Object recommendation method and device, computer equipment and storage medium

Abstract

The invention relates to the technical field of big data processing, and particularly discloses an object recommendation method and device, computer equipment, a storage medium and a computer program product. The method comprises the steps of obtaining a user identifier of a target user, a service identifier of a to-be-recommended service and a plurality of pre-configured recommendation models in response to an object recommendation event triggered for the target user; for each recommendation model, obtaining model use data of the recommendation model from a data storage area corresponding to the recommendation model; inputting the model use data into a corresponding recommendation model to obtain candidate recommendation results of the target user corresponding to the candidate objects; and fusing the respective candidate recommendation results of the recommendation models, and determining a target recommendation object of the target user from the candidate objects. By adopting the method, the object recommendation accuracy can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of big data processing, and in particular to an object recommendation method, apparatus, computer equipment, storage medium and computer program product. Background Art

[0002] As the business scope of financial institutions, such as banks, gradually expands, users have more and more product consumption choices. How to recommend corresponding products that meet user needs and improve user satisfaction is an issue that needs to be focused on in current banking business processing.

[0003] Traditional business product recommendations are usually based on rules, that is, pre-set rules and conditions to recommend products to users. Taking fund products as an example, they are generally based on factors such as user risk preferences, investment goals or investment periods, and corresponding rules are set to screen out suitable fund products and recommend them to users. Alternatively, collaborative filtering algorithms can be used to make product recommendations based on the user's historical behavior data and the association between similar users. However, traditional recommendation methods are often prone to cold start due to the lack of user feature data, which greatly reduces the accuracy of recommendations. Summary of the invention

[0004] Based on this, it is necessary to provide an object recommendation method, apparatus, computer device, computer-readable storage medium and computer program product that can improve the accuracy of object recommendation in response to the above technical problems.

[0005] In a first aspect, the present application provides an object recommendation method, the method comprising:

[0006] In response to an object recommendation event triggered for a target user, obtaining a user identifier of the target user, a service identifier of a service to be recommended, and a plurality of pre-configured recommendation models;

[0007] For each of the recommendation models, obtaining model usage data of the recommendation model from a data storage area corresponding to the recommendation model; the model usage data includes user information matching the user identifier and service information matching the service identifier; the service information includes object information of each of a plurality of candidate objects belonging to the service to be recommended;

[0008] Inputting the model usage data into a corresponding recommendation model to obtain a candidate recommendation result corresponding to each candidate object for the target user;

[0009] The candidate recommendation results of each of the recommendation models are integrated, and the target recommendation object of the target user is determined from the candidate objects.

[0010] In one embodiment, the method further comprises:

[0011] Obtaining a model identifier of the recommendation model;

[0012] Based on the mapping relationship between the model identifier and each storage area of ​​the database, the storage area matching the model identifier is determined as the data storage area corresponding to the recommendation model.

[0013] In one embodiment, the recommendation model includes a classification recommendation model, and the classification recommendation model includes multiple feature extraction paths, and different feature extraction paths correspond to different feature extraction methods;

[0014] The step of inputting the model usage data into a corresponding recommendation model to obtain a candidate recommendation result corresponding to each candidate object for the target user includes:

[0015] Determining, according to the data types of the sub-data in the model usage data, the feature extraction paths corresponding to the sub-data in the classification recommendation model;

[0016] Determine the associated data of each of the sub-usage data based on the associated relationship between the sub-data;

[0017] For each of the sub-data, the sub-data and the associated data of the sub-data are input into the classification recommendation model based on the feature extraction path corresponding to the sub-data to obtain the candidate recommendation results corresponding to each of the candidate objects for the target user.

[0018] In one of the embodiments, the classification recommendation model includes a first feature extraction path based on a convolutional neural network layer and a second feature extraction path based on a long short-term memory neural network layer;

[0019] Determining, according to the data types of the sub-data, the feature extraction paths of the sub-data in the classification recommendation model, includes:

[0020] For each of the sub-data, when the data type of the sub-data is spatial data, determining the first feature extraction path as the feature extraction path of the sub-data;

[0021] When the data type of the sub-data is time series data, the second feature extraction path is determined as the feature extraction path of the sub-data.

[0022] In one of the embodiments, the first feature extraction path includes a first feature embedding layer and a convolutional neural network layer; the model usage data of the classification recommendation model includes user social network data;

[0023] The step of inputting the sub-data and the associated data of the sub-data into the classification recommendation model based on the feature extraction path corresponding to the sub-data to obtain the candidate recommendation result corresponding to each candidate object for the target user includes:

[0024] Inputting the user social network data and associated data of the user social network data into the first feature embedding layer to obtain embedded features;

[0025] Using the convolutional neural network layer to perform convolution processing on the embedded features to obtain convolution features;

[0026] Based on the convolutional features, a candidate recommendation result corresponding to each of the candidate objects for the target user is determined.

[0027] In one embodiment, the second feature extraction path includes a positive feature extraction path and a negative feature extraction path; the model usage data of the classification recommendation model includes user positive behavior data and user negative behavior data;

[0028] The step of inputting the sub-data and the associated data of the sub-data into the classification recommendation model based on the feature extraction path corresponding to the sub-data includes:

[0029] Inputting the user positive behavior data and associated data of the user positive behavior data into the classification recommendation model based on the positive feature extraction path;

[0030] The user negative behavior data and associated data of the user negative behavior data are input into the classification recommendation model based on the negative feature extraction path.

[0031] In a second aspect, the present application further provides an object recommendation device, the device comprising:

[0032] A response module, configured to obtain, in response to an object recommendation event triggered for a target user, a user identifier of the target user, a service identifier of the service to be recommended, and a plurality of pre-configured recommendation models;

[0033] A data acquisition module, for acquiring, for each of the recommendation models, model usage data of the recommendation model from a data storage area corresponding to the recommendation model; the model usage data includes user information matching the user identifier and service information matching the service identifier; the service information includes object information of each of a plurality of candidate objects belonging to the service to be recommended;

[0034] A candidate recommendation module, used for inputting the model usage data into a corresponding recommendation model to obtain a candidate recommendation result corresponding to each candidate object for the target user;

[0035] The fusion recommendation module is used to fuse the candidate recommendation results of each of the recommendation models and determine the target recommendation object of the target user from the candidate objects.

[0036] In a third aspect, the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0037] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of the above method when executed by a processor.

[0038] In a fifth aspect, the present application also provides a computer program product, including a computer program, which implements the steps of the above method when executed by a processor.

