Training method, device, equipment, medium and program product of recommendation model

Abstract

This application discloses a training method, apparatus, device, medium, and program product for a recommendation model, relating to the field of artificial intelligence. The method includes: acquiring multiple interaction data sets generated by a target account within a first historical time period; performing feature cross-processing on the multiple interaction data sets to obtain cross-feature representations of the interaction data sets; predicting the interest level between the target account and virtual items using a candidate item recommendation model, wherein the candidate item recommendation model is a model trained based on multiple interaction data sets generated by the target account within a second historical time period; and training the candidate item recommendation model based on the interest level prediction results to obtain an item recommendation model. This method improves the recommendation accuracy when recommending items using the model.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a method, apparatus, device, medium, and program product for training a recommendation model. Background Technology

[0002] Games typically feature a virtual shop where players can exchange virtual resources for virtual items, such as virtual props, virtual characters, and virtual skins. When displaying items for exchange in the virtual shop, a recommendation system is integrated. This system predicts an account's interest in virtual items based on their preferences and then displays and recommends items accordingly.

[0003] In related technologies, the training of the recommendation prediction model in the aforementioned recommendation system typically employs a data window training mode. This involves acquiring historical behavior data of an account based on a set data window, using this historical behavior data as sample data to train the initial model, thereby obtaining the required recommendation model. For example, when the data window is one month, historical behavior data from the previous month is acquired for model training when the recommendation model needs to be updated.

[0004] However, in the training process of the above recommendation model, if the selected data window is too large, the features learned from earlier behavioral data may be invalid features in actual needs. If the selected data window is too small, the model learns from a smaller sample size, resulting in insufficient feature learning. Therefore, the above recommendation model suffers from low recommendation accuracy in application. Summary of the Invention

[0005] This application provides a training method, apparatus, device, medium, and program product for a recommendation model, which can improve the recommendation accuracy of virtual items. The technical solution is as follows:

[0006] On the one hand, a method for training a recommendation model is provided, the method comprising:

[0007] Acquire multiple sets of interactive data generated by the target account within a first historical time period. Each set of interactive data includes multiple sets of interactive data, which are used to represent the interactive relationship between the target account and virtual items.

[0008] Perform feature cross-processing on multiple interactive data in the interactive data group to obtain the cross-feature representation of the interactive data group;

[0009] Interest prediction is performed on the cross-feature representation by the candidate item recommendation model to obtain the interest prediction result between the target account and the virtual item. The candidate item recommendation model is a model trained based on multiple interaction data groups generated by the target account in the second historical time period. The end time of the second historical time period is located before the end time of the first historical time period in time sequence.

[0010] The candidate item recommendation model is trained based on the interest prediction results to obtain an item recommendation model, which is used to recommend virtual items to the target account in the game store.

[0011] On the other hand, a training apparatus for a recommendation model is provided, the apparatus comprising:

[0012] The acquisition module is used to acquire multiple groups of interactive data generated by the target account within a first historical time period. The interactive data group includes multiple interactive data, which are used to represent the interactive relationship between the target account and virtual items.

[0013] The processing module is used to perform feature cross-processing on multiple interactive data in the interactive data group to obtain the cross-feature representation of the interactive data group;

[0014] The training module is used to predict the interest degree of the cross feature representation through the candidate item recommendation model, so as to obtain the interest degree prediction result between the target account and the virtual item. The candidate item recommendation model is a model trained based on multiple interaction data groups generated by the target account in the second historical time period. The end time of the second historical time period is located before the end time of the first historical time period in time sequence.

[0015] The training module is further used to train the candidate item recommendation model based on the interest prediction results to obtain an item recommendation model, which is used to recommend virtual items to the target account in the game store.

[0016] On the other hand, a computer device is provided, the terminal including a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set, the at least one instruction, the at least one program, the code set or instruction set being loaded and executed by the processor to implement the training method of any of the recommendation models described in the embodiments of this application.

[0017] On the other hand, a computer-readable storage medium is provided, wherein at least one piece of program code is stored in the computer-readable storage medium, the program code being loaded and executed by a processor to implement the training method of any of the recommendation models described in the embodiments of this application.

[0018] On the other hand, a computer program product or computer program is provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the training method of the recommended model described in any of the above embodiments.

[0019] The technical solution provided in this application includes at least the following beneficial effects:

[0020] When incrementally training an item recommendation model for predicting item recommendation rates, multiple sets of interaction data within a first historical time period are acquired. Feature cross-processing is performed on the interaction data within these sets to obtain cross-feature representations. These cross-feature representations are then used to train a candidate recommendation model, resulting in a new item recommendation model. This candidate model is trained based on multiple sets of interaction data within a second historical time period, the end of which is chronologically prior to the end of the first historical time period. In other words, during model training, an iterative update approach is used to incrementally learn from the previous item recommendation model. This allows the model to learn from recent data directly based on the already trained model parameters, resulting in an item recommendation model sensitive to recent data. This improves the accuracy of item recommendations when using the model and maintains synchronization between the model and changes in account interest preferences. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of window learning in related technologies;

[0023] Figure 2 This is a schematic diagram illustrating incremental learning provided in an exemplary embodiment of this application;

[0024] Figure 3 This is a schematic diagram of the interface for virtual skin recommendation provided in an exemplary embodiment of this application;

[0025] Figure 4 This is a schematic diagram of an implementation environment provided by an exemplary embodiment of this application;

[0026] Figure 5This is a flowchart of a training method for a recommendation model provided in an exemplary embodiment of this application;

[0027] Figure 6 This is a flowchart of a training method for a recommendation model provided in another exemplary embodiment of this application;

[0028] Figure 7 This is a diagram illustrating the application process corresponding to the interactive features in related technologies;

[0029] Figure 8 This is a diagram illustrating the application process of the cross feature provided in an exemplary embodiment of this application;

[0030] Figure 9 This is a flowchart of a training method for a recommendation model provided in another exemplary embodiment of this application;

[0031] Figure 10 This is a schematic diagram of a personalized recommendation system for a skin e-commerce platform provided in an exemplary embodiment of this application;

[0032] Figure 11 This is a flowchart of a training method for a recommendation model provided in an exemplary embodiment of this application;

[0033] Figure 12 This is a flowchart illustrating data processing provided in an exemplary embodiment of this application;

[0034] Figure 13 This is a block diagram of a training device for a recommendation model provided in an exemplary embodiment of this application;

[0035] Figure 14 This is a block diagram of a training device for a recommendation model provided in an exemplary embodiment of this application;

[0036] Figure 15 This is a schematic diagram of the structure of a server provided in an exemplary embodiment of this application. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0038] First, a brief introduction to the terms used in the embodiments of this application:

[0039] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce new intelligent machines that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess perception, reasoning, and decision-making capabilities.

[0040] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, as well as machine learning / deep learning, autonomous driving, and intelligent transportation.

[0041] Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instruction-based learning.

[0042] With the research and advancement of artificial intelligence (AI) technology, AI is being studied and applied in various fields, such as smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, autonomous driving, drones, robots, smart healthcare, smart customer service, vehicle networking, and intelligent transportation. It is believed that with the development of technology, AI will be applied in more fields and play an increasingly important role.

[0043] Recommendation Systems: With the development of the internet, more and more content is being presented to users. To prevent users from churning due to information overload and browsing too much content that doesn't interest them, personalized recommendation systems have emerged. Personalized recommendation systems are intelligent platforms built on massive amounts of data, improving the accuracy of content recommendations.

[0044] Recommendation systems are also used in gaming scenarios. For example, in-game virtual stores, a variety of virtual items are offered, such as virtual props, virtual characters, and virtual skins. Different players may show different interests in different virtual items. To improve the efficiency of players acquiring virtual items, the virtual store can incorporate a recommendation system to prioritize displaying virtual items that match players' interests.

[0045] Because player interests change over time, the recommendation system for the virtual store needs to be updated. Related technologies often employ window learning to update the recommendation system. That is, the server acquires historical data based on a set window size, uses this historical data as samples to train an initial prediction model, and obtains the corresponding prediction model for application in the recommendation system. Please refer to [reference needed]. Figure 1 The diagram illustrates window learning 100 in related technologies. Taking a window size of 30 days as an example, model {T-2}121 is trained based on historical data {T-31}~{T-2}111, model {T-1}122 is trained based on historical data {T-30}~{T-1}112, and model {T}123 is trained based on historical data {T-29}~{T}113.

[0046] However, the update method for recommendation systems implemented through window learning described above has the following problems:

[0047] (1) The data span is limited, that is, the model only learns the features of the data within the window, and its learning results are limited by the window size;

[0048] (2) In order to ensure sufficient sample data so that the model training can converge, the window size cannot be too small, that is, each model update requires a large amount of data to support it.

[0049] (3) The trained model is not sensitive to recent data. That is, the model is updated based on historical data within a large data window. All sample data have an equal probability of being learned, so the model parameters are equally sensitive to all data within the window.

[0050] (4) There is a significant sample crossover problem. That is, since the sample data within the window are mixed together for sampling training, when data from an earlier time is sampled, the model may have already been trained on recent data.

