Virtual scene interaction mode processing method and device, equipment and storage medium

Abstract

The application provides a processing method, device and equipment of a virtual scene interaction mode, and a storage medium. The method comprises: obtaining interaction mode data and recommendation object data of a plurality of candidate interaction modes; performing feature extraction on the interaction mode data and the recommendation object data respectively to obtain interaction mode features of each candidate interaction mode and object features of a recommendation object; performing natural target prediction of a plurality of target tasks based on the object features, determining a first recommendation index based on a first prediction probability of each target task obtained by the prediction; performing intervention target prediction of the plurality of target tasks for each interaction mode feature and object feature respectively to obtain a second recommendation index corresponding to each candidate interaction mode; and selecting a target interaction mode from the plurality of candidate interaction modes based on the first recommendation index and the second recommendation index of each candidate interaction mode. Through the application, the accuracy of the recommended interaction mode in the virtual scene can be improved.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and computer technology, and in particular to a method, apparatus, device and storage medium for processing virtual scene interaction modes. Background Technology

[0002] In the current gaming industry, to enhance user experience and increase user engagement, suitable interactive content is typically intelligently selected and pushed to players within the virtual environment based on factors such as player behavior and preferences. Interactive mode recommendations are applied across various online games, using different interaction paradigms and storyline progressions to increase user interest. However, current technologies rely on the experience of game operators or developers, manually setting rules to determine the timing and recipients of recommendations. This lacks flexibility and adaptability, making it difficult to cope with complex and ever-changing game environments, and it fails to consider the relevance between different types of interactive modes and players.

[0003] Currently, there is no better way to consider the impact of interaction patterns on users from multiple perspectives and achieve personalized interaction pattern recommendations for different users. Summary of the Invention

[0004] This application provides a method, apparatus, device, and storage medium for processing virtual scene interaction modes, which can improve the accuracy of recommended interaction modes in virtual scenes.

[0005] The technical solution of this application embodiment is implemented as follows:

[0006] This application provides a method for processing virtual scene interaction modes, the method comprising:

[0007] Acquire interaction pattern data and recommendation object data for multiple candidate interaction patterns, wherein the candidate interaction patterns are interaction patterns between virtual objects in the virtual scene;

[0008] Feature extraction is performed on the interaction pattern data and the recommendation object data respectively to obtain the interaction pattern features of each candidate interaction pattern and the object features of the recommendation object;

[0009] Based on the object features, natural target prediction is performed for various target tasks, and a first recommendation index is determined based on the first prediction probability of each target task obtained from the prediction.

[0010] For each interaction mode feature and object feature, the intervention target prediction of the multiple target tasks is performed to obtain the second recommendation index corresponding to each candidate interaction mode;

[0011] Based on the first recommendation index and the second recommendation index for each of the candidate interaction modes, a target interaction mode is selected from the plurality of candidate interaction modes, wherein the target interaction mode is used to perform recommendation processing on the recommended object.

[0012] This application provides a processing device for a virtual scene interaction mode, comprising:

[0013] The feature extraction module is used to acquire interaction pattern data and recommendation object data of multiple candidate interaction patterns, wherein the candidate interaction patterns are interaction patterns between virtual objects in the virtual scene; and to extract features from the interaction pattern data and the recommendation object data respectively to obtain the interaction pattern features of each candidate interaction pattern and the object features of the recommendation object.

[0014] The recommendation module is used to predict natural targets for multiple target tasks based on the object features, and determine a first recommendation index based on a first prediction probability for each target task obtained from the prediction; to predict intervention targets for the multiple target tasks for each interaction pattern feature and the object features respectively, and obtain a second recommendation index corresponding to each candidate interaction pattern; and to select a target interaction pattern from the multiple candidate interaction patterns based on the first recommendation index and the second recommendation index for each candidate interaction pattern, wherein the target interaction pattern is used to perform recommendation processing on the recommended object.

