parameters corresponding to the multi-path recall branch model.
[0078] In some embodiments of the present invention, the switching parameter of the triggering branch is set to M, and the hyperparameters of each branch construct a parameter space S, wherein M = {M1, M2, ... M}. m}、S={s1,s2,…s m The two parameters are combined to generate the parameter HP_s = {a1, a2, ... a}. m Each 'a' in HP_s can be considered an action. Due to the large number of parameters and the large parameter space, the number of actions is on the order of hundreds of millions. Therefore, actions are initialized at the beginning.
[0079] S1022. Perform crossover and mutation processing on multiple initial recall parameters to determine the selectable parameter space.
[0080] In some embodiments of the present invention, multiple initial recall parameters are cross-crossed and mutated to obtain a selectable parameter space.
[0081] In some embodiments of the present invention, multiple initial recall parameters are cross-crossed and mutated to select a computable parameter space (Selected Action Space), which contains multiple recall parameters after cross-crossing and mutation.
[0082] S1023. Based on multiple user profiles, determine the user characteristics corresponding to each user; wherein, each user profile corresponds to one user.
[0083] In some embodiments of the present invention, user characteristics corresponding to each user are constructed based on each user's profile.
[0084] S1024. Based on the selectable parameter space and the user characteristics corresponding to each user, determine the optimal recall parameters corresponding to each user.
[0085] In some embodiments of the present invention, the optimal recall parameters for each user are calculated based on the multiple recall parameters after crossover and mutation processing contained in the selectable parameter space and the user characteristics corresponding to each user.
[0086] In some embodiments of the present invention, S1024 can be implemented by S201 to S206, which will be described in detail in conjunction with the following steps.
[0087] S201. Based on multiple processed recall parameters and the user characteristics corresponding to each user, determine multiple current evaluation results for each user.
[0088] In some embodiments of the present invention, multiple current evaluation results for each user are obtained based on multiple processed recall parameters and user characteristics corresponding to multiple users.
[0089] In some embodiments of the present invention, the evaluation result is equivalent to predicting the benefit that a certain recall parameter can bring as the optimal recall parameter for a user. Multiple current evaluation results can be calculated for each user based on multiple processed recall parameters. The optimal current evaluation result for each user is then determined by selecting from these multiple current evaluation results. Multiple current evaluation results for each user can be calculated using the following formula (1):
[0090] S u =argmaxf(u,S) (1)
[0091] In the above formula, S u The evaluation result is represented by u; the user is represented by S; the parameter space is optional; the function's form is not unique and should be set according to the specific situation.
[0092] S202. Based on multiple current evaluation results, determine the current recall parameters corresponding to each user from multiple processed recall parameters.
[0093] In some embodiments of the present invention, after calculating multiple current evaluation results for each user, the multiple current evaluation results are processed to obtain the optimal current evaluation result, and the recall parameter corresponding to the optimal current evaluation result is used as the current recall parameter for the user.
[0094] S203. Based on the current recall parameters corresponding to each of the multiple users, determine the optimal recall parameters for the current iteration corresponding to each of the multiple users.
[0095] In some embodiments of the present invention, after obtaining the current recall parameters corresponding to each user, the optimal recall parameters for the current time corresponding to each user are determined based on the current recall parameters corresponding to each user.
[0096] S204. Based on the current optimal recall parameters, adjust the parameters in the selectable parameter space to obtain the next recall parameters based on the next selectable parameter space.
[0097] In some embodiments of the present invention, the selectable parameter space is adjusted based on multiple current optimal recall parameters, and the next recall parameters are obtained using the adjusted selectable parameter space.
[0098] In some embodiments of the present invention, the optimal recall parameter for the current iteration is the optimal recall parameter obtained by the user in the first iteration of a preset number of training iterations. The optimal recall parameter obtained in the first iteration is used as the starting data for the second iteration for crossover and mutation processing to obtain multiple newly processed recall parameters stored in the selectable space. Then, the evaluation result corresponding to each user is calculated. Based on the multiple evaluation results, the second recall parameter corresponding to each user is determined from the multiple newly processed recall parameters.