[0051] In this embodiment, incremental learning technology from artificial intelligence is integrated into the recommendation system to update the item recommendation model in the recommendation system through incremental learning.

[0052] Incremental learning refers to a learning system that continuously learns new knowledge from new samples while retaining most of the previously learned knowledge. That is, whenever new data is added, it is not necessary to rebuild the entire knowledge base; instead, the existing knowledge base is updated only to reflect the changes caused by the new data.

[0053] Indicative, such as Figure 2 As shown, it illustrates an incremental learning 200 provided in an exemplary embodiment of this application, with a daily update frequency. Model {T-2}221 is trained based on historical data {T-2}211 and model {T-3}, model {T-1}222 is trained based on historical data {T-1}212 and model {T-2}221, and model {T}223 is trained based on historical data {T}213 and model {T-1}222.

[0054] Compared to window learning, incremental learning has at least the following advantages:

[0055] (1) The training data span is longer, that is, incremental learning is based on all historical data to complete the training, and the features of historical data are passed to the next training process through model parameters;

[0056] (2) The model update efficiency is high, that is, incremental learning only needs to train the data of the day, the amount of sample data is small, and the corresponding training time is short.

[0057] (3) Recent data has a greater impact on the model. That is, since the model is mainly trained iteratively on recent data during incremental learning, the model parameters are more representative of recent data.

[0058] (4) Fewer sample crossing problems: Incremental learning updates the model only for data within a short period of time. The model parameters are more inclined to the features of recent data, while the features of earlier data are weakened, thus reducing the sample crossing problem.

[0059] Based on the above definitions, the application scenarios provided in the embodiments of this application are illustrated below:

[0060] This method can be applied to the item recommendation process in a virtual store within a game scenario. For example, consider the recommendation of virtual skins. Figure 3As shown, this is a schematic diagram of a virtual skin recommendation interface provided by an exemplary embodiment of this application. When the user enters the main game interface 310 through the game application, a virtual store entrance 311 is displayed on the main game interface 310. When the virtual store entrance 311 receives a trigger operation, a virtual store interface 320 is displayed. The virtual store interface 320 displays multiple recommended skins 321 displayed in a recommendation order list as instructed by the server. The recommendation order list is a list generated by the server based on the recommended order of virtual skins predicted by the item recommendation model updated on the same day after the terminal sends a skin recommendation request to the server. The terminal displays virtual skins according to the recommendation order list.

[0061] It is worth noting that this method can also be applied to other recommendation scenarios where item recommendations are needed. In another example, this method can also be applied to the item recommendation process in e-commerce scenarios, which will not be elaborated here.

[0062] Please refer to Figure 4 This diagram illustrates an implementation environment provided by an exemplary embodiment of this application. The computer system of this implementation environment includes: a terminal 410, a server 420, and a communication network 430.

[0063] Terminal 410 includes various forms of devices such as mobile phones, tablets, desktop computers, laptops, smart voice interaction devices, smart home appliances, vehicle terminals, and aircraft. Illustratively, terminal 410 runs a target application, which can be at least one of the following: a game application, an e-commerce application, or a social application. In one example, taking a recommendation scenario applied to a game as an example, the aforementioned game application can be any one of the following: a third-person shooter (TPS) game, a first-person shooter (FPS) game, a multiplayer online battle arena (MOBA) game, a massively multiplayer online role-playing game (MMORPG), or a strategy game.

[0064] Server 420 is used to provide training and application functions for the item recommendation model. It is worth noting that server 420 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud security, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0065] Cloud technology refers to a hosting technology that unifies hardware, software, and network resources within a wide area network (WAN) or local area network (LAN) to achieve data computation, storage, processing, and sharing. Based on the cloud computing business model, cloud technology encompasses network technology, information technology, integration technology, management platform technology, and application technology. It can form resource pools, providing flexible and convenient on-demand access. Cloud computing technology will become a crucial support. Backend services of technical network systems require substantial computing and storage resources, such as video websites, image websites, and many portal websites. With the rapid development and application of the internet industry, every item may have its own identification mark in the future, requiring transmission to backend systems for logical processing. Data at different levels will be processed separately, and various industry data will require robust system support, which can only be achieved through cloud computing.

[0066] In some embodiments, the server 420 described above can also be implemented as a node in a blockchain system. Blockchain is a new application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms.

[0067] Indicatively, terminal 410 and server 420 are connected via communication network 430, which can be a wired network or a wireless network, and is not limited here.

[0068] In some embodiments, server 420 includes a model update service and a model application service, wherein the model update service operates offline and the model application service operates online.

[0069] The aforementioned model update service acquires multiple interaction data sets generated by the target account within the first historical time period. It performs feature cross-processing on these interaction data sets to obtain cross-feature representations. An interest prediction is then performed on these cross-feature representations using a candidate item recommendation model to obtain the interest prediction result between the target account and the virtual items. The candidate item recommendation model is trained based on interaction data sets generated by the target account at multiple time points within the second historical time period. The model is then trained based on the interest prediction result to obtain the item recommendation model. After completing model training, the model update service updates the model parameters of the trained item recommendation model to the online parameter storage area between the model update service and the model application service.

[0070] After receiving the item recommendation request from terminal 410, the model application service retrieves the latest model parameters from the online parameter storage area, loads them into the online service model, predicts the recommendation score of the candidate items using the online service model, and arranges the candidate items according to the recommendation score to obtain a recommendation order list. The recommendation order list is then sent to terminal 410, and terminal 410 displays the candidate items according to the recommendation order list.

[0071] In other embodiments, the item recommendation model is installed as an application component in the target application of terminal 410. That is, when terminal 410 has an item recommendation need, it calls the item recommendation model in the target application to predict item recommendations and obtain the corresponding recommendation order list, thereby realizing a localized item recommendation process. After server 420 completes the update of the item recommendation model, server 420 sends the updated model parameters of the item recommendation model to terminal 410 as update data. Terminal 410 initiates an application update based on the update data, updating the application component corresponding to the item recommendation model in the target application.

[0072] In some embodiments, the method of this application can be applied to an in-vehicle scenario. That is, when the terminal 410 is implemented as an in-vehicle terminal, illustratively, the target application can be implemented as a game auxiliary application corresponding to a game application. The game auxiliary application is used to provide auxiliary functions to the game application. For example, a player logs into the game application on a mobile terminal with a target account, and simultaneously logs into the auxiliary game application on the in-vehicle terminal with the target account. The auxiliary game application provides the player with auxiliary functions such as battle record query function, game store function, and social function associated with the game application. When the auxiliary game application on the in-vehicle terminal has an item recommendation requirement, it obtains a recommendation order list from the server and displays the candidate items in the recommendation order list in the auxiliary game application. For example, the auxiliary game application provides a recommendation function for virtual skins within the game application. The auxiliary game application on the in-vehicle terminal obtains a skin recommendation order list from the server that meets the target account's preference for virtual skins. The in-vehicle terminal pushes the first N virtual skins in the skin recommendation order list or provides an acquisition path associated with the game application.

[0073] Please refer to Figure 5 It illustrates a flowchart of a training method for a recommendation model according to an embodiment of this application. In this embodiment, the method is applied to, for example... Figure 4 In the server shown, the method includes:

[0074] Step 510: Obtain multiple groups of interactive data generated by the target account within the first historical time period.

[0075] Schematic representation: Each of the aforementioned multiple interaction data groups includes multiple interaction data sets, which represent the interaction relationship between the target account and virtual items. In some embodiments, the multiple interaction data sets in the aforementioned interaction data groups are discrete data sets.

[0076] In some embodiments, the multiple interactive data included in the aforementioned interactive data group include interactive data characterizing item characteristics and interactive data characterizing account characteristics. That is, in the embodiments of this application, the interactive data in the interactive data group are discrete from each other.

[0077] Optionally, the interaction data representing account characteristics may include at least one of the following: account acquisition (purchase) status, click status, usage status, collection status, like status, sharing status, etc.

[0078] In some embodiments, each of the multiple interactive data groups may be an interactive data group corresponding to multiple time points within a first historical time period. For example, the multiple interactive data groups may include a first interactive data group {virtual items acquired by the target account within 7 days, target virtual items}, a second interactive data group {virtual items acquired by the target account within 15 days, target virtual items}, and a third interactive data group {virtual items acquired by the target account within 30 days, target virtual items}.

[0079] In some embodiments, the combination of interactive data within multiple interactive data groups differs; that is, the data combinations of multiple interactive data differ in different interactive data groups. For example, multiple interactive data groups may include a first interactive data group {virtual items liked by the target account within 7 days, target virtual items}, a second interactive data group {virtual items acquired by the target account within 7 days, target virtual items}, and a third interactive data group {virtual items added to the target account's wishlist within 7 days, target virtual items}.

[0080] In one example, taking the recommendation of virtual skins in a game as an example, as shown in Table 1, it illustrates an exemplary set of interactive data provided in this application, wherein each set of interactive data includes two different sets of interactive data.

