A universal TOP-N intelligent recommendation system and method

Abstract

The present invention discloses a general TOP‑N intelligent recommendation system and method, including an application layer, a service layer, a data layer and a storage layer. The recommendation method includes recall (initialization of recommendation result generation), correction (filtering / supplementing), personalized sorting (result ranking), and recommendation display (final recommendation result). The present invention can support the rapid deployment of various recommendation scenario solutions. By using the general recommendation engine capabilities, the engine-based model building capabilities and algorithm strategy configuration capabilities, it can form a recommendation model at the scenario application level after accessing user data, user behavior data and item data, and can perform model management and indicator observation and analysis through the user interaction interface, select different recall and filtering recommendation algorithms, and generate APIs for each recommendation system scenario for use by the mobile phone APP end or the vehicle-mounted system, so as to realize end-to-end recommendation management and application.

Description

Technical Field

[0001] The present invention belongs to the technical field of vehicle networking and automobile intelligent cockpit, and more specifically relates to a universal TOP-N intelligent recommendation system and method. Background Art

[0002] The recommendation system analyzes the user's historical behavior and models the user's interests, so as to actively recommend information that may meet their needs. In the era of information overload, the task of the recommendation system is to connect users and information, help users find valuable information for themselves, and at the same time make the information available to users who are interested in it, so as to achieve a win-win situation for information consumers and information producers. As a closed personal space, the car has become an indispensable part of various business scenarios in the car's intelligent cockpit based on personalized recommendation systems, such as car product recommendations, music recommendations, and article recommendations. Deploying a new intelligent recommendation system often requires huge manpower, time and energy, and the current mainstream recommendation system is not universal and cannot meet the needs of a single system to adapt to multiple recommendation application scenarios.

[0003] For example, my country's patent CN201610643458.8 discloses an online taxi recommendation method based on collaborative filtering, which includes the following steps: 1) raw data preprocessing, the actual GPS data needs a preprocessing process, mainly the extraction of the starting point and the end point of the passenger; 2) gridding, including regional filtering, grid segmentation, and user filtering; 3) recommendation algorithm, introduce the collaborative filtering algorithm, use the taxibased algorithm and the latent factor model to give the recommendation results, and construct a reasonable scoring matrix; 4) evaluation and result application: use the traditional topN recommendation for verification, and the indicators are accuracy and recall. The purpose of this invention is to analyze the GPS data of taxis in real time, construct a scoring matrix that describes the behavior of taxis through the analysis of the running status, and provide personalized recommendations for taxis to find empty passengers. The advantage of the present invention is that the traditional collaborative filtering method is introduced, the recommendation is personalized, and the model based on mobile characteristics is more in line with reality. However, it only applies a single recommendation scenario, and the recommendation algorithm is single and fixed and has no flexibility.

[0004] Another example is my country's patent CN201610083749.6, which discloses a topN recommendation method based on a cloud model for social networks, belonging to the field of data mining and information retrieval. The method uses a massive data set of social networks to collect recommended user information, score item information, and recommend user information; extract user attributes and obtain user scores from recommended user information and recommending user information, and generate a multidimensional cloud by a multidimensional cloud reverse cloud generator; calculate the similarity of user attribute clouds through multidimensional cloud similarity; generate a score cloud from the scores of the recommending user and the recommended user, merge all score clouds to generate a parent cloud, and generate a new score in the parent cloud; combine similarity and score to generate the interest of the recommended user in the new item, and take the top N items with high interest to generate a topN recommendation set. This method solves the obstacles brought to the recommendation system by the uncertainty of setting the weights of single or multiple attributes, and generates different recommendation sets for different users for personalized recommendations, and sends the recommendation-related information to the user, saving network resources. However, it does not establish a background management platform for the recommendation system to realize visual operations on the interface such as algorithm model selection and data training. Summary of the invention