[0099] S205. Based on the next recall parameters, continue to determine the next optimal recall parameters for each of the multiple users until the first preset iteration condition is met, and obtain the P-th optimal recall parameters for each of the multiple users; where P is a positive integer greater than or equal to 2.
[0100] In some embodiments of the present invention, after obtaining the second recall parameters corresponding to each user according to S204, multiple evaluation results are calculated and obtained. Among the multiple evaluation results, the optimal evaluation result for each user in the current instance is selected, thereby determining the optimal recall parameters for each user in the current instance, i.e., the second instance. Using the optimal recall parameters for the second instance corresponding to the user, the optimal recall parameters for the third instance corresponding to each user are calculated, and so on, continuously iterating until a preset iteration condition is reached, thus obtaining the p-th recall parameters for each user.
[0101] In some embodiments of the present invention, the first preset iteration condition may be the number of iterations or a certain threshold, and the present invention does not impose any limitations.
[0102] S206. Determine the optimal recall parameters for the Pth iteration for each of the multiple users as the optimal recall parameters for each of the multiple users.
[0103] In some embodiments of the present invention, after obtaining the optimal recall parameters for each user in the p-th iteration, the optimal recall parameters obtained in the p-th iteration are used as the final optimal recall parameters for each user. Here, P is a preset number of iterations.
[0104] In some embodiments of the present invention, S202 can be implemented by S2021 to S2022, which will be described in detail in conjunction with the following steps.
[0105] S2021. Compare the multiple current evaluation results for each user to determine the current optimal evaluation result for each user.
[0106] In some embodiments of the present invention, multiple current evaluation results for each user are compared, and the optimal evaluation result for each user is determined from among the multiple current evaluation results.
[0107] In some embodiments of the present invention, multiple current evaluation results can be sorted to select the optimal evaluation result.
[0108] S2022. Among the multiple processed recall parameters, the recall parameter corresponding to the current best evaluation result of each user is determined as the current recall parameter corresponding to each user.
[0109] In some embodiments of the present invention, among multiple processed recall parameters, the recall parameter corresponding to the current best evaluation result of each user is determined as the current recall parameter corresponding to each user.
[0110] In some embodiments of the present invention, S203 can be implemented by S2031 to S2033, which will be described in detail in conjunction with the following steps.
[0111] S2031. Determine the similarity between multiple users based on their respective user characteristics.
[0112] In some embodiments of the present invention, the similarity between each user is calculated based on the characteristics corresponding to each user.
[0113] In some embodiments of the present invention, the following formula (2) can be used for calculation:
[0114]
[0115] In the above formula, x i x j Here, denoted as user features; dist represents similarity; and T is the device matrix.
[0116] In some embodiments of the present invention, the similarity between each user can be calculated using Euclidean distance, cosine similarity, Pearson correlation, modified cosine similarity, Hamming distance, and Manhattan distance, and the present invention does not impose any limitations.
[0117] S2032. Based on the similarity between multiple users, cluster the multiple users to obtain multiple clusters; each cluster includes at least one user.
[0118] In some embodiments of the present invention, after obtaining the similarity between each user, the similar users are clustered according to the multiple similarities to obtain multiple clusters, and each cluster includes at least one user.
[0119] S2033. Compare the current recall parameters corresponding to at least one user in each cluster to determine the current optimal recall parameters for at least one user in each cluster, thereby obtaining the current optimal recall parameters for multiple users belonging to multiple clusters.
[0120] In some embodiments of the present invention, the current recall parameters corresponding to users in each cluster are compared, and the optimal recall parameter is determined among multiple current recall parameters. The optimal recall parameter is then used as the current optimal recall parameter for each user in the cluster.
[0121] In some embodiments of the present invention, S103 can be implemented by S1031 to S1034, which will be described in detail in conjunction with the following steps.
[0122] S1031. Based on the optimal model parameters corresponding to each user, determine the target recall branch model corresponding to each user and the item recall quantity corresponding to the target recall branch model.