[0081] Table 1

[0082]

[0083] In some embodiments, when multiple sets of interactive data groups are used for model updates, the first historical time period corresponding to each set of interactive data groups can be a historical time period with a first time point as the end time point and different specified durations as the period length.

[0084] In some embodiments, the first historical time period is a historical time period with a first moment as the end moment and a specified duration as the period.

[0085] Optionally, the specified duration can be a duration associated with the model's update cycle (frequency). In one example, the specified duration can be N times the model update cycle, where N is a positive integer.

[0086] Optionally, the target account can be an account logged in on a specified terminal, all accounts in the target application, or a subset of accounts in the target application. For example, when the trained item recommendation model serves only a single account, the target account is the account logged in on the specified terminal; when the trained item recommendation model serves multiple accounts, the target account is all accounts in the target application or a subset of accounts in the target application. The subset of accounts can be accounts that have frequent interactions with virtual items, selected by the server based on the amount of interaction data corresponding to those accounts.

[0087] For illustrative purposes, the aforementioned virtual items include items to be recommended within the target application. In one example, if the target application is a game application and the recommendation scenario is a virtual store recommendation, then the aforementioned virtual items may include virtual props, virtual skins, virtual characters, etc., to be recommended.

[0088] In some embodiments, the aforementioned interaction data sets are obtained through log data recorded by the server. For example, log data of the target account within a first historical time period is obtained. The log data is used to record the interaction between the target account and virtual items. At least one preprocessing operation is performed on the log data to obtain multiple interaction data sets.

[0089] The preprocessing operations include bucketing, normalization, truncation, and hashing. Bucketing is used to split the continuous information represented by log data, thus discretizing the data; normalization is used to unify the values ​​of different categories of information in the log data to a specified range; truncation is used to filter out abnormal data in the log data; and hashing is used to convert the log data into a vector or matrix representation.

[0090] Step 520: Perform feature cross processing on multiple interactive data in the interactive data group to obtain the cross feature representation of the interactive data group.

[0091] In illustrative terms, the above feature cross-processing instruction crosses the features indicated by the interactive data to obtain feature combinations, which are then used as features in the input model.

[0092] Optionally, when crossing the features indicated by the interactive data, the crossing method may include at least one of self-crossing and mutual crossing. Specifically, self-crossing indicates that each feature of interactive data crosses with itself; for example, feature A corresponding to interactive data A crosses with itself to obtain a feature combination {feature A, feature A}. Mutual crossing indicates that features of different interactive data cross; for example, feature A corresponding to interactive data A and feature B corresponding to interactive data B cross to obtain a feature combination {feature A, feature B}.

[0093] Schematic, the crossing of features corresponds to a crossing order, which indicates the number of features involved in the crossing. For example, a crossing order of two indicates that two features are crossing; a crossing order of three indicates that three features are crossing. In some embodiments, the crossing order is determined by the number of interactive data in the interactive data set.

[0094] In some embodiments, the above feature cross processing can be implemented by performing at least one of the inner product, Hadamard product, and Cartesian product on the data feature representations corresponding to multiple interactive data.

[0095] In other embodiments, the feature crossing process described above can also be implemented automatically by a model, that is, a dedicated feature crossing layer is designed in the model to perform feature crossing on multiple interactive data. In one example, the feature crossing process described above is implemented using a Multilayer Perceptron (MLP).

[0096] Step 530: Use the candidate item recommendation model to predict the interest level of the cross-feature representation to obtain the interest level prediction result between the target account and the virtual item.

[0097] The aforementioned candidate item recommendation model is a model trained based on multiple interaction data sets generated by the target account within the second historical time period. The second historical time period is a historical time period with a specified duration as the period ending at the second moment. The second moment is located before the first moment in time sequence, that is, the end moment of the second historical time period is located before the end moment of the first historical time period in time sequence.

[0098] Optionally, when the specified duration corresponding to the first historical time period and the second historical time period is equal to the model update cycle, the first historical time period and the second historical time period can be two adjacent time periods; or, when the specified duration is a specified multiple of the model update cycle, the first historical time period and the second historical time period can be time periods with overlapping time periods.

[0099] Optionally, the above-mentioned candidate item recommendation model can be implemented as at least one of Deep Neural Networks (DNN), Convolutional Neural Networks (CNN), Graph Neural Networks (GNN), etc.

[0100] In some embodiments, when multiple sets of interaction data exist, the multiple sets of interaction data are randomly sampled to determine the target interaction data set. The cross-feature representation of the target interaction data set is then input into the candidate item recommendation model to predict interest and obtain the interest prediction result.

[0101] Step 540: Train the candidate item recommendation model based on the interest prediction results to obtain the item recommendation model.

[0102] As an illustration, the above item recommendation model is used to recommend virtual items to a target account in the game store.

[0103] In some embodiments, the training of the above-mentioned candidate item recommendation model is implemented as supervised training, that is, the above-mentioned interaction data group corresponds to sample labels, and the sample labels indicate the interest of the target account corresponding to the interaction data group in the item.

[0104] Optionally, the sample labels mentioned above can be manually labeled, or they can be automatically generated during the interactive data set generation process.

[0105] In a schematic way, after inputting the cross-feature representation into the candidate item recommendation model and predicting the interest prediction result, the model parameters of the candidate item recommendation model are iteratively adjusted according to the difference between the interest prediction result and the sample label, thereby obtaining the item recommendation model.

[0106] To illustrate, when determining the difference between the interest prediction result and the sample label, a loss function can be specified. That is, the interest prediction result and the sample label are input into the specified loss function to determine the loss value corresponding to this prediction process. Based on the above loss value, the model parameters of the candidate item recommendation model are iteratively adjusted to obtain the item recommendation model.

[0107] Optionally, the specified loss function can be implemented as at least one of the following: cross-entropy loss function, mean squared error loss function, log loss function, etc., without limitation.

[0108] In summary, the training method for the recommendation model provided in this application, during incremental training of the item recommendation model used to predict item recommendation degree, acquires multiple sets of interaction data within a first historical time period, performs feature cross-processing on the interaction data in the interaction data sets to obtain cross-feature representations, and trains candidate recommendation models using these cross-feature representations to obtain new item recommendation models. These candidate recommendation models are trained based on multiple sets of interaction data within a second historical time period, the end time of which is chronologically prior to the end time of the first historical time period. That is, during model training, an iterative update approach is used to incrementally learn from the previous model used for item recommendation, enabling the model to directly learn from recent data based on the already trained model parameters, resulting in an item recommendation model sensitive to recent data. This improves the recommendation accuracy when using the item recommendation model and maintains the synchronization between the model and changes in the account's interest tendencies.

[0109] Please refer to Figure 6 This document illustrates a flowchart of a training method for a recommendation model provided in an exemplary embodiment of this application. In this embodiment, the feature crossover process during the feature construction stage is illustrated, specifically steps 521-522. Figure 5 The next step after step 520. This method is implemented as follows:

[0110] Step 521: Extract features from multiple interactive data in the interactive data group to obtain the data feature representation corresponding to each interactive data.

[0111] In some embodiments, the aforementioned interactive data may be data indicated by an identifier. For example, the interactive data group includes interactive data A and interactive data B, wherein interactive data A is a list of IDs of existing virtual characters in the target account, and interactive data B is the ID of the target virtual skin.

[0112] For illustrative purposes, when the aforementioned interactive data is data indicated by an identifier, the interactive data can be directly used as a data feature representation; or, the identifier indicating the interactive data can be encoded to obtain the aforementioned data feature representation, wherein the aforementioned encoding can be implemented as one-hot encoding.

[0113] In other embodiments, the aforementioned interactive data may be data indicated by descriptive text. For example, the interactive data may include interactive data A and interactive data B, wherein interactive data A is a list of skin names corresponding to the virtual skins already existing in the target account, and interactive data B is the skin name of the target virtual skin.

[0114] Schematic, when the aforementioned interactive data is data indicated by descriptive text, the text content of the interactive data can be encoded using a specified encoding method to obtain the aforementioned data feature representation. The specified encoding method can be implemented as one-hot encoding. Alternatively, text features can be extracted from the descriptive text indicating the interactive data to obtain the aforementioned data feature representation. The text feature extraction can be implemented by calculating the Term Frequency-Inverse Document Frequency (TF-IDF) index corresponding to the descriptive text.

[0115] Step 522: Determine the cross feature representation based on the Cartesian product between the data feature representations of multiple interactive data.

[0116] In this embodiment, the cross feature representation is obtained by calculating the Cartesian product between the data feature representations of different interactive data in the interactive data group.

[0117] Optionally, the Cartesian product between the above-mentioned data feature representations can be calculated in at least one of the following ways:

[0118] The first method involves performing a continuous Cartesian product on multiple interactive data sets within the interactive data group.

[0119] To illustrate, the first candidate feature representation corresponding to the first i interactive data in the interactive data group is obtained. The first candidate feature representation is the continuous Cartesian product corresponding to the first i interactive data. The data feature representation corresponding to the (i+1)th interactive data and the first candidate feature representation are combined into a Cartesian product to obtain the cross feature representation, where i is a positive integer.