[0015] This application provides an electronic device, the electronic device comprising:

[0016] Memory is used to store executable instructions or computer programs.

[0017] The processor, when executing computer-executable instructions or computer programs stored in the memory, implements the processing method for the virtual scene interaction mode provided in the embodiments of this application.

[0018] This application provides a computer-readable storage medium storing a computer program or computer-executable instructions, which, when executed by a processor, implements the processing method for the virtual scene interaction mode provided in this application.

[0019] This application provides a computer program product, including a computer program or computer executable instructions. When the computer program or computer executable instructions are executed by a processor, they implement the processing method for the virtual scene interaction mode provided in this application.

[0020] The embodiments of this application have the following beneficial effects:

[0021] By extracting features from interaction pattern data and recommendation object data of multiple candidate interaction patterns, player preferences and needs can be more accurately identified and understood. Natural target prediction for multiple target tasks is performed based on object features. A first recommendation index is determined based on the first predicted probability of each target task. Furthermore, intervention target prediction for multiple target tasks is performed for each interaction pattern feature and object feature, yielding a second recommendation index for each candidate interaction pattern. This allows for the evaluation of the impact of different interaction patterns on the recommended object. Based on the first recommendation index and the second recommendation index for each candidate interaction pattern, the target interaction pattern is selected from multiple candidate interaction patterns to determine the most attractive target interaction pattern for recommendation, reducing ineffective interaction pattern recommendations and saving time and resources. Accurate determination of the target interaction pattern helps improve user engagement and retention rates, thereby enhancing the accuracy of interaction pattern recommendations. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating the application mode of the virtual scene interaction mode processing method provided in the embodiments of this application;

[0023] Figure 2 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application;

[0024] Figure 3A This is a schematic diagram of the first process of the virtual scene interaction mode processing method provided in the embodiments of this application;

[0025] Figure 3B This is a second flowchart illustrating the processing method for the virtual scene interaction mode provided in the embodiments of this application;

[0026] Figure 3C This is a schematic diagram of the third process of the virtual scene interaction mode processing method provided in the embodiments of this application;

[0027] Figure 3D This is a schematic diagram of the fourth process of the virtual scene interaction mode processing method provided in the embodiments of this application;

[0028] Figure 4 This is a schematic diagram of the fifth process of the virtual scene interaction mode processing method provided in the embodiments of this application;

[0029] Figure 5A This is a first structural schematic diagram of the neural network model provided in the embodiments of this application;

[0030] Figure 5B This is a schematic diagram of the second structure of the neural network model provided in the embodiments of this application;

[0031] Figure 6This is a schematic diagram of the sixth process of the virtual scene interaction mode processing method provided in the embodiments of this application.

[0032] It should be noted that the terms "first" and "second" mentioned above are only used to distinguish between different options and do not represent the degree of superiority or inferiority of the options or their priority in the implementation process. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0035] In the following description, the terms "first, second, third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0036] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0037] Unless otherwise defined, all technical and scientific terms used in the embodiments of this application have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in the embodiments of this application is for the purpose of describing the embodiments of this application only and is not intended to limit this application.

[0038] In this application embodiment, the collection and processing of relevant data (e.g., recommendation object data and interaction mode data) should be strictly in accordance with the requirements of relevant laws and regulations, obtain the informed consent or separate consent of the personal information subject, and carry out subsequent data use and processing within the scope of laws and regulations and the authorization of the personal information subject.

[0039] Before providing a further detailed description of the embodiments of this application, the nouns and terms involved in the embodiments of this application will be explained, and the nouns and terms involved in the embodiments of this application shall be interpreted as follows.

[0040] 1) Virtual scenes: These are scenes that differ from the real world, which are output by devices. Visual perception of virtual scenes can be formed with the naked eye or with the assistance of devices. For example, two-dimensional images are output through a display screen, and three-dimensional images are output through stereoscopic display technologies such as stereoscopic projection, virtual reality, and augmented reality. In addition, various possible hardware can be used to form various perceptions that simulate the real world, such as auditory perception, tactile perception, olfactory perception, and motion perception.