[0005] To solve the above problems, the present invention provides a universal TOP-N intelligent recommendation system and method, which can support the rapid deployment of various recommendation scenario solutions. By using the universal recommendation engine capabilities, the engine-based model building capabilities and the algorithm strategy configuration capabilities, it can form a recommendation model at the scenario application level after accessing user data, user behavior data and item data, and can manage the model and observe and analyze indicators through the user interaction interface, select different recall and filtering recommendation algorithms, and generate APIs for various recommendation system scenarios for use by mobile APP or vehicle-mounted systems, thereby realizing end-to-end recommendation management and application.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is as follows: a general TOP-N intelligent recommendation system, comprising:

[0007] The application layer includes the recommendation scenario module, user data module, item data module and system management module, which are used to provide a management platform and a visual interactive interface for operation and display;

[0008] The service layer includes the recommendation engine, model and algorithm management, and monitoring modules. It is used to process data through various services, train the recommendation model based on the data processing results, provide interfaces corresponding to multiple recommendation scenarios, and monitor and manage the recommendation model and algorithm.

[0009] The data layer includes business data, model data, system logs, and metadata; it is used to store model training data, model files, and logs on the big data platform;

[0010] Storage layer, used to save data.

[0011] As an optimization, the recommendation scenario module includes data selection, model selection, cold start strategy selection, model management and model hosting environment, which are used to build and implement recommendation scenarios;

[0012] Data selection, used to select the training data required by the recommendation model corresponding to each recommendation scenario through offline or online multi-source methods, the training data including user data, behavior data, and item data;

[0013] Model selection, which is used to select a corresponding algorithm from the algorithm library as the algorithm of the recommendation model. The recommendation algorithm includes personalized recommendation, popular recommendation, and similar recommendation;

[0014] Cold start strategy selection, used to select a strategy for a recommendation model from the cold start strategies, wherein the cold start strategies include a hot recommendation strategy and a time-ordered recommendation strategy;

[0015] Model management is used to configure the update cycle of the recommendation model, including immediate training or periodic training. It is also used to count the AUC offline index and the exposure, click count, and click rate online indexes of the recommendation model, and count the number of calls and call success rate of the recommendation model.

[0016] The model hosting environment is used to specify the service environment where the recommendation model is located, which includes a test environment or a formal environment, and configure relevant permissions based on the service environment.

[0017] As an optimization, the user data module includes a user data unit, a user data table management unit, a behavior data unit and a behavior data table management unit, which are used to access, display and manage user information and behavior information; wherein,

[0018] User data unit, used for searching, querying, previewing and visualizing user data;

[0019] User data table management unit, used to add, delete, modify and query user data;

[0020] Behavior data unit, used for searching, querying, previewing, and visualizing behavior data;

[0021] The behavior data management unit is used to add, delete, modify and query behavior data.

[0022] As an optimization, the item data module includes an item library list, item library management, and item classification management, which are used to access, display, and manage item information;

[0023] Item library list, used to visualize the multi-cascade relationship of the item library;

[0024] Item library management, used to add, delete, modify and check the item library;

[0025] Item classification management is used to add, delete, modify and check item classifications.

[0026] As an optimization, the system management module includes log management and alarm management, which is used to perform operation and maintenance management on the system;

[0027] Log management, which is used to manage the system backend to support the management of online and offline service logs, data and instance logs;

[0028] Alarm management is used to provide alarms for model training, data consistency, and recommendation system status.

[0029] As an optimization, the recommendation engine includes recall service, filtering / supplementing service, sorting service and recommendation service; it is used to support the periodic training generation model of the model and apply the model to the interface provided by the external service; wherein,

[0030] Recall service, used to filter recommended items and generate a candidate list;

[0031] Filtering / supplementing service, used to add and filter the candidate list generated by the recall service through the setting of blacklist and whitelist;

[0032] Sorting service, which is used to generate a user-item-behavior feature matrix based on user information, behavior information, and item information, and to make predictions on the feature matrix;

[0033] Recommendation service is used to call recall service, filtering / supplementing service, and sorting service through the grpc protocol, and provide HTTP request interface and HTTP response interface for external services.