[0120] That is, the cross feature representation x combine This can be represented as Formula 1.

[0121] Formula 1: x combine = x1×x2×...×x n

[0122] Where n is a positive integer, determined by the number of interactive data in the interactive data group, x n This represents the data features corresponding to each interactive data in the interactive data group.

[0123] The second method involves performing pairwise product (Cartesian product) on multiple interactive data sets within the interactive data group.

[0124] To illustrate, the data feature representations of the i-th and j-th interactive data in the interactive data group are Cartesian products to obtain the second candidate feature representation, where i and j are unequal positive integers; the second candidate feature representations between every two interactive data in the interactive data group are feature concatenated to obtain the cross feature representation.

[0125] The feature space of the cross feature representation is the product of the feature spaces of all discrete data feature representations before the cross, thus resulting in a huge feature space. In some embodiments, to avoid the cross feature representation being too sparse and detrimental to model learning, a second-order cross is employed. In one example, taking an interaction data set that includes interaction data representing account features and interaction data representing item features, the data feature representation includes the account feature representation x. user and item characteristic representation x item Then the cross feature represents x combine This can be represented as Formula 2.

[0126] Formula 2: x combine =x user ×x item

[0127] In one example, taking the recommendation of in-game virtual skins as an example, x user For the target account, the sequence of hero IDs used in the last 7 days, x user = [103,282,125,138], x item x is the ID of the target skin item =12503, then x combine = [103°12503, 282°12503, 125°12503, 138°12503].

[0128] In some embodiments, weights can be incorporated when calculating the Cartesian product between data feature representations, i.e., feature crosses can be implemented as weighted Cartesian products.

[0129] Optionally, the weights of each data feature can be preset by the system, specified by the terminal, or determined according to preset conversion rules.

[0130] Optionally, the above-mentioned preset conversion rules can be implemented in at least one of the following ways:

[0131] The first method is to determine the order in which the target account acquires virtual items.

[0132] In illustrative terms, the virtual items corresponding to the interaction data include a first category of items and a second category of items. The first category of items are those to be recommended, and the second category of items are those that depend on the first category of items. The interaction data set includes at least one of the first interaction data corresponding to the first category of items and the second interaction data corresponding to the second category of items. Taking the recommendation of virtual skins in a game as an example, the first category of items mentioned above are virtual skins, and the second category of items are the virtual characters or virtual items that the virtual skins affect. Therefore, when determining the weights corresponding to the data feature representations, it can be determined based on the order preference of the account when acquiring the first and second categories of items.

[0133] Indicatively, the sequential preference information of the target account within the first historical time period is obtained. The sequential preference information indicates the sequential preference of the target account when acquiring the first type of items and the second type of items with the aforementioned dependency relationship. The weights corresponding to the first interaction data and the second interaction data are determined based on the sequential preference information. The data feature representations of multiple interaction data are processed by weighted Cartesian product according to the weights to obtain the cross feature representation.

[0134] In one example, the server pre-stores a first weight combination and a second weight combination. When it is determined, based on order preference information, that the target account prefers to acquire the first type of item before acquiring the second type of item, the first weight combination is used, where the first weight corresponding to the first interaction data in the first weight combination is higher than the second weight corresponding to the second interaction data. When it is determined, based on order preference information, that the target account prefers to acquire the second type of item before acquiring the first type of item, the second weight combination is used, where the second weight corresponding to the second interaction data in the second weight combination is higher than the first weight corresponding to the first interaction data.

[0135] The second method is to determine the similarity between the interactive data.

[0136] Indicatively, the similarity between interactive data is determined; based on the above similarity, the weight corresponding to each interactive data is determined; according to the weight, the data feature representations of multiple interactive data are processed by weighted Cartesian product to obtain cross feature representation.

[0137] In some embodiments, the similarity between the interactive data is determined by calculating the feature distance between the corresponding data feature representations. Optionally, the feature distance can be at least one of cosine distance, Euclidean distance, Manhattan distance, etc., between the data feature representations.

[0138] In one example, the similarity between the obtained data feature representations is converted into weights using a specified transformation formula, where the weights are positively correlated with the similarity. In another example, the server pre-stores a mapping table between similarity intervals and weights, which is used to determine the weight corresponding to each data feature representation.

[0139] In summary, the training method for the recommendation model provided in this application, during incremental training of the item recommendation model used to predict item recommendation degree, acquires multiple sets of interaction data within a first historical time period, performs feature cross-processing on the interaction data in the interaction data sets to obtain cross-feature representations, and trains candidate recommendation models using these cross-feature representations to obtain new item recommendation models. These candidate recommendation models are trained based on multiple sets of interaction data within a second historical time period, the end time of which is chronologically prior to the end time of the first historical time period. That is, during model training, an iterative update approach is used to incrementally learn from the previous model used for item recommendation, enabling the model to directly learn from recent data based on the already trained model parameters, resulting in an item recommendation model sensitive to recent data. This improves the recommendation accuracy when using the item recommendation model and maintains the synchronization between the model and changes in the account's interest tendencies.

[0140] In related technologies, the data features commonly used in recommendation systems for game scenarios are interaction features, such as... Figure 7 The diagram illustrates the application process of interactive features in related technologies. Account identifier 701 and virtual item identifier 702 are manually processed to obtain interactive feature 703. Interactive feature 703 is stored in storage area 710. When model training is required, the interactive feature 703 is retrieved from storage area 710 and input into model 720 for training. Interactive features are manually designed features, such as the number of times an account is used on a virtual skin / virtual character; that is, text content expressing the interaction process between the account and the virtual item. Problems with interactive features include a large number of features (|account| × |item|), and significant data lag due to the large computational and time required for pre-database processing, resulting in low feature timeliness.

[0141] Compared to interactive features in related technologies, the cross features used in this application embodiment have a feature space of similar size to interactive features. However, cross features do not require complex manual processing before application; they are calculated based on raw log data. They also do not require designing interaction methods between features; instead, the model learns the interaction between accounts and items on its own. Therefore, cross features are superior to interactive features in terms of feature representation, scalability, and timeliness. Figure 8 As shown, it illustrates the application process of cross features provided in an exemplary embodiment of this application. In this case, the account identifier 801 and the item identifier 802 of the virtual item are directly stored in the storage area 810. When model training is required, the account identifier 801 and the item identifier 802 are retrieved from the storage area 810, and feature cross is performed to obtain the cross feature representation 803. The cross feature representation 803 is then input into the model 820 for training.

[0142] In the embodiments of this application, when performing feature cross processing on the interactive data in the interactive data group, Cartesian product is used to fully cross the features represented by the interactive data, thereby enabling the model to fully learn the feature representation and improve the accuracy of the model in the prediction process.

[0143] In this embodiment, the server employs an offline-online mode for training and applying the item recommendation model. That is, the training process of the candidate item recommendation model is implemented offline, while the application process is implemented online. For illustrative examples, please refer to [reference needed]. Figure 9 The diagram illustrates a flowchart of a training method for a recommendation model provided in an exemplary embodiment of this application, the method comprising:

[0144] Step 910: Obtain multiple groups of interactive data generated by the target account within the first historical time period.

[0145] Schematic illustration: The aforementioned interaction data group includes multiple interaction data points representing the interaction relationship between the target account and the item, wherein these multiple interaction data points are discrete data. The aforementioned first historical time period is a historical time period with a first moment as the end moment and a specified duration as the period.

[0146] Step 920: Perform feature cross processing on multiple interactive data in the interactive data group to obtain the cross feature representation of the interactive data group.

[0147] In illustrative terms, the above feature cross-processing instruction crosses the features indicated by the interactive data to obtain feature combinations, which are then used as features in the input model.

[0148] In some embodiments, features are extracted from multiple interactive data in the interactive data group to obtain a data feature representation corresponding to each interactive data; the cross feature representation is determined based on the Cartesian product of the data feature representations of multiple interactive data. That is, the above-mentioned cross feature representation is determined by calculating the Cartesian product of the data feature representations corresponding to the interactive data.

[0149] Step 931: Obtain the historical model parameters corresponding to the second historical time period from the online parameter storage area.

[0150] In this embodiment, the server includes a model update service and a model application service. The model update service is implemented through an offline portion of the server, and the model application service is implemented through an online portion of the server.

[0151] As an illustration, an online parameter storage area exists between the model update service and the model application service. This online parameter storage area provides model parameters for real-time item recommendation predictions. In one example, the online parameter storage area can be implemented as a parameter server, the model update service as a model update sub-server, and the model application service as a model application sub-server. Together, these three form a server cluster that provides virtual item recommendation functionality for the target application.

[0152] In this embodiment of the application, when the model update service is started, the historical model parameters corresponding to the second historical time period are obtained from the online parameter storage area.

[0153] Step 932: Load the historical model parameters into the model to be trained to obtain the candidate item recommendation model.

[0154] In this embodiment of the application, a model to be trained is deployed in the model update service. The model to be trained deployed in the model update service is used to indicate the model structure for implementing virtual item recommendation prediction. That is, the model to be trained in the model update service is a model structure used to record the item recommendation model in the service.