[0123] In some embodiments of the present invention, after each user obtains their respective optimal model parameters, the target recall branch model and the item recall quantity corresponding to the target recall branch model are determined in the multi-branch model based on the optimal model parameters.
[0124] In some embodiments of the present invention, the target recall branch model corresponding to each user is determined by the value of the switch parameter included in the user's optimal model parameters. If the switch parameter includes a value of (1,2), then the user corresponds to the first and second recall branch models.
[0125] S1032. If each user has a target recall branch model, then based on a target recall branch model and the corresponding item recall volume, determine a recall strategy for each user.
[0126] In some embodiments of the present invention, if each user has one target recall branch model, then each user has one recall strategy.
[0127] S1033. If each user has multiple target recall branch models, then the multiple target recall branch models are arranged and combined to determine multiple sets corresponding to each user; wherein, each set includes at least one target recall branch model and at least one item recall quantity corresponding to the target recall branch model.
[0128] In some embodiments of the present invention, if there are multiple target recall branch models corresponding to each user, then the multiple target recall branch models are arranged and combined to determine multiple sets corresponding to each user. Each set contains at least one recall branch model and the item recall quantity corresponding to that model.
[0129] In some embodiments of the present invention, if the value of the switch parameter is (1,2,3), then the user corresponds to the 1st, 2nd, and 3rd recall branch models. By arranging and combining these three recall branch models, seven sets are obtained: (1,2,3), (1,2), (1,3), (2,3), (1), (2), and (3).
[0130] S1034. Based on the multiple sets corresponding to each user, determine multiple recall strategies corresponding to each user; wherein, one set corresponds to one recall strategy.
[0131] In some embodiments of the present invention, multiple recall strategies corresponding to each user are determined based on multiple sets corresponding to each user.
[0132] In some embodiments of the present invention, as can be seen from S1033, there are 7 recall strategies for each user.
[0133] In some embodiments of the present invention, S104 can be implemented by S1041 to S1043, which will be described in detail in conjunction with the following steps.
[0134] S1041. Obtain at least one recall strategy corresponding to each user.
[0135] In some embodiments of the present invention, at least one recall strategy corresponding to each user is obtained.
[0136] S1042. Using the recall branch model corresponding to at least one recall strategy, analyze each user profile, make item recommendations, and determine the initial recommendation results for each user.
[0137] In some embodiments of the present invention, based on the recall branch model corresponding to at least one recall strategy for each user, the user profile is analyzed, item recommendations are performed, and the initial recommendation result for each user is determined.
[0138] S1043. Sort the initial recommendation results to obtain the final recommendation results.
[0139] In some embodiments of the present invention, the initial recommendation results are sorted to obtain the user's final recommendation results.
[0140] It is understood that, in some embodiments of the present invention, by modeling user profiles, behaviors and triggering branch switches, and the direct relationship between hyperparameters, a personalized fusion strategy is constructed to generate personalized strategies for each user, thereby improving the accuracy of item recommendations and maximizing platform revenue.
[0141] This invention also provides a recommendation method that adds a learnable trigger strategy controller in the multi-path recall fusion stage. This controller models the user's profile, behavior, trigger branch switching, and the relationship between hyperparameters. It can be used in real-world scenarios for recommending everyday items to users. The specific steps are as follows.
[0142] S1. Combine the parameters HP (including switch parameters and hyperparameters) of the branch model to generate the final parameter set HP_s = {a1; a2; ... am}. Each a in HP_s is called an action.
[0143] S2. Perform cross-mutation processing on the state layer of the triggering policy controller for the obtained multiple actions to obtain a selectable parameter space;
[0144] S3. For a given user, calculate the optimal action for each user based on a limited action space;
[0145] S4. In the strategy layer, users are clustered according to their respective characteristics to obtain multiple clusters, and each cluster contains at least one user;
[0146] S5. Compare the actions corresponding to users in each cluster, find the optimal action, and use the optimal action as the action for users in each cluster.