[0034] As an optimization, the model and algorithm management includes a model library, an algorithm library and a model life cycle, which are used to provide model data management, algorithm management, the life cycle of trained models, versions, and online and offline management. The model library comes from the model generated by the recommendation engine training, and the algorithm library comes from the input of algorithm personnel.

[0035] As an optimization, the monitoring module includes log extraction, aggregate query and log management; it is used to provide performance indicators of the request interface and model effect indicator display, and its data comes from the log collection of the recommendation engine prediction interface.

[0036] As an optimization, the data layer includes business data, model data, system logs, and metadata, all of which are stored in the big data platform;

[0037] The business data is the training data that the recommendation scenario relies on and is used for model training and model generation;

[0038] The model data is a model file generated after training and is used for prediction;

[0039] The system log is the request and response log of the model, which is used to feedback the performance of the model instance.

[0040] The present invention also provides a general TOP-N intelligent recommendation method, comprising the following steps:

[0041] S1. Obtain user data, item data and behavior data, and store them in the database;

[0042] S2. Process the data obtained in S1 to obtain three basic data tables: user data table, item data table, and user-item interaction behavior table;

[0043] S3, generating a user table, an item table, an event-user table, and an event-item table according to the basic data table by statistical methods, and storing them in a database to form a data set;

[0044] S4, processing the data set through multiple recall strategies to generate a set of candidate items and obtain an initial recommendation result;

[0045] S5, filtering the initialization recommendation result by blacklist and supplementing the whitelist;

[0046] S6. Sort the filtered and supplemented results, and obtain recommendation results by combining user characteristics and item characteristics;

[0047] S7. After receiving the user request, return the recommendation result according to the user information, the user's geographical location information and the IP address information.

[0048] Compared with the prior art, the present invention has the following advantages:

[0049] The present invention adopts a universal recommendation system technical solution: just abstract a scene that needs TOPN recommendation into three data tables with specified formats and store them in the database, and then perform visual operations in the management background interface, perform data training, and form a unique API for the scene business, so as to obtain personalized recommendation service content. The operation management platform does not require professional technicians, and both product and operation personnel can operate according to the steps.

[0050] The present invention can flexibly configure and customize the algorithm strategies of modules such as recall, filtering, supplementation, and sorting; the algorithm usage strategy of the recommendation system can be flexibly adjusted through the configuration file; the deployment of services such as recall, rough sorting, and fine sorting can be quickly completed; and the logs of the corresponding modules can be recorded for easy monitoring and debugging. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a system architecture diagram of the present invention;

[0052] Figure 2 It is a system flow chart of the present invention. DETAILED DESCRIPTION

[0053] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0054] Example: See Figure 1-Figure 2 ,

[0055] A general TOP-N intelligent recommendation system includes four parts: application layer, service layer, data layer and storage layer.

[0056] 1. The application layer includes a recommendation scenario module, a user data module, an item data module, and a system management module, which are used to provide a management platform and a visual interactive interface for operation and display. Specifically, the application layer is a management platform that supports the training of the recommendation model after operation on the interactive interface to form an API. The recommended content for the corresponding user can be obtained based on user information, item library ID, item classification ID, and recommendation mode, and is suitable for various recommendation scenarios.

[0057] The recommendation scenario module includes data selection, model selection, cold start strategy selection, model management and model hosting environment, which are used to build and implement recommendation scenarios;

[0058] Data selection is used to select the training data required for the recommendation model corresponding to each recommendation scenario through an offline or online multi-source method. The training data needs to be classified into three categories (i.e., three tables), including user data, behavior data, and item data.