[0155] In some embodiments, the model parameters corresponding to the model to be trained are empty, so as to load historical model parameters obtained from the online parameter storage area.

[0156] Step 933: Use the candidate item recommendation model to predict the interest level of the cross-feature representation to obtain the interest level prediction result between the target account and the virtual item.

[0157] Optionally, the above-mentioned candidate item recommendation model can be implemented as at least one of DNN, CNN, GNN, etc.

[0158] In one example, when the above candidate item recommendation model is implemented as a DNN, the above candidate item recommendation model can be represented as shown in Formula 3.

[0159] Formula 3:

[0160] in, x represents the interest prediction result output by the model. user x represents the interactive data in the interactive data set that characterizes account features. itemx represents the interactive data that characterizes the features of virtual items in the interactive data. combine For x user and x item The cross feature representation obtained after feature crossing, Θ DNN Recommend the model parameters corresponding to the candidate item model.

[0161] Step 940: Train the candidate item recommendation model based on the interest prediction results to obtain the item recommendation model.

[0162] As an illustration, the above item recommendation model is used to recommend items to a target account.

[0163] In this embodiment of the application, the model parameters of the candidate item recommendation model are iteratively updated by determining the difference between the interest prediction result and the sample label corresponding to the interaction data group by specifying a loss function, thereby obtaining the above-mentioned item recommendation model.

[0164] In this embodiment, incremental learning is used to update the model. Compared to the data window-based model update method in related technologies, this method makes the trained item recommendation model more sensitive to recent data, resulting in more accurate predictions during application. Illustratively, the data window-based model update in related technologies can be represented as shown in Equation 4.

[0165] Formula 4:

[0166] Where T refers to the date of model update, Θ T X represents the model parameters of the DNN corresponding to this update cycle. {T-29,...,T-1,T} Θ0 represents the sample data within the data window, Θ0 represents the initialized model parameters, and y represents the sample label corresponding to the sample data.

[0167] The model update corresponding to the incremental learning used in this application embodiment can be expressed as shown in Formula 5.

[0168] Formula 5:

[0169] Where T refers to the date of model update, Θ T X represents the model parameters of the DNN corresponding to this update cycle. {T} For the sample data corresponding to the current update period, Θ T-1 Here, y represents the model parameters corresponding to the previous update cycle, and y represents the sample label corresponding to the sample data.

[0170] Step 950: Update the model parameters of the item recommendation model to the online parameter storage area.

[0171] After the model update service completes the model update for the current update cycle, the model parameters corresponding to the item recommendation model will be updated to the online parameter storage area mentioned above.

[0172] When the online parameter storage area receives the latest model parameters from the model update service, it may optionally overwrite the historical model parameters stored in the original storage with the latest model parameters; or, the online parameter storage area may not delete the historical model parameters stored in the original storage, but instead store the latest model parameters in correspondence with the time period identifier corresponding to the first historical time period.

[0173] Step 961: In response to the item recommendation request, retrieve the model parameters from the online parameter storage area.

[0174] When the model application service has an item recommendation requirement—that is, when the server receives an item recommendation request—the model application service is started and retrieves the latest model parameters from the online parameter storage area. The item recommendation request includes the account information corresponding to the target account. In one example, the account information included in the item recommendation request can be the target account's account ID.

[0175] Optionally, when the online parameter storage area stores model parameters in a mode where new parameters overwrite old parameters, the model application service directly reads the aforementioned model parameters from the online parameter storage area; when the online parameter storage area stores model parameters according to the corresponding historical time period, the model shadow service reads the corresponding model parameters from the online parameter storage area based on the current time.

[0176] In this embodiment of the application, the aforementioned item recommendation request is a request sent by the terminal to the server. Optionally, the terminal triggering the sending of the item recommendation request can be implemented in at least one of the following ways:

[0177] The first method involves automatically sending an item recommendation request to the server when the target application is launched on the terminal.

[0178] That is, when the terminal receives an application activation command for the target application, the target application automatically sends an item recommendation request to the server through the terminal.

[0179] In some embodiments, when the terminal first opens the target application within the model's update cycle, it sends the aforementioned item recommendation request to the server. For example, if the target application records that the model's update cycle is one day, then the target application sends an item recommendation request to the server via the terminal the first time it is opened each day.

[0180] The second method involves sending an item recommendation request to the server when the redemption interface for virtual items in the target application is opened.

[0181] As an illustration, the target application provides a redemption interface for virtual items. In response to receiving an activation command for the redemption interface within the target application, the terminal sends an item recommendation request to the server. For example, after the target application is launched, it enters the game's main interface, which includes various entry controls for game functions, such as game match entry, virtual store entry, and social list entry. When the control corresponding to the virtual store entry receives a trigger operation, the target application sends the aforementioned item recommendation request to the server via the terminal.

[0182] The third method involves sending an item recommendation request to the server when the recommended activation action is received on the redemption interface corresponding to the virtual item.

[0183] In some embodiments, a recommendation mode control is provided in the redemption interface. In one example, when the target application enters the redemption interface, the recommendation mode control is in a closed state, and the redemption interface displays virtual items in a default order. When the recommendation mode control receives an enable operation, the target application sends the item recommendation request to the server through the terminal, and refreshes the interface according to the recommended order list of virtual items returned by the server, displaying the virtual items arranged in the recommended order list.

[0184] Step 962: Load the model parameters into the online service model to obtain the item recommendation model.

[0185] The model application service deploys the aforementioned online service model, which is used to provide real-time predictions of item recommendation scores. In other words, this online service model deployed in the model application service indicates the model structure for implementing virtual item recommendation score predictions.

[0186] To illustrate, the model application service loads the model parameters obtained from the online parameter storage area onto the online service model deployed in the model application service, thereby obtaining the item recommendation model.

[0187] In some embodiments, the item recommendation model is used to recommend items to a target account within a specified time period, where the start time of the specified time period is after the end time of a first historical time period. Within the specified time period, an item recommendation request for the target account is received. The item recommendation request includes account information corresponding to the target account. Item information corresponding to candidate items is obtained. The account information and item information are input into the item recommendation model for prediction to obtain a recommendation score for the candidate item. Based on the recommendation score of the candidate item, item recommendations are made to the target account.

[0188] In some embodiments, when the above-mentioned item recommendation model serves the global account of the target application, in order to reduce the overhead of inter-service communication data, the model application service caches the model parameters after obtaining the model parameters for the first time within the same update cycle of the model, so as to process item recommendation requests corresponding to other accounts.

[0189] In some embodiments, when the above-mentioned item recommendation model serves a target account, the model application service, after processing the item recommendation request of the target account, resets the model parameters loaded into the online service model so as to load the model parameters in response to item recommendation requests from other accounts.

[0190] Step 971: Input the account information and item information into the item recommendation model for prediction to obtain the recommendation score corresponding to the candidate item.

[0191] In some embodiments, the model application service obtains the item information corresponding to the candidate item. In one example, the item information of the candidate item can be implemented as the item ID of the candidate item.

[0192] Optionally, the aforementioned candidate items are items selected from the item library for recommendation. In some embodiments, the items in the item library are filtered based on the items already owned by the target account; that is, items not owned by the target account in the item library are selected as candidate items.

[0193] In some embodiments, when the target application needs to display all virtual items not owned in the item library, the virtual items in the item library other than those already owned by the target account are used as the candidate items; in other embodiments, when the target application only needs to display a specified number of virtual items, the candidate items are obtained by random sampling after filtering based on the items already owned by the target account.

[0194] To illustrate, account information and item information are used as input to the item recommendation model. The item recommendation model predicts the target account's interest in each candidate item and outputs the recommendation score for each candidate item.

[0195] Step 972: Recommend items to the target account based on the recommendation scores of the candidate items.

[0196] In some embodiments, the relationship between the recommendation score of the candidate item and the recommendation threshold is used to determine whether to recommend the candidate item to the target account. That is, in response to the candidate item's recommendation score being higher than the recommendation threshold, the candidate item is indicated as a recommended item to the terminal corresponding to the target account.

[0197] In other embodiments, multiple candidate items are sorted according to their recommendation scores to obtain a recommendation order list, which is used to indicate the display order of the candidate items, wherein candidate items with higher recommendation scores are displayed first.

[0198] In summary, the training method for the recommendation model provided in this application, during incremental training of the item recommendation model used to predict item recommendation degree, acquires multiple sets of interaction data within a first historical time period, performs feature cross-processing on the interaction data in the interaction data sets to obtain cross-feature representations, and trains candidate recommendation models using these cross-feature representations to obtain new item recommendation models. These candidate recommendation models are trained based on multiple sets of interaction data within a second historical time period, the end time of which is chronologically prior to the end time of the first historical time period. That is, during model training, an iterative update approach is used to incrementally learn from the previous model used for item recommendation, enabling the model to directly learn from recent data based on the already trained model parameters, resulting in an item recommendation model sensitive to recent data. This improves the recommendation accuracy when using the item recommendation model and maintains the synchronization between the model and changes in the account's interest tendencies.