[0041] 2) Uplift Model: A model used to estimate the increment of an individual intervention, used to predict the causal effect of an intervention on an individual's state or behavior.

[0042] 3) Extraction Network: This refers to a network layer specifically designed to extract useful information from input data. In the gain model, the extraction network helps to identify and process key information in player features.

[0043] 4) Mixture of Experts (MoE): This is a sparse-gated deep learning model, primarily composed of a set of expert networks and a gating network. The basic idea of ​​MoE is to divide the input data into multiple regions based on the task type and assign one or more expert networks to each region. Each expert network can focus on processing its assigned region of input data, thereby improving the overall performance of the model.

[0044] 5) Expert Networks: These are pre-trained sub-networks (neural networks or layers) that specialize in handling specific data or tasks. During training, the input data is assigned to different expert networks by a gating model for processing. During inference, the experts selected by the gating model generate corresponding outputs based on the input data. These outputs are then weighted and combined with the weights assigned to each expert network based on its ability to process that feature, forming the final prediction result. Each expert network specializes in processing a specific type of data and is responsible for learning complex patterns in the data.

[0045] 6) Gating Network: A selector that routes input data to relevant experts. In a hybrid expert network, the "gate" is a sparse gate network that takes a single data element as input and outputs a weight that represents the contribution of each expert network to processing the input data.

[0046] 7) User Retention Rate: The ratio of users who continue to use an application to new users of that application. Users who start using an application within a certain period and continue to use it after a period of time are considered retained users. The proportion of these users to the new users at that time is the retention rate, which is calculated at intervals of 1 unit of time (e.g., day, week, month). In this embodiment, the short-term user retention rate is the next-day retention rate, which is (the number of new users on the day that still logs in on the first day after the initial login) / the total number of new users on the first day. The long-term user retention rate is the seventh-day retention rate, which is (the number of new users on the day that still logs in on the seventh day after the initial login) / the total number of new users on the first day.

[0047] In related technologies, the recommendation scheme for interaction modes is determined by manually formulated rules. These rules rely on the human experience of game operators or developers, manually setting rules to determine the timing and target audience for interaction modes. This approach lacks flexibility and adaptability, failing to consider the impact of interaction modes on different players from the perspective of the entire game market and all user groups. It struggles to cope with complex and ever-changing game environments and player behaviors, and does not consider the correlation between different types of interaction modes and players.

[0048] This application provides a method for processing virtual scene interaction modes, a device for processing virtual scene interaction modes, an electronic device, a computer-readable storage medium, and a computer program product, which can improve the accuracy of recommended interaction modes in virtual scenes.

[0049] The following describes exemplary applications of the electronic devices provided in the embodiments of this application. These devices can be implemented as various types of terminals such as laptops, tablets, desktop computers, set-top boxes, smartphones, smart speakers, smartwatches, smart TVs, and in-vehicle terminals, or as servers. Exemplary applications when the device is implemented as a terminal or server will be described below.

[0050] See Figure 1 , Figure 1 This is a schematic diagram illustrating the application mode of the virtual scene interaction mode processing method provided in the embodiments of this application. It is an example of how to implement a processing method for supporting a virtual scene interaction mode. Figure 1 The system involves server 200, network 300, terminal device 400 and database 500. Terminal device 400 is connected to server 200 through network 300. Network 300 can be a wide area network or a local area network, or a combination of both.

[0051] In some embodiments, the user may be a person skilled in the art or a player, the server 200 is a server for executing a processing method for virtual scene interaction modes, the terminal device 400 is a terminal operated by the user, the terminal device 400 is pre-configured with an application for displaying virtual scenes, and the database 500 stores interaction mode data and recommended object data.