[0039] The object recommendation method, device, computer equipment, computer-readable storage medium and computer program product described above can obtain the user ID of the target user, the service ID of the service to be recommended, and a plurality of pre-configured recommendation models in response to an object recommendation event triggered for a target user. The recommendation method used by each recommendation model is different. For each recommendation model, the model usage data required by the recommendation model is directly obtained from the data storage area corresponding to the recommendation model. The recommendation model can recommend objects to the target user through the user information and service information contained in the model usage data, and obtain the candidate recommendation results of each recommendation model. Then, the candidate recommendation results of each recommendation model are fused to determine the target recommended object of the target user from the candidate objects. By using multiple recommendation models to recommend objects to the target user based on their respective model usage data, the situation in which the recommendation accuracy is reduced due to the cold start problem when a single model is used can be effectively avoided. By fusion of the candidate recommendation results of each recommendation model to determine the target recommended object in a hybrid recommendation manner, the target recommended object can be more matched with the target user, and the accuracy of object recommendation is effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments of the present application or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0041] FIG1 is a diagram of an application environment of an object recommendation method in some embodiments;

[0042] FIG2 is a schematic diagram of a flow chart of an object recommendation method in some embodiments;

[0043] FIG3 is a schematic diagram of a process of inputting model usage data into a corresponding recommendation model to obtain candidate recommendation results corresponding to each candidate object for a target user in some embodiments;

[0044] FIG4 is a schematic diagram of the structure of a classification recommendation model in some embodiments;

[0045] FIG5 is a schematic diagram of a process of determining a feature extraction path for each sub-data in a classification recommendation model according to the data type of each sub-data in some embodiments;

[0046] 6 is a flow chart of, in some embodiments, inputting sub-data and associated data of the sub-data into a classification recommendation model based on a feature extraction path corresponding to the sub-data to obtain candidate recommendation results corresponding to each candidate object for a target user;

[0047] FIG7 is a diagram of an application environment of an object recommendation method in some other embodiments;

[0048] FIG8 is a schematic diagram of detailed feature distribution stored in a database in some embodiments;

[0049] FIG9 is a schematic diagram of model distribution of a hybrid recommendation model in some embodiments;

[0050] FIG10 is a flowchart of an object recommendation method in some other embodiments;

[0051] FIG11 is a schematic diagram of the structure of a deep learning classification model in some embodiments;

[0052] FIG12 is a structural block diagram of an object recommendation device in some embodiments;

[0053] FIG. 13 is a diagram of the internal structure of a computer device in some embodiments. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application. It should be noted that in the embodiments of the present application, certain software, groups, models and other existing solutions in the industry may be mentioned, which should be considered as exemplary, and their purpose is only to illustrate the feasibility of the implementation of the technical solution of the present application, but it does not mean that the applicant has or will necessarily use the solution.

[0055] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. The acquisition, storage, use, and processing of data in the technical solution of this application comply with the relevant provisions of national laws and regulations.

[0056] The object recommendation method provided in the embodiment of the present application can be applied to the application environment shown in FIG. 1. The recommendation system 102 communicates with the database 104 and the terminal 106 through the network. The data storage system can store the data that the recommendation system 102 needs to process. The data storage system can be integrated on the recommendation system 102, or placed on the cloud or other network servers. The recommendation system 102 obtains the user identification of the target user, the business identification of the business to be recommended, and multiple pre-configured recommendation models in response to the object recommendation event triggered for the target user. For each recommendation model, the model usage data of the recommendation model is obtained from the data storage area corresponding to the recommendation model in the database 104, and the model usage data includes user information matching the user identification and business information matching the business identification. The business information includes object information of multiple candidate objects belonging to the business to be recommended. The recommendation system 102 inputs the model usage data into the corresponding recommendation model, obtains the candidate recommendation results corresponding to each candidate object of the target user, fuses the candidate recommendation results of each recommendation model, and determines the target recommended object of the target user from each candidate object. Subsequently, the recommendation system 102 can send the target recommended object to the terminal 106 and display it to the user based on the terminal 106.

[0057] The recommendation system 102 may be a business system for performing recommendation tasks in a financial institution, such as a business recommendation system of a bank. The recommendation system 102 may be integrated on the terminal 106, or may be an independent server or a server cluster composed of multiple servers, or may be a cloud server providing cloud computing services.

[0058] The database 104 is a data repository for storing financial service related data. Taking a bank as an example, the database 104 can be used to store business data and user information data generated during the bank-related processing.

[0059] Terminal 106 can be a management terminal for business managers or a user terminal for target users. Terminal 106 can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers and portable wearable devices. Portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc.

[0060] In some embodiments, as shown in FIG. 2 , a method for recommending an object is provided, and the method is described by taking the method applied to the recommendation system 102 in FIG. 1 as an example, including the following steps:

[0061] S202 , in response to an object recommendation event triggered for a target user, obtaining a user identifier of the target user, a service identifier of the service to be recommended, and a plurality of pre-configured recommendation models.

[0062] The target user is a business user who has a need for object recommendation. The object recommendation event is a trigger event for indicating that an object is recommended to the target user.

[0063] In some embodiments, the target user can trigger an object recommendation event based on the use of a terminal by the target user or a business manager. For example, the user can use a terminal to enter the recommendation system, click or enter an object recommendation in the recommendation system, and the object recommendation event for the target user can be triggered.

[0064] In some of the embodiments, the recommendation system may automatically trigger an object recommendation event for the target user based on the historical behavior data of the target user, such as business information of related businesses that have been browsed multiple times recently.

[0065] The user ID of the target user is the unique identification information used to mark the target user. Similarly, the service ID is also the unique identification information used to mark the service to be recommended. The service to be recommended refers to the type of service to which the object to be recommended belongs. For example, the service to be recommended may be a fund service, and the candidate objects under the service to be recommended are the fund products under the fund service.

[0066] The recommendation model is a pre-configured model used to recommend objects to target users based on user information and business information. It can be understood that there are multiple recommendation models pre-configured in the recommendation system, and each recommendation model corresponds to a different recommendation method or recommendation algorithm. For example, the recommendation system can pre-configure a rule-based recommendation model, a collaborative filtering recommendation model, a deep learning classification recommendation model, a clustering recommendation model, a reinforcement learning recommendation model, etc., and each recommendation model corresponds to a different recommendation method or recommendation algorithm.

[0067] Specifically, in response to an object recommendation event triggered for a target user, the recommendation system obtains a user identifier of the target user, a service identifier of the service to be recommended, and a plurality of pre-configured recommendation models.