[0059] Model selection, which is used to select a corresponding algorithm from the algorithm library as the algorithm of the recommendation model. The recommendation algorithm includes personalized recommendation, popular recommendation, and similar recommendation;

[0060] Cold start strategy selection, used to select a strategy for a recommendation model from the cold start strategies, wherein the cold start strategies include a hot recommendation strategy and a time-ordered recommendation strategy;

[0061] Model management is used to manually configure the update cycle of the recommendation model, including immediate training or periodic training. It is also used to count the AUC offline indicator and the exposure, click count, and click rate online indicators of the recommendation model, and count the number of calls and call success rate of the recommendation model.

[0062] The model hosting environment is used to specify the service environment where the recommendation model is located, which includes a test environment or a formal environment, and configure relevant permissions based on the service environment.

[0063] The user data module includes a user data unit, a user data table management unit, a behavior data unit and a behavior data table management unit, which are used to access, display and manage user information and behavior information; wherein,

[0064] User data unit, used for searching, querying, previewing and visualizing user data;

[0065] User data table management unit, used to add, delete, modify and query user data;

[0066] Behavior data unit, used for searching, querying, previewing, and visualizing behavior data;

[0067] The behavior data management unit is used to add, delete, modify and query behavior data.

[0068] The item data module includes an item library list, item library management, and item classification management, which are used to access, display and manage item information;

[0069] Item library list, used to visualize the multi-cascade relationship of the item library;

[0070] Item library management, used to add, delete, modify and check the item library;

[0071] Item classification management is used to add, delete, modify and check item classifications.

[0072] The system management module includes log management and alarm management, which are used to manage the system operation and maintenance and monitor and manage online services;

[0073] Log management, which is used to manage the system backend to support the management of online and offline service logs, data and instance logs;

[0074] Alarm management is used to provide alarms for model training, data consistency, and recommendation system status.

[0075] 2. Service layer, including recommendation engine, model and algorithm management, and monitoring module; used to process data through multiple services, train recommendation models based on the data processing results, provide interfaces corresponding to multiple recommendation scenarios, and monitor and manage recommendation models and algorithms; specifically, the service layer is mainly divided into three parts: recommendation engine, model algorithm management, and monitoring. The recommendation engine supports the periodic training of models to generate models and applies models to interfaces provided by external services.

[0076] The recommendation engine includes recall service, filtering / supplement service, sorting service and recommendation service; it is used to support the periodic training of the model to generate the model, and apply the model to the interface provided by the external service.

[0077] The recall service is used to filter recommended items and generate a candidate list. Specifically, the recall service is responsible for filtering out some data from the massive data for use in the subsequent sorting stage. The recall process is added to quickly filter all the recommended items using a small number of features and simple models or rules, further reducing the time overhead of the subsequent sorting stage. The design of the recall service needs to consider both the recall rate and the calculation speed to ensure the recall of data that users are interested in and the timeliness.

[0078] The filtering / supplementing service is used to add and filter the candidate list generated by the recall service through the setting of black and white lists.

[0079] Sorting service, which is used to generate a user-item-behavior feature matrix based on user information, behavior information, and item information, and to make predictions on the feature matrix;

[0080] The recommendation service calls the recall server, correction server, and ranking server through the grpc protocol. The recommendation service provides an HTTP request interface and an HTTP response interface for external services.

[0081] The model and algorithm management includes a model library, an algorithm library and a model life cycle, which are used to provide model data management, algorithm management, the life cycle of trained models, versions, and online and offline management. The model library comes from the model generated by the recommendation engine training, and the algorithm library comes from the input of algorithm personnel.

[0082] The monitoring module includes log extraction, aggregate query and log management; it is used to provide performance indicators of the request interface and model effect indicator display, and its data comes from the log collection of the recommendation engine prediction interface.

[0083] The data layer includes business data, model data, system logs, and metadata; it is used to store model training data, model files, and logs on the big data platform.

[0084] The data layer includes business data, model data, system logs, and metadata, all of which are stored in the big data platform;

[0085] The business data is the training data that the recommendation scenario relies on and is used for model training and model generation; the storage location is the big data platform (Hive).