[0052] For example, a user triggers a terminal device to send an interaction mode processing command. The terminal device 400 sends the interaction mode processing command to the server 200 via the network 300. The server 200 reads interaction mode data and recommended object data from the database 500 and performs feature extraction. Based on the extracted interaction mode features and object features, it performs natural target prediction to obtain a first recommendation index and intervention target prediction to obtain a second recommendation index for each candidate interaction mode. The target interaction mode is selected based on the difference between the first recommendation index and the second recommendation index. The server 200 sends the target interaction mode to the terminal device 400 via the network 300. Based on the target interaction mode, the interaction processing between virtual objects is displayed on the terminal device 400.

[0053] In some embodiments, the virtual scene interaction mode processing method of this application can also be applied to the following application scenarios:

[0054] 1. In the virtual game mode recommendation scenario: the target interaction mode obtained by the virtual scene interaction mode processing method of this application embodiment is used as the recommended game mode and recommended to the player in the virtual game scene. Different game modes are recommended to different players in a personalized way to improve player interest and user retention rate.

[0055] 2. In the application scenarios of online education platforms: the target interaction mode obtained by the virtual scene interaction mode processing method of this application embodiment is used as a recommended course mode and recommended to students on the education platform. Different course modes are recommended in a personalized way according to the learning styles and abilities of different students, thereby improving students' interest in the course and learning efficiency.

[0056] 3. In the application scenario of e-commerce platforms: The target interaction mode obtained by the virtual scene interaction mode processing method of this application embodiment is used as a discount scheme to recommend to consumers. It is recommended to consumers on the e-commerce platform, and different discount schemes are recommended in a personalized way according to the shopping habits and preferences of different consumers, so as to meet the consumers' purchasing intentions and improve the accuracy of recommendation processing.

[0057] This application embodiment can be implemented using database technology. A database, simply put, can be viewed as an electronic filing cabinet storing electronic files, where users can perform operations such as adding, querying, updating, and deleting data. A "database" is a collection of data stored together in a certain way, capable of being shared by multiple users, having minimal redundancy, and being independent of application programs.

[0058] A Database Management System (DBMS) is a computer software system designed to manage databases, generally possessing basic functions such as storage, retrieval, security, and backup. DBMSs can be classified according to the database model they support, such as relational or XML (Extensible Markup Language); or according to the type of computer they support, such as server clusters or mobile devices; or according to the query language used, such as Structured Query Language (SQL) or XQuery; or according to performance priorities, such as maximum scale or maximum operating speed; or other classification methods. Regardless of the classification method used, some DBMSs can cross categories, for example, simultaneously supporting multiple query languages.

[0059] See Figure 2 , Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may be... Figure 1 Server 200 in the middle, Figure 2 The server 200 shown includes at least one processor 410, memory 450, and at least one network interface 420. The various components of server 200 are coupled together via a bus system 440. It is understood that the bus system 440 is used to implement communication between these components. In addition to a data bus, the bus system 440 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 2 The general labeled all buses as Bus System 440.

[0060] The processor 410 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0061] The memory 450 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage, hard disk drives, optical disk drives, etc. The memory 450 may optionally include one or more storage devices physically located away from the processor 410.

[0062] The memory 450 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 450 described in this application embodiment is intended to include any suitable type of memory.

[0063] In some embodiments, memory 450 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0064] Operating system 451 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks;

[0065] The network communication module 452 is used to reach other electronic devices via one or more (wired or wireless) network interfaces 420, exemplary network interfaces 420 including Bluetooth, Wi-Fi, and Universal Serial Bus (USB).

[0066] In some embodiments, the apparatus provided in this application can be implemented in software. Figure 2 A processing device 455 is shown that stores virtual scene interaction modes in memory 450. This can be software in the form of programs and plugins, including the following software modules: a feature extraction module 4551, a recommendation module 4552, and a model training module 4553. These modules are logically linked and can therefore be arbitrarily combined or further split according to their implemented functions. Figure 2 For ease of explanation, all the above modules are shown at once. However, this should not be interpreted as the processing device 455 in the virtual scene interaction mode excluding the implementation that may only include the feature extraction module 4551 and the recommendation module 4552. The functions of each module will be explained below.