[0068] S204: For each recommendation model, obtain the model usage data of the recommendation model from the data storage area corresponding to the recommendation model.

[0069] The model usage data includes user information matching the user identifier and service information matching the service identifier.

[0070] User information is relevant information data used to characterize user characteristics of a target user. In some embodiments, the user information may include user basic information, user behavior information, user space information, user time information, etc.

[0071] Among them, user basic information may include user gender, age, user social relationship, user population, etc. User social relationship may include association relationship among friends, colleagues, relatives, etc. User behavior information may include user click information, user browsing information, user purchase information, user transaction information, user comment information, user published information, etc. Among them, user click information may include click information generated by click operations on notification bars, text messages, advertisements, etc. of the business interface. User published information may include pictures, texts, voices, videos, etc. published by users in the business system. User spatial information may be information data used to characterize the geographic location of the target user, such as address information. User time information may be information data used to characterize the attribute time of the target user.

[0072] The business information includes object information of each of the multiple candidate objects belonging to the business to be recommended, such as object income classification, object basic attributes, object management personnel attributes, etc. For example, if the business to be recommended is a fund business, and the corresponding candidate objects are fund products under the fund business, the corresponding business information may include product income classification, product basic attributes, product management manager attributes, etc. of each fund product.

[0073] Among them, the data storage area of ​​the recommendation model is the data storage location pre-configured in the database for each recommendation model. Since the model usage data required by each recommendation model may be different, in order to achieve end-to-end object recommendation, the corresponding data storage area can be pre-configured for each recommendation model in the database, and the relevant data required for each recommendation model to perform object recommendation processing is stored in the data storage area corresponding to each recommendation model. The data storage area can be regarded as the data center of the recommendation model. When using the recommendation model for object recommendation, the model usage data can be directly obtained from the corresponding data storage area.

[0074] It is understandable that the model usage data corresponding to different recommendation models may be different. For example, when the recommendation model is a rule-based recommendation model, the corresponding model usage data may be all user information and business information. When the recommendation model is a collaborative filtering model, the corresponding model usage data is the relevant data used to calculate the association between users and objects, such as user behavior information, user social network information, product basic information, etc. When the recommendation model is a clustering recommendation model, the corresponding model usage data may be relevant information data that facilitates clustering.

[0075] Specifically, when the recommendation system obtains the user ID of the target user, the business ID of the business to be recommended, and multiple pre-configured recommendation models, for each recommendation model, it can first determine the corresponding data storage area of ​​the recommendation model in the database, and then obtain the user information matching the user ID and the business information matching the business ID from the data storage area based on the user ID and the business ID, and determine the user information and business information as model usage data for the recommendation model.

[0076] In some embodiments, the object recommendation method further includes: obtaining a model identifier of the recommendation model, and determining a storage area matching the model identifier as a data storage area corresponding to the recommendation model based on a mapping relationship between the model identifier and each storage area of ​​the database.

[0077] The model identifier is the unique identification information used to mark the recommended model. The mapping relationship between the model identifier and each storage area of ​​the database is obtained by the designer binding the model identifier of the recommended model and the data storage area corresponding to each recommended model in the database in advance. The mapping relationship can be used to represent the one-to-one correspondence between the recommended model and the data storage area in the database.

[0078] Specifically, after obtaining multiple recommendation models, the recommendation system can obtain the model identifier of each recommendation model, and then based on the model identifier of the recommendation model, call the pre-configured mapping relationship between the model identifier and each storage area of ​​the database, and determine the storage area in the database that matches the model identifier of the recommendation model as the data storage area of ​​the recommendation model. By pre-setting the mapping relationship between the model identifier and each storage area of ​​the database, it is possible to quickly locate the data storage area for each recommendation model, providing a data basis for subsequent end-to-end object recommendation, and also improving the recommendation efficiency of object recommendation.

[0079] S206, inputting the model usage data into the corresponding recommendation model to obtain the candidate recommendation results corresponding to each candidate object for the target user.

[0080] The candidate recommendation result is the output result obtained by each recommendation model after performing object recommendation processing on the target user based on the corresponding model usage data. The candidate recommendation result can be the recommendable value of each candidate object calculated by each recommendation model, or it can be a single or multiple candidate recommendation objects determined by each recommendation model.

[0081] Specifically, after obtaining the model usage data of each recommendation model, the recommendation system can input each model usage data into the corresponding recommendation model, perform object recommendation processing through each recommendation model, and obtain candidate recommendation results corresponding to each candidate object for the target user.

[0082] In some embodiments, before inputting the usage data of each model into the corresponding recommendation model, the recommendation system may also perform data preprocessing on the usage data of each model by a preset data preprocessing method to obtain data that is more suitable for use by each recommendation model. It is understandable that the data preprocessing method may include data anomaly cleaning, blank data filling, data normalization processing, etc.

[0083] S208, integrating the candidate recommendation results of each recommendation model, and determining the target recommendation object for the target user from the candidate objects.

[0084] Specifically, after obtaining the candidate recommendation results, the recommendation system can fuse the candidate recommendation results of each recommendation model to improve the accuracy of object recommendation, and determine the target recommendation object that best matches the target user from the candidate objects.

[0085] In some of the embodiments, the candidate recommendation results may be the recommendable values ​​of each candidate object. The recommendation system determines the target recommendation value of each candidate object through the recommendable values ​​of each candidate object output by each recommendation model, and determines the candidate object with the largest target recommendation value as the target recommendation object.

[0086] In some embodiments, the candidate recommendation result may be a single or multiple candidate recommendation objects output by the recommendation model. The recommendation system may count the number of occurrences of each candidate recommendation object and determine the candidate recommendation object with the largest number of occurrences as the target recommendation object.

[0087] In some embodiments, a fusion classification recommendation model is also pre-set in the recommendation system. After obtaining the candidate recommendation results output by each recommendation model, the recommendation system can fuse the candidate recommendation results to obtain candidate recommendation data, and input the candidate recommendation data into the pre-set fusion classification recommendation model for object recommendation processing to obtain the final target recommendation object.