[0199] In this embodiment, offline training and online prediction are used to create a closed data flow loop for the overall personalized recommendation system. In one example, considering the recommendation of virtual skins in a game scene, such as... Figure 10The diagram illustrates a personalized recommendation system for a skin store provided in an exemplary embodiment of this application. The recommendation system 1000 includes a game server 1010 that interacts with a terminal, an offline model update service 1020, an online model application service 1030, an online parameter storage service (PS) 1040, a request splitting service 1050, a framework service 1060, and a remote dictionary service (Redis) 1070. Indicatively, the game server 1010 interacts with the terminal, receives item recommendation requests from the terminal, and forwards the item recommendation requests to the request splitting service 1050. The request splitting service 1050 splits the item recommendation requests to obtain item sorting requests indicating the sorting of virtual items. These item sorting requests are transmitted to the framework service 1060, which performs traffic distribution and caching based on configuration information obtained from the configuration center through whitelisting and A / B testing, and encapsulates them into service requests that can be received and processed by the model application service 1030. The model application service 1030 associates the account ID and candidate virtual skin ID in the request with the corresponding information by accessing the remote dictionary service 1070. The model application service 1030 extracts the latest model parameters from the online parameter storage service 1040, loads the model parameters into the online service model, and obtains the item recommendation model. The target features are then input into the item recommendation model to score and rank the candidate virtual skins. The resulting recommendation order list is then returned to the game server 1010 via the framework service 1060 and the request splitting service 1050. The game server 1010 sends the list to the terminal, allowing the terminal to display virtual skins according to the recommendation order list. The target features are also transmitted as sample features to the sample library 1021 in the model update service 1020.

[0200] For the offline portion, the game server 1010 also records the interaction data of the target account with the virtual skin, thereby determining the sample tags and interaction data. The game server 1010 sends the interaction data, including account features and virtual skin features, to the remote dictionary service 1070 through the feature injection library 1024, and sends the sample tags to the sample processing unit 1022 in the model update service 1020. The sample processing unit 1022 obtains the returned sample features from the sample library 1021, performs concatenation and sampling of the sample features and sample tags, and sends the resulting sample set to the model training unit 1023. The model training unit 1023 performs incremental learning of the model parameters based on the historical model parameters, and updates the trained model parameters to the online parameter storage service 1040.

[0201] In some embodiments, the server stores the interactive data sets used for model training in each update cycle for later training. For illustrative purposes, please refer to [reference needed]. Figure 11 The diagram illustrates a flowchart of a training method for a recommendation model provided in an exemplary embodiment of this application, the method comprising:

[0202] Step 1110: Store the interaction data group as a historical interaction data group in the sample storage area.

[0203] To illustrate, the aforementioned sample storage area is used to store the data used to train the model in all historical time periods. That is, the interactive data groups that serve as sample data in each update cycle of the model update service are stored in this sample storage area.

[0204] In some embodiments, when an interaction data group is stored as a historical interaction data group, it can be stored according to the historical time period corresponding to the interaction data group, that is, the historical interaction data is marked with time period information; in other embodiments, when an interaction data group is stored as historical interaction data, it can be stored according to the update cycle corresponding to the interaction data group, that is, the historical interaction data is marked with an update cycle identifier.

[0205] This is illustrative; please refer to it. Figure 12 It illustrates a flowchart of data processing provided by an exemplary embodiment of this application, in which log data 1201 is preprocessed to obtain interactive data set X as training samples. T 1202, will interact with data group X T 1202 is stored in the sample storage area 1210 and simultaneously input to the model training unit 1220. The model training unit 1220 obtains the model parameters to be updated from the online parameter storage area 1230 and updates the trained model parameters to the online parameter storage area 1230.

[0206] Step 1120: In response to the online service model's model structure being updated from the first structure to the second structure, the initial item recommendation model is constructed using the second structure.

[0207] In this embodiment of the application, when the model structure of the online service model in the model application service changes, that is, when the model structure of the prediction model indicated by the online service model when making item recommendations is updated from the first structure to the second structure, the model update service needs to build an initial item recommendation model with the second structure to initialize the model parameters.

[0208] In some embodiments, the designer updates the model structure of the second structure to the online service model in the model application service. When the model application service detects a change in the model structure of the online service model, it instructs the model update service on the second structure. The model update service then constructs the initial item recommendation model based on the second structure.

[0209] Step 1130: Obtain the historical interaction data set from the sample storage area.

[0210] Schematic illustration: The model update service retrieves historical interaction data sets from the sample storage area. In some embodiments, the model update service retrieves historical interaction data sets corresponding to all update periods in the sample storage area; in other embodiments, since the characteristics of data that is too old have a low contribution to the current virtual item recommendation, a specified time window can be set when retrieving historical interaction data sets. When the update period falls within the time window, the historical interaction data set corresponding to that update period is retrieved.

[0211] Step 1140: Based on the historical interaction data set, the initial item recommendation model is supplemented and trained to obtain the candidate item recommendation model with the second structure.

[0212] In some embodiments, when supplementing the training of the initial item recommendation model using historical interaction data sets, the historical interaction data corresponding to the update cycle can be used as a whole sample set for iterative training of the initial item recommendation model.

[0213] In other embodiments, incremental learning is also employed during the supplementary training process. Illustratively, time period information corresponding to multiple historical interaction data groups is determined. This time period information indicates the training time period of the historical interaction data groups when training the candidate item recommendation model of the first structure. Based on the temporal order of the multiple historical interaction data groups in the historical training process indicated by the time period information, the initial item recommendation model of the second structure is iteratively trained to obtain the candidate item recommendation model. The training time sequence between historical interaction data is positively correlated with the time sequence indicated by the time period information; that is, the historical interaction data groups with earlier training time periods indicated by the time period information are used first to train the model.

[0214] In summary, the model supplementation training method provided in this application stores the interaction data group for each update cycle. When the model structure of the item recommendation model in the server changes, the model with the new model structure is supplemented and trained using the stored historical interaction data. This enables rapid updating of the model structure, facilitates model optimization during application, and improves the accuracy of virtual item recommendations.

[0215] To verify the recommendation performance of virtual items achieved by fusing cross-features and incremental learning in the embodiments of this application, offline and online experiments were conducted. The main metric for the offline experiments was the Area Under the Curve (AUC) of the comparison model. A higher AUC indicates better model prediction performance. Illustratively, the training data of the current day was used as the validation set, and the data from the previous day and earlier was used as the training set. The method provided in this embodiment achieved an overall AUC of 0.9428 on the test set, representing a 0.72% improvement over the baseline model.

[0216] In one example, the effect of cross features was verified through a model ablation experiment, and the verification results are shown in Table 2. The table shows that removing cross features reduces model performance, indicating that incorporating cross features can improve model performance. Combining cross features with other features that can characterize account interactions can also lead to corresponding improvements.

[0217] Table 2

[0218]

[0219]

[0220] In another example, to verify the effect of incremental learning, an AUC comparison was also performed. As shown in Table 3, incremental learning has a relatively stable AUC improvement compared to window learning, with an average improvement of 0.24%.

[0221] Table 3

[0222] Dataset identifier AUC of incremental learning AUC of window learning 20220429 0.9387 0.9378 20220428 0.9408 0.9378 20220427 0.9345 0.9416 20220426 0.9448 0.9432 20220425 0.9475 0.9441 20220424 0.9471 0.9435 20220423 0.9462 0.9351

[0223] Meanwhile, incremental learning only requires processing one day's worth of sample data, and the model updates faster, taking less than 1 / 4 of the time of window learning. In one example, as shown in Table 4, a comparison of training time between window training and incremental training is presented.

[0224] Table 4

[0225]

[0226] In another example, the diversity of the methods in this application can also be analyzed according to Formula 6.

[0227] Formula Six:

[0228] Where, q ab(L) indicates the number of identical virtual items in the TOP-L recommendation results for accounts a and b, where U is the set of accounts. A higher Personalization@L value indicates greater diversity in the recommendation results. The baseline model is used for comparison, and the results are shown in Table 5. The comparison shows that the model provided by the method in this embodiment has better diversity in leg and foot results, and the model's recommendations are more personalized.

[0229] Table 5

[0230] Baseline This method personalization@6 0.8723 0.9126 personalization@12 0.8284 0.8724

[0231] It should be noted that this application may display prompt interfaces, pop-ups, or output voice prompts before and during the collection of user-related data (e.g., account interaction data with virtual items). These prompt interfaces, pop-ups, or voice prompts are used to inform the user that their data is being collected. This ensures that the application only begins the steps for collecting user-related data after receiving confirmation from the user regarding the prompt interface or pop-up; otherwise (i.e., without user confirmation), the steps for collecting user-related data end, meaning no user-related data is collected. In other words, all user data collected in this application is collected with the user's consent and authorization, and the collection, use, and processing of related user data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0232] Please refer to Figure 13 The diagram illustrates a structural block diagram of a training apparatus for a recommendation model provided in an exemplary embodiment of this application. The apparatus includes the following modules:

[0233] The acquisition module 1310 is used to acquire multiple interaction data groups generated by the target account within a first historical time period. The interaction data group includes multiple interaction data, which are used to represent the interaction relationship between the target account and virtual items.