[0067] In some embodiments, the terminal or server can implement the virtual scene interaction mode processing method provided in this application embodiment by running various computer-executable instructions or computer programs. For example, computer-executable instructions can be microprogram-level commands, machine instructions, or software instructions. Computer programs can be native programs or software modules in an operating system; they can be native applications (APPs), i.e., programs that need to be installed in the operating system to run; or they can be applets that can be embedded in any APP, i.e., programs that only need to be downloaded to a browser environment to run. In summary, the aforementioned computer-executable instructions can be any form of instruction, and the aforementioned computer programs can be any form of application, module, or plugin.

[0068] The processing method of the virtual scene interaction mode provided in the embodiments of this application will be described in conjunction with the exemplary application and implementation of the terminal provided in the embodiments of this application.

[0069] The following describes the processing method for the virtual scene interaction mode provided in the embodiments of this application. As mentioned above, the electronic device implementing the processing method for the virtual scene interaction mode in the embodiments of this application can be a terminal, a server, or a combination of both. Therefore, the executing entity of each step will not be described again below.

[0070] See Figure 3A , Figure 3A This is a first flowchart illustrating the processing method for virtual scene interaction mode provided in the embodiments of this application, which will be combined with... Figure 3A The steps shown are explained.

[0071] In step 301, interaction pattern data and recommendation object data of multiple candidate interaction patterns are obtained.

[0072] Here, the candidate interaction mode is the interaction mode between virtual objects in the virtual scene.

[0073] For example, this application uses a game virtual scene as an example for explanation. The recommended object can be a user. In the game virtual scene, the virtual objects controlled by the user interact in different interaction modes, generating various types of interaction data. The interaction mode in the virtual scene is used to guide the behavior and dialogue of virtual objects, including the storyline and virtual object settings, as well as the rules and logic for how virtual objects react according to the virtual scene and behavior. In the interaction mode, virtual objects are allowed to communicate and interact dynamically. Based on the preset plot and real-time scenario of the interaction mode, a user experience combining the plot and virtual object selection is constructed. Each user has an account for logging into the virtual scene and controls the virtual object corresponding to the account to select different other virtual objects to interact with in the interaction mode.

[0074] The system acquires interaction pattern data from multiple candidate interaction patterns. These candidate interaction patterns are used for user recommendations, and the system also retrieves data from user accounts as recommendation target data. Recommendation target data includes: general basic information (e.g., frequently used virtual objects, virtual object usage rate, virtual object win rate, account match data in the past 14 days, ranking, game duration, etc.), account level, gender, age, etc. Interaction pattern data includes: storyline data, virtual object settings, interaction rules and logical feedback, interaction pattern type, etc.

[0075] In step 302, feature extraction is performed on the interaction pattern data and the recommended object data to obtain the interaction pattern features of each candidate interaction pattern and the object features of the recommended object.

[0076] For example, feature extraction is performed on interaction pattern data and recommendation object data respectively. Feature extraction is achieved through the extraction network in a hybrid expert network. The interaction pattern data and recommendation object data are taken as input, and the extraction network is a network layer that extracts useful information from the input data. This helps to identify and process key information in the interaction pattern data and recommendation object data. The extraction network obtains the interaction pattern features of each candidate interaction pattern and the object features of the recommendation object. A hybrid expert network is a sparse-gated deep learning model, mainly composed of a set of expert networks and a gating network. The basic concept of a hybrid expert network is to divide the input data into multiple regions according to the task type and assign one or more expert networks to the data in each region. Each expert network can focus on processing this part of the input data, thereby improving the overall performance of the model. In this embodiment, a set of expert networks and a gating network are used as the extraction network, and the extraction network can be stacked in multiple layers.