[0088] In the above object recommendation method, in response to the object recommendation event triggered for the target user, the user identifier of the target user, the business identifier of the business to be recommended, and multiple pre-configured recommendation models can be obtained. The recommendation method used by each recommendation model is different. For each recommendation model, the model usage data required by the recommendation model is directly obtained from the data storage area corresponding to the recommendation model. The recommendation model can recommend objects to the target user through the user information and business information contained in the model usage data, and obtain the candidate recommendation results of each recommendation model. Then, the candidate recommendation results of each recommendation model are fused to determine the target recommended object of the target user from the candidate objects. By using multiple recommendation models to recommend objects to the target user based on their respective model usage data, the situation where the recommendation accuracy is reduced due to the cold start problem when using a single model can be effectively avoided. By fusing the candidate recommendation results of each recommendation model to determine the target recommended object in a hybrid recommendation manner, the target recommended object can be more matched with the target user, effectively improving the accuracy of object recommendation.

[0089] In order to further improve the accuracy of object recommendation, in some embodiments, the recommendation model may include a classification recommendation model, and the classification recommendation model includes multiple feature extraction paths, and different feature extraction paths correspond to different feature extraction methods.

[0090] As shown in FIG. 3 , S206 , the model usage data is input into the corresponding recommendation model to obtain the candidate recommendation results corresponding to each candidate object for the target user, including:

[0091] S302, determining the feature extraction paths corresponding to the sub-data in the classification recommendation model according to the data types of the sub-data in the model usage data.

[0092] The model usage data can be composed of multiple sub-data, and each sub-data has its own data type. The data type is a type parameter obtained by classifying the sub-data according to the time-related attributes of the sub-data. It can be understood that the data type can include time series data and spatial data. Different sub-data types have different corresponding feature extraction paths in the classification recommendation model.

[0093] The classification recommendation model is a deep learning recommendation model that uses a multi-layer perceptron to classify target users and then uses the model classification results to recommend objects to target users. As shown in Figure 4, the classification recommendation model can include multiple feature extraction paths, and the feature output end of each feature extraction path is connected to the multi-layer perceptron. The extracted features can be input into the multi-layer perceptron for processing, and the output end of the multi-layer perceptron is connected to the fully connected layer. The features processed by the multi-layer perceptron will enter the fully connected layer for recommendation processing. Finally, the output layer connected to the output end of the fully connected layer outputs the final candidate recommendation results of the classification recommendation model.

[0094] The feature extraction path in the classification recommendation model is an extraction path for extracting data features from the input model usage data. The feature extraction path can be composed of a feature embedding layer and a feature extraction layer.

[0095] In some of these embodiments, the feature embedding layer may be a Bert model.

[0096] Specifically, after obtaining the model usage data of the recommendation model, the recommendation system can determine the data type of each sub-data according to the data identifier of each sub-data in the model usage data and the correspondence between the pre-set data identifier and the data type. Then, based on the data type of each sub-data, the feature extraction path corresponding to each sub-data in the classification recommendation model is determined.

[0097] S304: Determine the associated data of each sub-usage data based on the associated relationship between each sub-data.

[0098] Among them, the association relationship between each sub-data can be determined in advance by the designer based on the data characteristics of each sub-data. For example, user basic information can be considered as basic data, and the user basic information will have an association relationship with other sub-data in the model usage data, and the basic attributes of the object can have an association relationship with the user context description information.

[0099] Specifically, after determining the feature extraction paths corresponding to each sub-data in the classification recommendation model, the recommendation system also needs to call the association relationship between each sub-data, and determine the associated data of each sub-data based on the association relationship between each sub-data.

[0100] S306 , for each sub-data, the sub-data and the associated data of the sub-data are input into a classification recommendation model based on the feature extraction path corresponding to the sub-data to obtain candidate recommendation results corresponding to each candidate object for the target user.

[0101] Specifically, for each sub-data, the recommendation system not only needs to input the sub-data into the classification recommendation model based on the feature extraction path corresponding to the sub-data, but also needs to input the associated data of the sub-data together with the sub-data, based on the feature extraction path corresponding to the sub-data, into the classification recommendation model, so that when making object recommendations, the recommendation accuracy will not be reduced due to the isolation of features. The classification recommendation model will recommend objects for the target user based on the data features extracted by the feature extraction path, and obtain the candidate recommendation results corresponding to each candidate object for the target user.

[0102] Taking the model usage data including sub-data A, B, and C as an example, if A and B are related data, and B is related data with A and C, the recommendation system needs to input A and B into the feature extraction path corresponding to A at the same time, input A, B, and C into the feature extraction path corresponding to B at the same time, and input B and C into the feature extraction path corresponding to C at the same time for feature extraction.

[0103] In the above embodiment, by determining the corresponding feature extraction path for each sub-data according to the data type of each sub-data, the accuracy of data feature extraction can be improved. By inputting the sub-data and the corresponding associated data into the feature extraction path corresponding to the sub-data for feature extraction, a certain correlation can be created between the obtained data features, thereby avoiding the problem of reduced recommendation accuracy due to isolation of features.

[0104] Furthermore, in some embodiments, the classification recommendation model includes a first feature extraction path based on a convolutional neural network layer and a second feature extraction path based on a long short-term memory neural network layer.

[0105] As shown in FIG5 , S302, according to the data types of the sub-data, determining the feature extraction paths of the sub-data in the classification recommendation model, includes:

[0106] S502: for each sub-data, when the data type of the sub-data is spatial data, determine the first feature extraction path as the feature extraction path of the sub-data.

[0107] Among them, the convolutional neural network is a feedforward neural network. By using the first feature extraction path composed of convolutional neural network layers, it has better feature extraction capabilities for data with similar grid structures. It generally uses convolution operations during calculation, which can effectively extract data features and reduce local feature dimensions.

[0108] Spatial data can also be considered as spatiotemporal data, which refers to data generated over time at a specific spatial location and can describe the information of objects in space.

[0109] Specifically, for each sub-data, when the data type of the sub-data is determined to be spatial data, the recommendation model can determine the first feature extraction path composed of a convolutional neural network layer as the feature extraction path of the sub-data to ensure the effective extraction of data features of the spatial data.

[0110] S504: When the data type of the sub-data is time series data, determine the second feature extraction path as the feature extraction path of the sub-data.

[0111] Among them, the long short-term memory neural network (LSTM) is a special recursive neural network. Different from the general feedforward neural network, LSTM can use time series to analyze the input. Therefore, the second feature extraction path composed of the long short-term memory neural network layer is suitable for corresponding feature extraction of time-series data.

[0112] Time series data refers to data that changes over time. This type of data is usually recorded at certain time intervals and has time series information.

[0113] Specifically, when the recommendation system determines that the data type of the sub-data is time series data, it can determine the second feature extraction path composed of the long short-term memory neural network layer as the feature extraction path of the sub-data to ensure the effective extraction of data features of the time series data.