[0147] S6. The feedback layer feeds back the action corresponding to each user obtained in the first iteration to the state layer for processing until the preset number of iterations is reached and the optimal action corresponding to the user is obtained.
[0148] S7. Based on the optimal recall parameters and the multi-path recall branch model, determine at least one recall strategy for each user;
[0149] S8. Based on at least one recall strategy and each user profile, perform item recommendation to obtain recommendation results.
[0150] It is understood that, in the embodiments of the present invention, by modeling the user profile, behavior and the on / off state of the triggering branch, and the relationship between the hyperparameters by the triggering policy controller, it is ensured that the trigger signal can be dynamically determined to be routed to which branches according to the user information when the service is in use, and the relevant hyperparameters of the activated branch are adjusted to generate personalized strategies for each user, so as to maximize the platform's revenue and ultimately improve the recall relevance.
[0151] This invention provides a recommended device, such as... Figure 5 As shown, Figure 5 This is a schematic diagram of a recommendation device provided in an embodiment of the present invention. The recommendation device includes: a determining unit 501 and a recommendation unit 502; wherein,
[0152] The determining unit 501 is used to determine multiple user profiles using the collected user information.
[0153] The determining unit 501 is further configured to combine the model parameters corresponding to the multi-path recall branch model using the multiple user profiles to determine the optimal recall parameters corresponding to each of the multiple users; wherein at least one user corresponds to one optimal recall parameter.
[0154] The determining unit 501 is further configured to determine at least one recall strategy for each user based on the optimal recall parameters and the multi-path recall branch model.
[0155] The recommendation unit 502 is used to recommend items based on the at least one recall strategy and each user profile, and obtain recommendation results.
[0156] In some embodiments of the present invention, the model parameters include switching parameters and hyperparameters; the determining unit 501 is further configured to fuse the switching parameters and the hyperparameters to determine multiple initial recall parameters corresponding to the multi-path recall branch model; and to perform cross-validation and mutation processing on the multiple initial recall parameters to determine a selectable parameter space; and to determine user features corresponding to each of the multiple user profiles; wherein each user profile corresponds to one user; and to determine the optimal recall parameters corresponding to each of the multiple users based on the selectable parameter space and the user features corresponding to each of the multiple users.
[0157] In some embodiments of the present invention, the selectable parameter space includes multiple processed recall parameters; the determining unit 501 is further configured to: determine multiple current evaluation results for each user based on the multiple processed recall parameters and the user characteristics corresponding to each of the multiple users; and determine the current recall parameters corresponding to each of the multiple users from among the multiple processed recall parameters according to the multiple current evaluation results; and determine the current optimal recall parameters corresponding to each of the multiple users based on the current recall parameters corresponding to each of the multiple users; and adjust the parameters of the selectable parameter space based on the current optimal recall parameters, thereby obtaining the next recall parameters based on the next selectable parameter space; and continue to determine the next optimal recall parameters corresponding to each of the multiple users based on the next recall parameters, until a first preset iteration condition is reached, thereby obtaining the Pth optimal recall parameters corresponding to each of the multiple users; where P is a positive integer greater than or equal to 2; and determine the Pth optimal recall parameters corresponding to each of the multiple users as the optimal recall parameters corresponding to each of the multiple users.
[0158] In some embodiments of the present invention, the determining unit 501 is further configured to compare multiple current evaluation results of each user to determine the current optimal evaluation result of each user; and to determine the recall parameter corresponding to the current optimal evaluation result of each user as the current recall parameter corresponding to each of the multiple processed recall parameters.
[0159] In some embodiments of the present invention, the determining unit 501 is further configured to: determine the similarity between the plurality of users based on the user characteristics corresponding to each of the plurality of users; and cluster the plurality of users based on the similarity between the plurality of users to obtain a plurality of clusters; each cluster includes at least one user; and compare the current recall parameters corresponding to each of the at least one user in each cluster to determine the current optimal recall parameter for each of the at least one user in each cluster, thereby obtaining the current optimal recall parameter corresponding to each of the plurality of users belonging to the plurality of clusters.