[0086] The model data is a model file generated after training and is used for prediction; the storage location is the big data platform (HDFS).

[0087] The system log is the model's request and response log, which is used to feedback the model instance performance and is stored in the big data platform (ES).

[0088] The storage layer is used to save data. The storage layer relies on the HDFS distributed file system for storage and the business metadata when Hive is available in the data warehouse. In the early stage, it is stored in the form of files on HDFS; the user management, model metadata, and monitoring data of the recommendation system itself are stored in MySQL; user request feature logs, prediction response logs, training logs, etc. are stored in Elasticsearch.

[0089] Based on the above system, the present invention also provides a general TOP-N intelligent recommendation method, comprising the following steps:

[0090] S1. Obtain user data, item data and behavior data, and store them in the database;

[0091] S2. Process the data obtained in S1 to obtain three basic data tables: user data table, item data table, and user-item interaction behavior table;

[0092] S3, generating a user table, an item table, an event-user table, and an event-item table according to the basic data table by statistical methods, and storing them in a database to form a data set;

[0093] S4, processing the data set through multiple recall strategies to generate a set of candidate items and obtain an initial recommendation result;

[0094] S5, filtering the initialization recommendation result by blacklist and supplementing the whitelist;

[0095] S6. Sort the filtered and supplemented results, and obtain recommendation results by combining user characteristics and item characteristics;

[0096] S7. After receiving the user request, return the recommendation result according to the user information, the user's geographical location information and the IP address information.

[0097] Specifically, the general recommendation system is mainly divided into four steps: recall (initialization of recommendation result generation), correction (filtering / supplementation), personalized sorting (result ranking), and recommendation display (final recommendation results).

[0098] like Figure 2 1-1 to 1-9 show the implementation method of the recall step.

[0099] A. The original data of users, items, and behaviors are pushed to the designated location (HDFS / DataBase, etc.) in a specified format. The push method is regular batch upload. The frequency of the regular update of the data set can range from half a day to two days, depending on the specific business needs.

[0100] B. The data adaptation module stores the uploaded raw data in a fixed format in the database. In the recommended business scenario, due to the need for algorithm and feature experiments, the raw data will be cleaned, completed, and annotated to form three basic tables with fixed formats, namely the user data table, the item data table, and the user-item interaction behavior table.

[0101] C. The data statistics module generates the user table, item table, event-user table, and event-item table based on the three basic information tables adapted by the data adapter, and stores the processed results in the database.

[0102] D. The above steps have resulted in a massive amount of processed data sets. The recall service will quickly obtain a relatively small amount of data from the data set based on different recall strategies such as content recall, similarity recall, and model recall, and call the candidate item set stored in the database based on the user ID. In view of the characteristics of multiple recall strategies, the candidate lists of multi-source recalls are deduplicated and merged at the same time, and the recall candidate set results are generated based on the real-time characteristics of users and items.

[0103] like Figure 2 2-1 to 2-4 show the implementation method of the filtering supplement step.

[0104] A. After the recall module processes the data, it obtains the initial recommendation results. Since the recalled items may be items that the user has already processed or browsed, it is necessary to blacklist the items recalled by the user; similarly, for users whose recalled items are empty, the whitelist needs to be supplemented.

[0105] B. A small number of items are obtained from the filtering and supplementing stage and handed over to the sorting stage. The sorting stage incorporates more user characteristics and product characteristics and uses complex models to accurately make personalized recommendations. Sorting emphasizes efficient feedback results and highly accurate recommendation results.

[0106] like Figure 2 3-1 to 3-8 show the implementation method of the sorting step.

[0107] A. Based on the conversion rate and exposure rate of the items, the XY supervised dataset is updated regularly and constructed as the training set for the sorting model.