[0114] In this embodiment, by configuring corresponding feature extraction paths for sub-data of different data types, data features of sub-data of various data types can be effectively extracted, which effectively improves the recommendation accuracy of subsequent classification recommendation models when recommending objects to target users.

[0115] In some embodiments, the first feature extraction path includes a first feature embedding layer and a convolutional neural network layer. The model usage data of the classification recommendation model includes user social network data.

[0116] As shown in FIG. 6 , in S306 , the sub-data and the associated data of the sub-data are input into the classification recommendation model based on the feature extraction path corresponding to the sub-data to obtain the candidate recommendation results corresponding to each candidate object for the target user, including:

[0117] S602: Input the user social network data and the associated data of the user social network data into a first feature embedding layer to obtain embedded features.

[0118] Among them, user social network data may include the number of relationship information between the target user and other users, such as friend relationships, family relationships, relationships in the same industry, school relationships, social media relationships, investor group relationships, business relationships, etc., and may also include user theme information, user pain point information, investment goals, etc. of the target user. User theme information can be information obtained by summarizing the theme of the user story. For example, if the user's main story is: the target user is a newly graduated white-collar worker with a busy life and work, who does not know much about the fund industry, but hopes to manage his finances through funds so that the funds can grow reasonably, then the user theme information can be summarized as "graduate, busy worker, fund novice", user needs: stable financial management, user pain points: introductory knowledge of funds, etc.

[0119] Specifically, when the recommendation system inputs user social network data, since the data type of the user social network data is spatial data, the user social network data and the associated data of the user social network data, such as user basic information, can be input into the first feature embedding layer in the first feature extraction path for processing to obtain embedded features.

[0120] S604, using a convolutional neural network layer to perform convolution processing on the embedded features to obtain convolution features.

[0121] Specifically, the embedded features output by the first embedding layer will be used as the input of the convolutional neural network layer and input into the convolutional neural network layer connected to the first embedding layer. The recommendation system continues to use the convolutional neural network layer to perform convolution processing on the embedded features to obtain convolutional features.

[0122] S606: Determine the candidate recommendation results corresponding to each candidate object for the target user based on the convolutional features.

[0123] Specifically, the classification recommendation model in the recommendation system recommends objects to the target user based on the convolutional features, and can determine the candidate recommendation results corresponding to each candidate object for the target user.

[0124] In some of the embodiments, the convolution features are input into the multi-layer perceptron of the classification recommendation model, and processed in the multi-layer perceptron together with the data features extracted by other feature extraction paths to obtain processed features. The processed features are then input into the fully connected layer of the classification recommendation model for classification recommendation processing to obtain the candidate recommendation results for the target user corresponding to each candidate object.

[0125] In the above embodiment, by using the first feature embedding layer and the convolutional neural network layer to form a first feature extraction path, effective feature extraction can be performed on the user social network data and the associated data of the user social network data, thereby improving the accuracy of object recommendation.

[0126] In some embodiments, the second feature extraction path includes a positive feature extraction path and a negative feature extraction path. The model usage data of the classification recommendation model includes user positive behavior data and user negative behavior data. The sub-data and the associated data of the sub-data are input into the classification recommendation model based on the feature extraction path corresponding to the sub-data, including: inputting the user's positive behavior data and the associated data of the user's positive behavior data into the classification recommendation model based on the positive feature extraction path. Inputting the user's negative behavior data and the associated data of the user's negative behavior data into the classification recommendation model based on the negative feature extraction path.

[0127] Among them, the positive feature extraction path is an extraction path used to extract features from user positive behavior data and associated data related to user positive behavior data. By extracting features from user positive behavior data and associated data of user positive behavior data, user positive features can be obtained.

[0128] In some embodiments, the forward feature extraction path includes a second feature embedding layer and a first long short-term memory neural network layer. The output of the second feature embedding layer is connected to the input of the first long short-term memory neural network layer.

[0129] User positive behavior data refers to the behavior data generated when users perform positive behaviors on the recommended business, which can be used to reflect the positive characteristics of users. Taking the recommended business as a fund business as an example, user positive behavior data may include following, comparing, searching, investing, purchasing, redeeming, checking fund information, checking account balances and positions, receiving reminders and notifications, checking market trends and news, participating in community discussions or exchanges, using investment tools and calculations, etc.

[0130] Among them, the negative feature extraction path is an extraction path used to extract features from the user's negative behavior data and the associated data related to the user's negative behavior data. By extracting features from the user's negative behavior data and the associated data of the user's negative behavior data, the user's negative features can be obtained.

[0131] In some of these embodiments, the negative feature extraction path includes a third feature embedding layer and a second long short-term memory neural network layer, and the output end of the third feature embedding layer is connected to the input end of the second long short-term memory neural network layer.

[0132] User negative behavior data refers to the behavior data generated when users perform negative behaviors on the recommended business, which can be used to reflect the negative characteristics of users. Taking the recommended business as a fund business as an example, user negative behavior data may include ignoring the push of funds, not being interested in clicking on fund knowledge, etc.

[0133] Specifically, the recommendation system can input the user's positive behavior data and the associated data of the user's positive behavior data into the classification recommendation model based on the positive feature extraction path, and input the user's negative behavior data and the associated data of the user's negative behavior data into the classification recommendation model based on the negative feature extraction path.

[0134] In the above embodiment, by inputting the user's positive behavior data and its associated data, and the user's negative behavior data and its associated data into the classification recommendation model through different feature extraction paths, the user's positive behavior characteristics and the user's negative behavior characteristics can be distinguished, thereby avoiding the situation where the accuracy of object recommendation is reduced due to feature confusion.

[0135] In some embodiments, an object recommendation method is provided, which is applied to an end-to-end recommendation architecture as shown in FIG7 , wherein the recommendation system is connected to the database and the terminal in communication, and can obtain the corresponding data from the database for model calculation and output to the terminal, while the database is also connected to the terminal in communication, and can collect the corresponding business data from the terminal. The terminal is user-oriented and can display the recommendation results generated by the recommendation system to the user.

[0136] Among them, the database is mainly used to store user and product information, and the detailed feature distribution is shown in Figure 8.

[0137] The recommendation system is pre-set with a hybrid recommendation model. The model distribution diagram of the hybrid recommendation model is shown in Figure 9. It can be seen that the hybrid recommendation model specifically consists of a parallel rule-based recommendation model, a collaborative filtering recommendation model, a deep learning classification recommendation model, a clustering recommendation model, a reinforcement learning recommendation model, and a fusion classification recommendation model connected in series with each parallel model. The database can be regarded as the overall data center, and each parallel model is connected to the data center, that is, each parallel model has a corresponding data storage area in the database.