[0160] In some embodiments of the present invention, the optimal model parameters include the item recall volume; the determining unit 501 is further configured to determine the target recall branch model and the item recall volume corresponding to the target recall branch model for each user based on the optimal model parameters corresponding to each user; and if there is one target recall branch model for each user, then a recall strategy for each user is determined based on one target recall branch model and the corresponding item recall volume; and if there are multiple target recall branch models for each user, then the multiple target recall branch models are arranged and combined to determine multiple sets for each user; wherein each set includes at least one target recall branch model and the item recall volume corresponding to the at least one target recall branch model; and multiple recall strategies for each user are determined based on the multiple sets for each user; wherein one set corresponds to one recall strategy.
[0161] In some embodiments of the present invention, the recommendation unit 502 is further configured to: acquire at least one recall strategy corresponding to each user; and analyze the profile of each user using the recall branch model corresponding to the at least one recall strategy, perform item recommendations, determine the initial recommendation result corresponding to each user; and sort the initial recommendation results to obtain the recommendation result.
[0162] It is understandable that in the above device implementation scheme, by combining the model parameters corresponding to the multi-path recall branch model using multiple user profiles, the optimal model parameters corresponding to each user are determined. Based on the optimal model parameters, a recall strategy and recall branch model are selected to recommend items to the user, so that each user can generate personalized strategies and improve the success rate of recommendation results.
[0163] Based on the methods described in the above embodiments, an embodiment of the present invention provides a structural schematic diagram, as follows: Figure 6 As shown, Figure 6 A schematic diagram of a recommended device provided in an embodiment of the present invention includes: a processor 601 and a memory 602; the memory 602 stores one or more programs executable by the processor 601, and when one or more programs are executed, the processor 601 executes a recommended method corresponding to the previous embodiment.
[0164] This invention provides a computer-readable storage medium storing executable instructions for inducing a processor to execute the recommended method.
[0165] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.
[0166] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. 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 illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0167] 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 an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0168] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment 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.
[0169] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention.
Claims
1. A recommendation method characterized by comprising: include: Multiple user profiles were determined by using the collected user information. Using the multiple user profiles, the model parameters corresponding to the multi-path recall branch model are combined to determine the optimal recall parameters for each of the multiple users; wherein at least one user corresponds to one optimal recall parameter. Based on the optimal recall parameters and the multi-path recall branch model, at least one recall strategy is determined for each user. Based on the at least one recall strategy and each user profile, item recommendations are made to obtain recommendation results; The model parameters include switching parameters and hyperparameters; the step of combining the model parameters corresponding to the multi-path recall branch model using the multiple user profiles to determine the optimal recall parameters for each of the multiple users includes: The switching parameters and the hyperparameters are fused to determine multiple initial recall parameters corresponding to the multi-path recall branch model; The multiple initial recall parameters are cross-crossed and mutated to determine the selectable parameter space; Based on the multiple user profiles, the user characteristics corresponding to each user are determined; wherein, each user profile corresponds to one user. Based on the selectable parameter space and the user characteristics corresponding to each of the multiple users, the optimal recall parameters corresponding to each of the multiple users are determined.
2. The method of claim 1, wherein, The selectable parameter space includes multiple processed recall parameters; The step of determining the optimal recall parameters for each of the multiple users based on the selectable parameter space and the user characteristics corresponding to each user includes: Based on the multiple processed recall parameters and the user characteristics corresponding to each user, multiple current evaluation results are determined for each user; Based on the multiple current evaluation results, the current recall parameters corresponding to each of the multiple processed recall parameters are determined; Based on the current recall parameters corresponding to each of the multiple users, determine the optimal recall parameters for the current iteration for each of the multiple users; Based on the current optimal recall parameters, the selectable parameter space is adjusted to obtain the next recall parameters based on the next selectable parameter space. Based on the next recall parameters, the next optimal recall parameters for each of the multiple users are determined until the first preset iteration condition is met, at which point the P-th optimal recall parameters for each of the multiple users are obtained; where P is a positive integer greater than or equal to 2. The optimal recall parameters for the Pth iteration corresponding to each of the multiple users are determined as the optimal recall parameters for each of the multiple users.