[0234] Processing module 1320 is used to perform feature cross-processing on multiple interactive data in the interactive data group to obtain the cross-feature representation of the interactive data group;

[0235] Training module 1330 is used to predict the interest degree of the cross feature representation through the candidate item recommendation model to obtain the interest degree prediction result between the target account and the virtual item. The candidate item recommendation model is a model trained based on multiple interaction data groups generated by the target account in the second historical time period. The end time of the second historical time period is located before the end time of the first historical time period in time sequence.

[0236] The training module 1330 is further configured to train the candidate item recommendation model based on the interest prediction result to obtain an item recommendation model, which is used to recommend virtual items to the target account in the game store.

[0237] In some optional embodiments, the processing model 1320 is further configured to extract features from multiple interactive data in the interactive data group to obtain a data feature representation corresponding to each interactive data; and to determine the cross feature representation based on the Cartesian product between the data feature representations of the multiple interactive data.

[0238] In some optional embodiments, the processing model 1320 is further configured to obtain a first candidate feature representation corresponding to the first i interactive data in the interactive data group, wherein the first candidate feature representation is a continuous Cartesian product corresponding to the first i interactive data; to perform a Cartesian product between the data feature representation corresponding to the (i+1)th interactive data and the first candidate feature representation to obtain the cross feature representation, where i is a positive integer; or, to perform a Cartesian product between the data feature representation of the i-th interactive data and the data feature representation of the j-th interactive data in the interactive data group to obtain a second candidate feature representation, where i and j are unequal positive integers; and to perform feature concatenation between the second candidate feature representations between every two interactive data in the interactive data group to obtain the cross feature representation.

[0239] In some optional embodiments, the virtual items include a first type of items and a second type of items, wherein the first type of items are items to be recommended, and the second type of items are virtual items that have a dependency relationship with the first type of items, and the interaction data group includes at least one of a first interaction data corresponding to the first type of items and a second interaction data corresponding to the second type of items;

[0240] like Figure 14 As shown, the processing model 1320 further includes:

[0241] The acquisition unit 1321 is used to acquire the sequential preference information of the target account in the first historical time period, wherein the sequential preference information indicates the sequential preference of the target account when acquiring the first type of item and the second type of item with the dependency relationship.

[0242] The determining unit 1322 is used to determine the weights corresponding to the first interactive data and the second interactive data respectively based on the order preference information;

[0243] The calculation unit 1323 is used to perform a weighted Cartesian product processing on the data feature representations of the multiple interactive data according to the weights to obtain the cross feature representation.

[0244] In some optional embodiments, the acquisition module 1310 is further configured to acquire historical model parameters corresponding to the second historical time period;

[0245] The training module 1330 is also used to load the historical model parameters onto the model to be trained to obtain the candidate item recommendation model.

[0246] In some optional embodiments, the item recommendation model is used to recommend items to the target account within a specified time period, wherein the start time of the specified time period is chronologically located after the end time of the first historical time period.

[0247] The device further includes:

[0248] The receiving module 1340 is configured to receive an item recommendation request for the target account within the specified time period, wherein the item recommendation request includes account information corresponding to the target account;

[0249] The acquisition module 1310 is also used to acquire item information corresponding to the candidate item;

[0250] The recommendation module 1350 is used to input the account information and the item information into the item recommendation model for prediction to obtain the recommendation score corresponding to the candidate item; and to recommend items to the target account based on the recommendation score of the candidate item.

[0251] In some alternative embodiments, the apparatus further includes:

[0252] The update module 1360 is used to update the model parameters of the item recommendation model to the online parameter storage area, which is used to provide model parameters for real-time item recommendation prediction.

[0253] In some optional embodiments, the acquisition module 1310 is further configured to acquire the model parameters from the online parameter storage area in response to the item recommendation request;

[0254] The recommendation module 1350 is further configured to load the model parameters into the online service model to obtain the item recommendation model, wherein the online service model is a model used to provide real-time prediction of item recommendation degree.

[0255] In some optional embodiments, the acquisition module 1310 is further configured to acquire log data of the target account during the first historical time period, wherein the log data is used to record the interaction between the target account and virtual items;

[0256] The processing module 1320 is further configured to perform at least one preprocessing operation on the log data to obtain the plurality of interactive data groups;

[0257] The preprocessing operations include bucketing, normalization, truncation, and hashing. The bucketing operation is used to split the continuous information represented by the log data to achieve data discretization. The normalization operation is used to unify the values ​​of different categories of information in the log data to a specified value range. The truncation operation is used to filter out abnormal data in the log data. The hashing operation is used to convert the log data into a vector representation or a matrix representation.

[0258] In some alternative embodiments, the apparatus further includes:

[0259] Storage module 1370 is used to store the interaction data group as a historical interaction data group in the sample storage area, and the sample storage area is used to store the training data of the model in all historical time periods;

[0260] The training module 1330 is also used to respond to the online service model's model structure being updated from a first structure to a second structure, and to construct an initial item recommendation model using the second structure. The online service model is a model in the server used to provide real-time prediction of item recommendation scores.

[0261] The acquisition module 1310 is further configured to acquire the historical interaction data group from the sample storage area;

[0262] The training module 1330 is further configured to supplement and train the initial item recommendation model based on the historical interaction data set to obtain the candidate item recommendation model of the second structure.

[0263] In some optional embodiments, the training module 1330 is further configured to determine time period information corresponding to multiple historical interaction data groups, wherein the time period information is used to indicate the training time period of the historical interaction data groups when training the candidate item recommendation model of the first structure; and to iteratively train the initial item recommendation model of the second structure according to the temporal order of the multiple historical interaction data groups in the historical training process indicated by the time period information to obtain the candidate item recommendation model.

[0264] In some optional embodiments, the training module 1330 is further configured to randomly sample the multiple sets of interaction data to determine the target interaction data set; input the cross-feature representation of the target interaction data set into the candidate item recommendation model to perform interest prediction and obtain the interest prediction result.

[0265] In summary, the training apparatus for the recommendation model provided in this application, during incremental training of the item recommendation model used to predict item recommendation rates, acquires multiple sets of interaction data within a first historical time period, performs feature cross-processing on the interaction data in the interaction data sets to obtain cross-feature representations, and trains candidate recommendation models using these cross-feature representations to obtain new item recommendation models. These candidate recommendation models are trained based on multiple sets of interaction data within a second historical time period, the end time of which is chronologically prior to the end time of the first historical time period. That is, during model training, an iterative update method is used to incrementally learn from the previous model used for item recommendation, enabling the model to directly learn from recent data based on the already trained model parameters, resulting in an item recommendation model sensitive to recent data. This improves the recommendation accuracy when using the item recommendation model and maintains the synchronization between the model and changes in the account's interest tendencies.

[0266] It should be noted that the training device for the recommendation model provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the training device for the recommendation model provided in the above embodiments and the method embodiments for recommending items belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0267] Figure 15 This illustration shows a schematic diagram of the structure of a server provided in an exemplary embodiment of this application. Specifically, it includes the following structure.

[0268] Server 1500 includes a Central Processing Unit (CPU) 1501, a system memory 1504 including Random Access Memory (RAM) 1502 and Read Only Memory (ROM) 1503, and a system bus 1505 connecting the system memory 1504 and the CPU 1501. Server 1500 also includes a mass storage device 1506 for storing an operating system 1513, application programs 1514, and other program modules 1515.

[0269] Mass storage device 1506 is connected to central processing unit 1501 via a mass storage controller (not shown) connected to system bus 1505. Mass storage device 1506 and its associated computer-readable media provide non-volatile storage for server 1500. That is, mass storage device 1506 may include computer-readable media (not shown) such as hard disk or compact disc read-only memory (CD-ROM) drive.

[0270] Without loss of generality, computer-readable media can include computer storage media and communication media. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer storage media include RAM, ROM, erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other solid-state memory technologies, CD-ROM, digital versatile disc (DVD) or other optical storage, magnetic tape cassettes, magnetic tape, disk storage, or other magnetic storage devices. Of course, those skilled in the art will recognize that computer storage media are not limited to the above-mentioned types. The system memory 1504 and mass storage device 1506 described above can be collectively referred to as memory.

[0271] According to various embodiments of this application, server 1500 can also be connected to a remote computer on a network, such as the Internet. That is, server 1500 can be connected to network 1512 via network interface unit 1511 connected to system bus 1505, or it can also use network interface unit 1511 to connect to other types of networks or remote computer systems (not shown).

[0272] The aforementioned memory also includes one or more programs, which are stored in the memory and configured to be executed by the CPU.

[0273] Embodiments of this application also provide a computer device including a processor and a memory. The memory stores at least one instruction, at least one program, code set, or instruction set. The processor loads and executes the at least one instruction, at least one program, code set, or instruction set to implement the training method of the recommended model provided in the above-described method embodiments. Optionally, the computer device may be a terminal or a server.

[0274] Embodiments of this application also provide a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the training method of the recommended model provided in the above-described method embodiments.

[0275] Embodiments of this application also provide a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the training method of any of the recommended models described in the above embodiments.