[0077] In some embodiments, step 302 can be implemented by the following method: for each candidate interaction pattern, the interaction pattern data is encoded to obtain a first encoding vector; the first encoding vector is subjected to different transformation processes to obtain multiple second encoding vectors; the multiple second encoding vectors are fused to obtain the interaction pattern feature of the candidate interaction pattern; the recommended object data is encoded to obtain a third encoding vector; the third encoding vector is subjected to different transformation processes to obtain multiple fourth encoding vectors; the multiple fourth encoding vectors are fused to obtain the object feature of the recommended object.

[0078] For example, for each candidate interaction pattern, the interaction pattern data is encoded using an embedding module composed of fully connected layers. The resulting embedded vector is used as the first encoding vector. This first encoding vector undergoes different transformations using different expert networks, yielding multiple second encoding vectors corresponding to different interaction patterns. These multiple second encoding vectors are then linearly fused using gating networks corresponding to each expert network. The encoded values ​​obtained from the multiple expert networks are weighted and fused into a single combined encoding value based on the weights of the gating networks, serving as the interaction pattern feature of the candidate interaction pattern. Similarly, the recommended object data is encoded using an embedding module composed of fully connected layers, yielding a third encoding vector. This third encoding vector undergoes different transformations using expert networks, yielding multiple fourth encoding vectors. These fourth encoding vectors are then linearly fused based on the weights of the gating networks to obtain the object features of the recommended object.

[0079] In some embodiments, see Figure 5A , Figure 5A This is a first structural diagram of the neural network model provided in this application embodiment. The input of the neural network model 501 includes interaction pattern data and recommendation object data. After receiving the input information data, the neural network model 501 calls the extraction network 502 to perform feature extraction on the input information. The extraction network 502 includes multiple expert networks and a shared expert network. Experts are trained sub-networks (neural networks or layers) that are specifically designed to handle specific data or tasks. In this application embodiment, the extraction network 502 includes multiple expert networks, each of which is used to process specific data, and a shared expert network for identifying and processing information used in multiple situations. Common information that can be used in multiple situations is compressed into the shared expert network, reducing parameter redundancy between other expert networks. After each expert network performs feature extraction processing, it corresponds to a gating structure. Each gating network is used to apply weights to different features. The gating network is a selector that routes the input data to the relevant experts, receives a single data element as input, and then outputs a weight. These weights represent the contribution of each expert network to processing the input data. The expert network in network 502 performs feature extraction processing on the input data in parallel to obtain interaction pattern features and object features with certain weight allocation. The interaction pattern features and object features obtained by the extraction network 502 are used as input into the multilayer extraction network 503 to provide a basis for target prediction in the gain model 504A.

[0080] In this embodiment, by assigning input data to different expert networks, each expert network can focus on processing a specific portion of the data, extracting features more accurately and reducing noise and errors in the feature extraction process. The gating network assigns weights to each expert network, clearly defining the contribution of each expert network to the final features. Accurate feature extraction helps to better understand user preferences and behavioral patterns, thereby providing more customized recommendations.

[0081] See also Figure 3A In step 303, natural target prediction for multiple target tasks is performed based on object features, and a first recommendation index is determined based on the first prediction probability of each target task obtained from the prediction.

[0082] For example, natural target prediction is a prediction process without applying candidate interaction patterns. Based on object features, natural target prediction is performed without any intervention. Natural target prediction corresponds to multiple types of target tasks. The prediction probabilities of multiple target tasks are linearly processed to obtain the first recommendation index obtained from natural prediction under multiple targets. Figure 5A In this embodiment, natural target prediction under multiple objectives is achieved through the natural target prediction module 5041A. In the game's virtual scene, the prediction process is implemented using a neural network model. The natural target prediction module within the neural network model predicts object features for multiple target tasks, obtaining a first prediction probability for each target task. These multiple target tasks include: short-term user retention rate, long-term user retention rate, and user's continued gaming probability.