3. The method of claim 2, wherein, The step of determining the current recall parameters corresponding to each of the multiple users from the multiple processed recall parameters based on the multiple current evaluation results includes: Compare multiple current evaluation results for each user to determine the current optimal evaluation result for each user; Among the multiple processed recall parameters, the recall parameter corresponding to the current best evaluation result of each user is determined as the current recall parameter corresponding to each of the multiple users.
4. The method of claim 2, wherein, The step of determining the optimal recall parameters for the current iteration for each of the multiple users based on their respective current recall parameters includes: Based on the user characteristics corresponding to each of the multiple users, the similarity between the multiple users is determined; Based on the similarity among the multiple users, the multiple users are clustered to obtain multiple clusters; each cluster includes at least one user; The current recall parameters corresponding to at least one user in each cluster are compared to determine the current optimal recall parameters for at least one user in each cluster, thereby obtaining the current optimal recall parameters for each of the multiple users belonging to the multiple clusters.
5. The method of claim 1, wherein, The optimal recall parameters include the item recall quantity; The step of determining at least one recall strategy for each user based on the optimal recall parameters and the multi-path recall branch model includes: Based on the optimal recall parameters corresponding to each user, determine the target recall branch model corresponding to each user and the item recall quantity corresponding to the target recall branch model; If each user has a target recall branch model, then a recall strategy for each user is determined based on the target recall branch model and the corresponding item recall volume. If each user corresponds to multiple target recall branch models, then the multiple target recall branch models are arranged and combined to determine multiple sets corresponding to each user; wherein, each set includes at least one target recall branch model and the item recall volume corresponding to the at least one target recall branch model; Based on the multiple sets corresponding to each user, multiple recall strategies are determined for each user; wherein, one set corresponds to one recall strategy.
6. The method according to any one of claims 1 to 5, characterized in that, The process of recommending items based on the at least one recall strategy and each user profile to obtain recommendation results includes: Obtain at least one recall strategy corresponding to each user; Using the recall branch model corresponding to the at least one recall strategy, analyze each user profile, make item recommendations, and determine the initial recommendation result for each user. The initial recommendation results are sorted to obtain the final recommendation result.
7. A recommendation device characterized by comprising: The recommendation device includes a recommendation unit and a determination unit; wherein... The determining unit is used to determine multiple user profiles using the collected user information. The determining unit is further configured to combine the model parameters corresponding to the multi-path recall branch model using the multiple user profiles to determine the optimal recall parameters corresponding to each of the multiple users; wherein at least one user corresponds to one optimal recall parameter. The determining unit is further configured to determine at least one recall strategy for each user based on the optimal recall parameters and the multi-path recall branch model. The recommendation unit is also used to recommend items based on the at least one recall strategy and each user profile to obtain recommendation results; The model parameters include switching parameters and hyperparameters; the determining unit is further configured to: fuse the switching parameters and the hyperparameters to determine multiple initial recall parameters corresponding to the multi-path recall branch model; perform cross-validation and mutation processing on the multiple initial recall parameters to determine a selectable parameter space; determine the user features corresponding to each of the multiple user profiles; wherein each user profile corresponds to one user; and determine the optimal recall parameters corresponding to each of the multiple users based on the selectable parameter space and the user features corresponding to each of the multiple users.
8. A recommendation apparatus characterized by comprising: The recommendation device includes: Memory, used to store executable instructions; A processor, when executing executable instructions stored in the memory, implements the recommended method according to any one of claims 1 to 6.
9. A computer-readable storage medium, characterized in that, It stores executable instructions for causing the processor to execute, thereby implementing the recommended method as described in any one of claims 1 to 6.
Citation Information
Patent Citations
Data processing method, system, medium and computing device
CN109087162A
Resource recommendation method and apparatus, parameter determination method and apparatus, device, and medium
US20210406327A1