[0138] As shown in FIG10 , the object recommendation method specifically includes the following steps:

[0139] First, in response to an object recommendation event triggered for a target user, the recommendation system obtains a user ID of the target user and a service ID of the service to be recommended.

[0140] The usage data of each recommendation model is obtained from the data center according to the user identifier of the target user and the service identifier of the service to be recommended, and the usage data is preprocessed to obtain the model usage data of each recommendation model.

[0141] The recommendation system inputs the usage data of each model into the corresponding recommendation model for calculation, and obtains the candidate recommendation results output by each recommendation model.

[0142] Specifically, the rule-based recommendation model only needs to match product information with user information, and recommend corresponding popular products according to the user's acceptable product risk level. The fund product risk level is divided into low, relatively low, medium, relatively high, and extremely high categories according to R1-R5. Users classify the acceptability level according to the risk level of fund products, and rank the fund products in each category according to factors such as fund managers, yields, investment industries, and fund investment allocations. Then, fund products with a risk level less than or equal to the acceptable level are recommended to users.

[0143] The collaborative filtering recommendation model uses the association between users and products to calculate the product with the highest score for recommendation. This embodiment introduces additional user online behavior features and user social network features to mine similarities between users and improve accuracy. The feature matrix after the introduction of features is as follows:

[0144]

[0145] Among them, Rnn represents the rating of Usern to Itemn.

[0146] Then, the similarity between users is calculated by weighted cosine similarity and Pearson correlation coefficient, and the similarity of text, speech, and image embedding is also calculated. After calculation, similar user groups are obtained, and then the products with the highest user ratings in the user group are selected. The weighted user similarity calculation formula is as follows:

[0147] =W0·C sim +W1·P sim +W2·M sim

[0148] =W0·C sim +W1·P sim +W2·M sim ;

[0149] Among them, W u,s represents the similarity between user u and user s, W i This similarity represents the weight, C sim Indicates the cosine similarity of the feature. The cosine similarity can reflect the cosine value of the angle of the feature, and the Pearson correlation coefficient can reflect the linear correlation of the feature vector. sim Represents the normalized value of Minkowski distance. Minkowski distance is a generalization of Euclidean distance and Manhattan distance. Euclidean distance is the linear distance of vectors, and Manhattan distance is the distance between calculated vectors. The similarity between users is obtained by weighted calculation of this horizontal feature similarity. Then the score of User to Item is calculated by the following formula:

[0150]

[0151] Among them, W u,s The similarity between user u and user s, R s,p is the score of user S on item P, H is the set of positive and negative feedback items of the target user, and W p,h is the similarity between items, R u,h is the rating of item h by user u.

[0152] The deep learning classification recommendation model is a recommendation model that uses a multi-layer perceptron to classify users after preprocessing the data in the data center, and uses the model classification results to make recommendations for users. Since the above recommendation algorithm will face the cold start problem, when only a small amount of user feature data is obtained, the user's social network relationship characteristics and user behavior characteristics are added and then classified using a multi-layer perceptron. Then the classification results are used for recommendation. The number of layers and neurons of the multi-layer perceptron can be determined by the number of features obtained and some heat products. The entire model structure diagram is shown in Figure 11.

[0153] Social network characteristics include the characteristics and quantity of relationships with other users, such as friendships, family relationships, industry relationships, school relationships, social media relationships, investor group relationships, and business relationships. Social network characteristics also include user theme stories, user pain points, and investment goals.

[0154] By building user social circles and user basic characteristics, we can filter out the real similar groups of users. User basic characteristics include age, gender, city, occupation, income level, education level, marital status, and family background. Contextual characteristics include: Investment goals: retirement savings, education funds. Investment knowledge level: novice, entry-level, general investor, investment veteran, investment expert. Risk preference, the degree of risk that can be accepted in investment. Investment preference: the degree of preference for product varieties. Investment period: short-term, medium-term, and long-term. Financial management goals: financial freedom, asset appreciation, and personal interests. Preferred channels: Internet, mobile applications, and traditional banks. Financial product preferences: funds, stocks, and bonds. Investment asset scale: trial, small capital, medium capital, large amount, and huge amount. Changes in user investment preferences: record user investment preferences and motivations. Investment history: investment experience and investment returns. Investor sources: personal investment and, machine purchase investors, and fund manager investors. Investor concerns: economic indicators, industry trends, and company performance. Sources of investment funds: salary income, other income, financial product income, and stock investment returns. Investment information channels: financial news, professional analyst reports, social media, official websites, and recommendations from friends. Positive user behavior characteristics: following, comparing, searching, investing, purchasing, redeeming, checking fund information, checking account balances and positions, receiving reminders and notifications, checking market trends and news, participating in community discussions or exchanges, and using investment tools and calculations. Negative user behavior characteristics: ignoring pushed funds. User emotional characteristics: positive, neutral, and negative (derived from user comments, pictures, and videos).

[0155] Then, Bert is used as a feature embedding model to build a feature dictionary for the business to be recommended, such as the fund industry, and initialize the embedding matrix. The entire structure can be divided into three modules. The left module is the characteristics of users and social relationships. This structure uses the social network characteristics of users to find similar groups of users as additional information input. The middle module combines user behavior characteristics, basic attribute characteristics, product characteristics, positive behavior characteristics, and emotional characteristics with user context characteristics to extract user interest products in a fine-grained manner. The right module extracts user negative feedback products. These two modules can understand user behavior patterns and habits, and have a more comprehensive understanding of user interests, preferences, and behavior habits. Then, after combining user positive and negative feedback, context, fund characteristics, and social network characteristics, a multi-layer perceptron is input for fine-grained calculation, and finally FC is used to calculate user preference fund products.

[0156] The clustering recommendation model is used to recommend users' features and then recommend popular products in the user cluster. The clustering recommendation model can not only cope with daily recommendations, but also improve the accuracy of clustering during cold start recommendations by adding users' spatiotemporal characteristics and social relationship characteristics, and can cluster users into user clusters more accurately.

[0157] In the reinforcement learning recommendation model, the definition of reinforcement learning recommendation is as follows:

[0158] State s represents the user's behavior characteristics towards fund products, such as click, follow, buy, ignore, sell, and invest regularly, and the state space is S. Action α represents the recommendation system recommending products, and the action space is Α. Strategy π(α, s) represents the function of the recommendation system recommending products to users based on user feedback, such as fund products. The state transition probability ρ(s'|α) is the probability that the user goes from s to s' after the recommendation system takes action α to the user.