[0276] Optionally, the computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), solid-state drives (SSDs), or optical discs, etc. The random access memory may include resistive random access memory (ReRAM) and dynamic random access memory (DRAM). The sequence numbers of the embodiments in this application are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0277] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0278] The above description is merely an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for training a recommendation model, characterized in that, The method includes: Obtain multiple interaction data groups generated by the target account within a first historical time period. Each interaction data group includes multiple interaction data, which represent the interaction relationship between the target account and virtual items. The multiple interaction data in the interaction data group are discrete among each other. The interaction data in the interaction data group is subjected to feature cross processing to obtain the cross feature representation of the interaction data group. The feature cross processing includes at least one of inner product, Hadamard product and Cartesian product on the data feature representations corresponding to the multiple interaction data. The historical model parameters corresponding to the second historical time period are obtained from the online parameter storage area. The online parameter storage area is used to provide model parameters for real-time item recommendation prediction. The end time of the second historical time period is located before the end time of the first historical time period in time sequence. The historical model parameters are trained based on multiple interaction data groups generated by the target account in the second historical time period. The historical model parameters are loaded into the model to be trained to obtain the candidate item recommendation model; The interest degree prediction result between the target account and the virtual item is obtained by using the candidate item recommendation model to predict the interest degree of the cross feature representation. The candidate item recommendation model is trained based on the interest prediction results to obtain an item recommendation model, which is used to recommend virtual items to the target account in the game store. Update the model parameters of the item recommendation model to the online parameter storage area.

2. The method according to claim 1, characterized in that, The step of performing feature cross-processing on multiple interactive data in the interactive data group to obtain the cross-feature representation of the interactive data group includes: Feature extraction is performed on multiple interactive data in the interactive data group to obtain the data feature representation corresponding to each interactive data; The cross feature representation is determined based on the Cartesian product of the data feature representations of the multiple interactive data.

3. The method according to claim 2, characterized in that, The Cartesian product of the data feature representations based on the multiple interactive data is used to determine the cross feature representation, including: Obtain the first candidate feature representation corresponding to the first i interactive data in the interactive data group. The first candidate feature representation is the continuous Cartesian product corresponding to the first i interactive data. Take the Cartesian product of the data feature representation corresponding to the (i+1)th interactive data and the first candidate feature representation to obtain the cross feature representation, where i is a positive integer. or, The second candidate feature representation is obtained by performing a Cartesian product on the data feature representation of the i-th interactive data and the data feature representation of the j-th interactive data in the interactive data group, where i and j are unequal positive integers; the second candidate feature representation between every two interactive data in the interactive data group is concatenated to obtain the cross feature representation.

4. The method according to claim 2, characterized in that, The virtual items include a first type of items and a second type of items. The first type of items are items to be recommended, and the second type of items are virtual items that have a dependency relationship with the first type of items. The interaction data group includes at least one of a first interaction data corresponding to the first type of items and a second interaction data corresponding to the second type of items. The Cartesian product of the data feature representations based on the multiple interactive data is used to determine the cross feature representation, including: Obtain the sequential preference information of the target account within the first historical time period, wherein the sequential preference information indicates the sequential preference of the target account when acquiring the first type of item and the second type of item with the dependency relationship; The weights corresponding to the first interactive data and the second interactive data are determined based on the order preference information. The data feature representations of the multiple interactive data are weighted and processed by a Cartesian product based on the weights to obtain the cross feature representation.

5. The method according to any one of claims 1 to 4, characterized in that, The item recommendation model is used to recommend items to the target account within a specified time period, wherein the start time of the specified time period is chronologically after the end time of the first historical time period, and the method further includes: Within the specified time period, receive item recommendation requests for the target account, the item recommendation requests including account information corresponding to the target account; Retrieve item information corresponding to candidate items; The account information and the item information are input into the item recommendation model for prediction to obtain the recommendation score corresponding to the candidate item; Based on the recommendation scores of the candidate items, items are recommended to the target account.

6. The method according to claim 5, characterized in that, Before inputting the account information and the item information into the item recommendation model for prediction to obtain the recommendation score corresponding to the candidate item, the method further includes: In response to the item recommendation request, the model parameters are obtained from the online parameter storage area; The model parameters are loaded into the online service model to obtain the item recommendation model, which is used to provide real-time prediction of item recommendation scores.

7. The method according to any one of claims 1 to 4, characterized in that, The acquisition of multiple interaction data groups generated by the target account within the first historical time period includes: Obtain log data of the target account within the first historical time period, wherein the log data is used to record the interaction between the target account and virtual items; Perform at least one preprocessing operation on the log data to obtain the plurality of interactive data groups; The preprocessing operations include bucketing, normalization, truncation, and hashing. The bucketing operation is used to split the continuous information represented by the log data to achieve data discretization. The normalization operation is used to unify the values ​​of different categories of information in the log data to a specified value range. The truncation operation is used to filter out abnormal data in the log data. The hashing operation is used to convert the log data into a vector representation or a matrix representation.

8. The method according to claim 7, characterized in that, The method further includes: The interaction data set is stored as a historical interaction data set in the sample storage area, which is used to store the training data of the model in all historical time periods. In response to the online service model, the model structure is updated from the first structure to the second structure, and the initial item recommendation model is constructed using the second structure. The online service model is a model in the server used to provide real-time prediction of item recommendation degree. Retrieve the historical interaction data group from the sample storage area; The initial item recommendation model is trained by supplementing the historical interaction data set to obtain the candidate item recommendation model of the second structure.

9. The method according to claim 8, characterized in that, The step of supplementing and training the initial item recommendation model based on the historical interaction data set to obtain the candidate item recommendation model includes: Determine the time period information corresponding to multiple historical interaction data groups respectively. The time period information is used to indicate the training time period of the historical interaction data group when training the candidate item recommendation model of the first structure. Based on the temporal order of the multiple historical interaction data groups indicated by the time period information during the historical training process, the initial item recommendation model of the second structure is iteratively trained to obtain the candidate item recommendation model.

10. The method according to any one of claims 1 to 4, characterized in that, The step of predicting the interest level between the target account and the virtual item by using a candidate item recommendation model to perform interest level prediction on the cross feature representation includes: Randomly sample the multiple interactive data groups to determine the target interactive data group; The cross-feature representation of the target interaction data group is input into the candidate item recommendation model to predict interest level, and the interest level prediction result is obtained.

11. A training device for a recommendation model, characterized in that, The device includes: The acquisition module is used to acquire multiple interaction data groups generated by the target account within a first historical time period. The interaction data group includes multiple interaction data, which are used to represent the interaction relationship between the target account and virtual items. The multiple interaction data in the interaction data group are discrete. The processing module is used to perform feature cross processing on multiple interactive data in the interactive data group to obtain the cross feature representation of the interactive data group. The feature cross processing includes at least one of inner product, Hadamard product, and Cartesian product on the data feature representations corresponding to the multiple interactive data. The acquisition module is further configured to acquire historical model parameters corresponding to the second historical time period from the online parameter storage area. The online parameter storage area is used to provide model parameters for real-time item recommendation prediction. The end time of the second historical time period is located before the end time of the first historical time period in time sequence. The historical model parameters are trained based on multiple interaction data groups generated by the target account in the second historical time period. The training module is used to load the historical model parameters into the model to be trained to obtain the candidate item recommendation model; The training module is also used to predict the interest degree of the cross feature representation through the candidate item recommendation model, so as to obtain the interest degree prediction result between the target account and the virtual item; The training module is also used to train the candidate item recommendation model based on the interest prediction result to obtain an item recommendation model, which is used to recommend virtual items to the target account in the game store. The update module is used to update the model parameters of the item recommendation model to the online parameter storage area.

12. The apparatus according to claim 11, characterized in that, The processing module is further configured to extract features from multiple interactive data in the interactive data group to obtain a data feature representation corresponding to each interactive data; and to determine the cross feature representation based on the Cartesian product of the data feature representations of the multiple interactive data.

13. The apparatus according to claim 12, characterized in that, The processing module is further configured to obtain a first candidate feature representation corresponding to the first i interactive data in the interactive data group, wherein the first candidate feature representation is a continuous Cartesian product corresponding to the first i interactive data; to obtain the cross feature representation by taking the Cartesian product of the data feature representation corresponding to the (i+1)th interactive data and the first candidate feature representation, where i is a positive integer; or, to obtain a second candidate feature representation by taking the Cartesian product of the data feature representation of the i-th interactive data and the data feature representation of the j-th interactive data in the interactive data group, where i and j are unequal positive integers; and to perform feature concatenation between the second candidate feature representations of every two interactive data in the interactive data group to obtain the cross feature representation.

14. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing at least one program, which is loaded and executed by the processor to implement the training method of the recommendation model as described in any one of claims 1 to 10.

15. A computer-readable storage medium, characterized in that, The storage medium stores at least one program segment, which is loaded and executed by a processor to implement the training method of the recommendation model as described in any one of claims 1 to 10.

16. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the training method of the recommendation model as described in any one of claims 1 to 10.

Images