[0108] B. After generating the candidate training set, the sorting model will score and sort the generated candidate objects to obtain the final list to be pushed. The recommendation system has recall queues from different sources, such as matrix decomposition model related, user related under various tags, etc. Since the scores of recommendations recalled by multiple recall strategies cannot be directly compared, the system combines these different sources into a common candidate library, and then a single model scores its memory, and combines user characteristics, item characteristics, and environmental information characteristics to personalize the item list.

[0109] C. Since the ranking model is based on the ratio of user exposure to clicks, in order to improve other indicators, the ranking score is multiplied by a weight coefficient to adjust the ranking of the recommended item list. For example, in order to increase the click rate of newly listed items, the ranking score is multiplied by a coefficient that is negatively correlated with the listing time.

[0110] The client initiates a request protocol, which carries information such as user ID, geographic location information, IP address, etc. The recommendation system's recommendation results are returned in the result feedback.

[0111] The present invention adopts a universal recommendation system technical solution: just abstract a scene that needs TOPN recommendation into three data tables with specified formats and store them in the database, and then perform visual operations in the management background interface, perform data training, and form a unique API for the scene business, so as to obtain personalized recommendation service content. The operation management platform does not require professional technicians, and both product and operation personnel can operate according to the steps.

[0112] The present invention can flexibly configure and customize the algorithm strategies of modules such as recall, filtering, supplementation, and sorting; the algorithm usage strategy of the recommendation system can be flexibly adjusted through the configuration file; the deployment of services such as recall, rough sorting, and fine sorting can be quickly completed; and the logs of the corresponding modules can be recorded for easy monitoring and debugging.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit the technical solution. Those skilled in the art should understand that those modifications or equivalent substitutions of the technical solution of the present invention that do not depart from the purpose and scope of the technical solution should be included in the scope of the claims of the present invention.

Claims

1. A general TOP-N intelligent recommendation system, characterized in that: include, The application layer includes a recommendation scenario module, a user data module, an item data module and a system management module, which are used to provide a management platform and a visual interactive interface for operation and display; the recommendation scenario module includes data selection, model selection, cold start strategy selection, model management and model hosting environment, which are used to build and implement recommendation scenarios; wherein, Data selection, used to select the training data required by the recommendation model corresponding to each recommendation scenario through offline or online multi-source methods, the training data including user data, behavior data, and item data; Model selection, used to select a corresponding algorithm from the algorithm library as the algorithm of the recommendation model, the algorithm of the recommendation model includes personalized recommendation, popular recommendation, and similar recommendation; Cold start strategy selection, used to select a strategy for a recommendation model from the cold start strategies, wherein the cold start strategies include a hot recommendation strategy and a time-ordered recommendation strategy; Model management is used to configure the update cycle of the recommendation model, including immediate training or periodic training. It is also used to count the AUC offline index and the exposure, click count, and click rate online indexes of the recommendation model, and count the number of calls and call success rate of the recommendation model. Model hosting environment, which is used to specify the service environment where the recommendation model is located, the service environment includes a test environment or a formal environment, and configure relevant permissions according to the service environment; The service layer includes a recommendation engine, model and algorithm management, and monitoring modules; it is used to process data through multiple services, train the recommendation model based on the data processing results, provide interfaces corresponding to multiple recommendation scenarios, and monitor and manage the recommendation model and algorithm; the recommendation engine includes recall service, filtering / supplement service, sorting service, and recommendation service; it is used to support the periodic training of the model to generate a model, and apply the model to the interface provided by external services; wherein, Recall service, used to filter recommended items and generate a candidate list; Filtering / supplementing service, used to add and filter the candidate list generated by the recall service through the setting of blacklist and whitelist; Sorting service, which is used to generate a user-item-behavior feature matrix based on user information, behavior information, and item information, and to make predictions on the feature matrix; Recommendation service, which is used to call recall service, filtering / supplementing service, and sorting service through grpc protocol, and provides HTTP request interface and HTTP response interface for external services; The data layer includes business data, model data, system logs, and metadata; it is used to store model training data, model files, and logs on the big data platform; Storage layer, used to save data.