[0159] Reward function: After the recommendation system takes action α to the user, when the user moves from s to s', the system should be given a reward r. The reward function is as follows:

[0160]

[0161] in, is the discount factor, and T is the set of user behaviors.

[0162] The objective function of the recommendation system can be defined as the product of ρ(s'|α) and the reward function:

[0163]

[0164] in, is the expectation of the Q function with respect to action α.

[0165] The recommendation system is to find an optimal strategy to get the highest reward. The target Q value can be trained using DQN. Reinforcement learning uses user behavior characteristics to find the strategy that the recommendation system should use to achieve the optimal solution of action-reward. Then find the products preferred by users.

[0166] After obtaining the above recommendation results, the candidate recommendation results output by each recommendation model are fused using model stacking, and then the fused recommendation results are input into the pre-trained fusion classification recommendation model, and the final prediction is made through meta-learning. The accuracy of the entire hybrid recommendation will be improved through Model stacking.

[0167] This embodiment proposes an end-to-end object hybrid recommendation method, including data collection, data storage, model calculation and result display in the entire process. Among them, the hybrid recommendation algorithm combines rules, collaborative filtering, clustering recommendation, deep learning recommendation, and reinforcement learning recommendation algorithm. Strengthening the weaknesses of a single algorithm in certain aspects, rule-based recommendation can better cope with cold start algorithms. Recommendations based on clustering and collaborative filtering can better find similar groups of users. Deep learning recommendation uses black box models to improve accuracy, and reinforcement learning makes better use of user behavior characteristics to improve accuracy. Finally, multiple recommendation results are aggregated by model stacking to obtain the final recommended products. Among them, in the model based on deep learning recommendation, a recommendation model based on social network and positive and negative interest model is proposed. The user stories, user needs, and user pain points of user portraits are also integrated into social networks and context features. User interest products are calculated through user behavior characteristics. Finally, the three are output to the multi-layer perceptron to calculate the recommended products.

[0168] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0169] Based on the same inventive concept, the embodiment of the present application also provides an object recommendation device for implementing the object recommendation method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more object recommendation device embodiments provided below can refer to the limitations on the object recommendation method above, and will not be repeated here.

[0170] In some embodiments, as shown in FIG. 12 , an object recommendation device 1200 is provided, including: a response module 1201, a data acquisition module 1202, a candidate recommendation module 1203 and a fusion recommendation module 1204, wherein:

[0171] The response module 1201 is used to obtain the user identifier of the target user, the service identifier of the service to be recommended, and a plurality of pre-configured recommendation models in response to an object recommendation event triggered for the target user.

[0172] The data acquisition module 1202 is used to obtain the model usage data of the recommendation model from the data storage area corresponding to the recommendation model for each recommendation model; the model usage data includes user information matching the user identifier and business information matching the business identifier; the business information includes the object information of each of the multiple candidate objects belonging to the business to be recommended.

[0173] The candidate recommendation module 1203 is used to input the model usage data into the corresponding recommendation model to obtain the candidate recommendation results corresponding to each candidate object for the target user.

[0174] The fusion recommendation module 1024 is used to fuse the candidate recommendation results of each recommendation model and determine the target recommendation object for the target user from the candidate objects.

[0175] In some embodiments, the object recommendation device also includes: a storage area determination module, which is used to obtain a model identifier of the recommendation model; based on the mapping relationship between the model identifier and each storage area of ​​the data center, the storage area matching the model identifier is determined as the data storage area corresponding to the recommendation model.

[0176] In some embodiments, the recommendation model includes a classification recommendation model, which includes multiple feature extraction paths, and different feature extraction paths correspond to different feature extraction methods; the candidate recommendation module includes:

[0177] A feature extraction path determination unit, used to determine the feature extraction path corresponding to each sub-data in the classification recommendation model according to the data type of each sub-data in the model usage data;

[0178] A related data determination unit, used for determining the related data of each sub-usage data based on the association relationship between each sub-data;

[0179] The model input unit is used to input the sub-data and the associated data of the sub-data into the classification recommendation model based on the feature extraction path corresponding to the sub-data for each sub-data, so as to obtain the candidate recommendation results corresponding to each candidate object for the target user.

[0180] In some embodiments, the classification recommendation model includes a first feature extraction path based on a convolutional neural network layer and a second feature extraction path based on a long short-term memory neural network layer. The feature extraction path determination unit is used to, for each usage data, determine the first feature extraction path as the feature extraction path of the sub-data when the data type of the sub-data is spatial data; and determine the second feature extraction path as the feature extraction path of the sub-data when the data type of the sub-data is time series data.

[0181] In some embodiments, the first feature extraction path includes a first feature embedding layer and a convolutional neural network layer; the model usage data of the classification recommendation model includes user social network data. The model input unit is used to input the user social network data and the associated data of the user social network data into the first feature embedding layer to obtain embedded features; based on the convolutional features obtained by convolution processing the embedded features using the convolutional neural network layer, determine the candidate recommendation results corresponding to each candidate object for the target user.

[0182] In some embodiments, the second feature extraction path includes a positive feature extraction path and a negative feature extraction path; the model usage data of the classification recommendation model includes user positive behavior data and user negative behavior data. The model input unit is used to input the user positive behavior data and the associated data of the user positive behavior data into the classification recommendation model based on the positive feature extraction path. The user negative behavior data and the associated data of the user negative behavior data are input into the classification recommendation model based on the negative feature extraction path.

[0183] Each module in the above-mentioned object recommendation device can be implemented in whole or in part by software, hardware or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each module.

[0184] In an exemplary embodiment, a computer device is provided, which may be a server integrated with a recommendation system, and its internal structure diagram may be shown in FIG13. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data such as user identification, business identification, multiple recommendation models, model usage data, candidate recommendation results and target recommendation objects. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, an object recommendation method is implemented.

[0185] Those skilled in the art will understand that the structure shown in FIG. 13 is merely a block diagram of a partial structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different arrangement of components.

[0186] In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, specific implementation steps of the above-mentioned object recommendation method embodiment are implemented.

[0187] In some embodiments, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the specific implementation steps of the above-mentioned object recommendation method embodiment are implemented.

[0188] In some embodiments, a computer program product is provided, including a computer program, which, when executed by a processor, implements the specific implementation steps of the above-mentioned object recommendation method embodiment.

[0189] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., but are not limited to this.