2. A general TOP-N intelligent recommendation system according to claim 1, characterized in that: The user data module includes a user data unit, a user data table management unit, a behavior data unit and a behavior data table management unit, which are used to access, display and manage user information and behavior information; wherein, User data unit, used for searching, querying, previewing and visualizing user data; User data table management unit, used to add, delete, modify and query user data; Behavior data unit, used for searching, querying, previewing, and visualizing behavior data; The behavior data management unit is used to add, delete, modify and query behavior data.

3. A general TOP-N intelligent recommendation system according to claim 1, characterized in that: The item data module includes an item library list, item library management, and item classification management, which are used to access, display and manage item information; Item library list, used to visualize the multi-cascade relationship of the item library; Item library management, used to add, delete, modify and check the item library; Item classification management is used to add, delete, modify and check item classifications.

4. A general TOP-N intelligent recommendation system according to claim 1, characterized in that: The system management module includes log management and alarm management, which are used to manage the operation and maintenance of the system; Log management, which is used to manage the system backend to support the management of online and offline service logs, data and instance logs; Alarm management is used to provide alarms for model training, data consistency, and recommendation system status.

5. A general TOP-N intelligent recommendation system according to claim 1, characterized in that: The model and algorithm management includes a model library, an algorithm library and a model life cycle, which are used to provide model data management, algorithm management, the life cycle of trained models, versions, and online and offline management. The model library comes from the model generated by the recommendation engine training, and the algorithm library comes from the input of algorithm personnel.

6. A general TOP-N intelligent recommendation system according to claim 1, characterized in that: The monitoring module includes log extraction, aggregate query and log management; it is used to provide performance indicators of the request interface and model effect indicator display, and its data comes from the log collection of the recommendation engine prediction interface.

7. A general TOP-N intelligent recommendation system according to claim 1, characterized in that: The data layer includes business data, model data, system logs, and metadata, all of which are stored in the big data platform; The business data is the training data that the recommendation scenario relies on and is used for model training and model generation; The model data is a model file generated after training and is used for prediction; The system log is the request and response log of the model, which is used to feedback the performance of the model instance.

8. A general TOP-N intelligent recommendation method, characterized in that: The following steps are included: S1. Obtain user data, item data and behavior data, and store them in the database; S2. Process the data obtained in S1 to obtain three basic data tables: user data table, item data table, and user-item interaction behavior table; S3, generating a user table, an item table, an event-user table, and an event-item table according to the basic data table by statistical methods, and storing them in a database to form a data set; S4. Processing the data set through multiple recall strategies to generate a candidate item set; the recall strategy includes obtaining a small amount of data from the data set based on content recall, similarity recall and / or model recall, calling the candidate item set stored in the database based on the user ID, deduplicating and fusing the candidate lists of multi-source recall at the same time according to the characteristics of multiple recall strategies, generating a recall candidate set result based on the real-time characteristics of users and items, and obtaining an initial recommendation result; S5. Perform blacklist filtering and whitelist supplement on the initial recommendation result; specifically, perform blacklist filtering on items that the user has processed or browsed in the initial recommendation result; and supplement the whitelist for users whose initial recommendation result is empty; S6. Sort the filtered and supplemented results, and obtain recommendation results by combining user characteristics and item characteristics; Specifically, the sorting steps include: A. Based on the conversion rate and exposure rate of the items, the XY supervised dataset is updated regularly and constructed as the training set for the sorting model; B. After generating the candidate training set, the sorting model will score and sort the generated candidate objects to obtain the final list to be pushed; among them, the different sources of recall by multiple recall strategies are combined into a common candidate library, which is then scored by a single model, and the item list is personalized and sorted by combining user characteristics, item characteristics, and environmental information characteristics; C. Multiply the ranking score by a weight coefficient to adjust the ranking of the recommended item list; S7. After receiving the user request, return the recommendation result according to the user information, the user's geographical location information and the IP address information.

Images