[0190] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0191] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. An object recommendation method, characterized in that: The method comprises: In response to an object recommendation event triggered for a target user, obtaining a user identifier of the target user, a service identifier of a service to be recommended, and a plurality of pre-configured recommendation models; For each of the recommendation models, obtaining model usage data of the recommendation model from a data storage area corresponding to the recommendation model; the model usage data includes user information matching the user identifier and service information matching the service identifier; the service information includes object information of each of a plurality of candidate objects belonging to the service to be recommended; Inputting the model usage data into a corresponding recommendation model to obtain a candidate recommendation result corresponding to each candidate object for the target user; The candidate recommendation results of each of the recommendation models are integrated, and the target recommendation object of the target user is determined from the candidate objects.

2. The method according to claim 1, characterized in that The method further comprises: Obtaining a model identifier of the recommendation model; Based on the mapping relationship between the model identifier and each storage area of ​​the database, the storage area matching the model identifier is determined as the data storage area corresponding to the recommendation model.

3. The method according to claim 1 or 2, characterized in that: The recommendation model includes a classification recommendation model, and the classification recommendation model includes multiple feature extraction paths, and different feature extraction paths correspond to different feature extraction methods; The step of inputting the model usage data into a corresponding recommendation model to obtain a candidate recommendation result corresponding to each candidate object for the target user includes: Determining, according to the data types of the sub-data in the model usage data, the feature extraction paths corresponding to the sub-data in the classification recommendation model; Determine the associated data of each of the sub-usage data based on the associated relationship between the sub-data; For each of the sub-data, the sub-data and the associated data of the sub-data are input into the classification recommendation model based on the feature extraction path corresponding to the sub-data to obtain the candidate recommendation results corresponding to each of the candidate objects for the target user.

4. The method according to claim 3, characterized in that The classification recommendation model includes a first feature extraction path based on a convolutional neural network layer and a second feature extraction path based on a long short-term memory neural network layer; Determining, according to the data types of the sub-data, the feature extraction paths of the sub-data in the classification recommendation model, includes: For each of the sub-data, when the data type of the sub-data is spatial data, determining the first feature extraction path as the feature extraction path of the sub-data; When the data type of the sub-data is time series data, the second feature extraction path is determined as the feature extraction path of the sub-data.

5. The method according to claim 4, characterized in that The first feature extraction path includes a first feature embedding layer and a convolutional neural network layer; the model usage data of the classification recommendation model includes user social network data; The step of inputting the sub-data and the associated data of the sub-data into the classification recommendation model based on the feature extraction path corresponding to the sub-data to obtain the candidate recommendation result corresponding to each candidate object for the target user includes: Inputting the user social network data and associated data of the user social network data into the first feature embedding layer to obtain embedded features; Using the convolutional neural network layer to perform convolution processing on the embedded features to obtain convolution features; Based on the convolutional features, a candidate recommendation result corresponding to each of the candidate objects for the target user is determined.

6. The method according to claim 4, characterized in that The second feature extraction path includes a positive feature extraction path and a negative feature extraction path; the model usage data of the classification recommendation model includes user positive behavior data and user negative behavior data; The step of inputting the sub-data and the associated data of the sub-data into the classification recommendation model based on the feature extraction path corresponding to the sub-data includes: Inputting the user positive behavior data and associated data of the user positive behavior data into the classification recommendation model based on the positive feature extraction path; The user negative behavior data and associated data of the user negative behavior data are input into the classification recommendation model based on the negative feature extraction path.

7. An object recommendation device, characterized in that: The device comprises: A response module, configured to obtain, in response to an object recommendation event triggered for a target user, a user identifier of the target user, a service identifier of the service to be recommended, and a plurality of pre-configured recommendation models; A data acquisition module, for acquiring, for each of the recommendation models, model usage data of the recommendation model from a data storage area corresponding to the recommendation model; the model usage data includes user information matching the user identifier and service information matching the service identifier; the service information includes object information of each of a plurality of candidate objects belonging to the service to be recommended; A candidate recommendation module, used for inputting the model usage data into a corresponding recommendation model to obtain a candidate recommendation result corresponding to each candidate object for the target user; The fusion recommendation module is used to fuse the candidate recommendation results of each of the recommendation models and determine the target recommendation object of the target user from the candidate objects.

8. The device according to claim 7, characterized in that The device also includes: The storage area determination module is used to obtain the model identification of the recommended model; based on the mapping relationship between the model identification and each storage area of ​​the data center, determine the storage area matching the model identification as the data storage area corresponding to the recommended model.

9. The device according to claim 7 or 8, characterized in that The recommendation model includes a classification recommendation model, and the classification recommendation model includes multiple feature extraction paths, and different feature extraction paths correspond to different feature extraction methods; The candidate recommendation module includes: A feature extraction path determination unit, used to determine the feature extraction path corresponding to each sub-data in the classification recommendation model according to the data type of each sub-data in the model usage data; A related data determining unit, configured to determine the related data of each of the sub-usage data based on the related relationship between the sub-data; The model input unit is used to input the sub-data and the associated data of the sub-data into the classification recommendation model for each of the sub-data based on the feature extraction path corresponding to the sub-data, so as to obtain the candidate recommendation results corresponding to each of the candidate objects for the target user.

10. The device according to claim 9, characterized in that The classification recommendation model includes a first feature extraction path based on a convolutional neural network layer and a second feature extraction path based on a long short-term memory neural network layer; The feature extraction path determination unit is used to determine the first feature extraction path as the feature extraction path of the sub-data for each of the usage data when the data type of the sub-data is spatial data; and to determine the second feature extraction path as the feature extraction path of the sub-data when the data type of the sub-data is time series data.

11. The device according to claim 10, characterized in that The first feature extraction path includes a first feature embedding layer and a convolutional neural network layer; the model usage data of the classification recommendation model includes user social network data; The model input unit is used to input the user social network data and the associated data of the user social network data into the first feature embedding layer to obtain embedded features; Based on the convolution features obtained by convolutionally processing the embedded features using the convolutional neural network layer, the candidate recommendation results corresponding to each of the candidate objects for the target user are determined.

12. The device according to claim 10, characterized in that The second feature extraction path includes a positive feature extraction path and a negative feature extraction path; the model usage data of the classification recommendation model includes user positive behavior data and user negative behavior data; The model input unit is used to input the user's positive behavior data and the associated data of the user's positive behavior data into the classification recommendation model based on the positive feature extraction path; and to input the user's negative behavior data and the associated data of the user's negative behavior data into the classification recommendation model based on the negative feature extraction path.

13. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

15. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images