[0083] In some embodiments, see Figure 3B , Figure 3B This is a second flowchart illustrating the processing method for the virtual scene interaction mode provided in the embodiments of this application. Figure 3A Step 303 in the process can be achieved through Figure 3B Steps 3031 to 3033 are implemented, and the details are explained below.

[0084] In step 3031, the object features are subjected to a nonlinear transformation to obtain the dimensionality-reduced object features.

[0085] For example, a non-linear transformation is performed on the object features. This non-linear transformation is achieved through an autoencoder. The autoencoder maps the original object features to a low-dimensional space through the encoder part, and then reconstructs the original object data through the decoder part to obtain the dimensionality-reduced object features.

[0086] In step 3032, for each target task, the pre-configured function corresponding to the target task is called based on the dimensionality-reduced object features to perform mapping processing, thereby obtaining the first predicted probability corresponding to the target task.

[0087] Here, the first prediction probability represents the probability that the recommended object will perform the target task.

[0088] For example, for each target task, a pre-configured function corresponding to the target task is called to perform mapping processing based on the dimensionality-reduced object features. The pre-configured function is an activation function (e.g., the sigmoid function), which maps the dimensionality-reduced features to values ​​between 0 and 1 to obtain the first predicted probability corresponding to the target task.

[0089] In step 3033, each first predicted probability is linearly transformed to obtain the first recommended index.

[0090] For example, the linear transformation process is based on the weighted fusion of multiple target tasks. The weights are pre-set according to the importance of each target task. The first prediction probability is weighted and fused based on the weights of multiple target tasks to obtain the first recommendation index.

[0091] The following explanation uses examples to illustrate the linear transformation of each first prediction probability. Assume that the user short-term retention rate is the primary prediction target, so the user short-term retention rate is assigned a weight of 1. The user long-term retention rate and the probability of the user continuing to play are secondary prediction targets, so the user long-term retention rate and the probability of the user continuing to play are assigned a weight of 0.02. Based on the weights, the first prediction probabilities obtained from each target prediction are weighted and fused to obtain the final first recommendation index under multi-target prediction.

[0092] In this embodiment, natural target prediction based on object features considers not only a single target but also other auxiliary targets, making the predicted first recommendation index more comprehensive. Dimensionality reduction of object features through nonlinear transformation helps extract key feature information, reduces data dimensionality, alleviates computational burden, and improves prediction accuracy.

[0093] See also Figure 3A In step 304, intervention target prediction for multiple target tasks is performed for each interaction mode feature and object feature to obtain the second recommendation index corresponding to each candidate interaction mode.

[0094] For example, intervention target prediction involves applying predictive processing under candidate interaction patterns. Different interaction patterns are applied to each interaction pattern feature and object feature to perform intervention target prediction processing. This intervention target prediction corresponds to multiple types of target tasks. The predicted probabilities of these multiple target task types are linearly processed to obtain a second recommendation index obtained from multi-target intervention target prediction. Figure 5A In the middle, the intervention target prediction module 5042A includes multiple intervention target prediction sub-modules (e.g.: Figure 5BThe intervention target prediction modules 505B and 506B in the actual scenario correspond to the number of candidate interaction modes. By applying interventions from candidate interaction mode 1 to candidate interaction mode n respectively, the intervention target prediction under multiple objectives is achieved.

[0095] In some embodiments, see Figure 3C , Figure 3C This is a schematic diagram of the third process of the virtual scene interaction mode processing method provided in the embodiments of this application. Figure 3A Step 304 in the process can be achieved through Figure 3C Steps 3041 to 3043 are implemented, and the details are explained below.

[0096] In step 3041, each interaction mode feature and object feature are concatenated to obtain the concatenated feature corresponding to each candidate interaction mode.

[0097] For example, each interaction mode feature and object feature are concatenated separately, and the first and last parts are concatenated by the embedding module to obtain the concatenated feature corresponding to each candidate interaction mode. For example, if the interaction mode feature is h1 and the object feature is d1, the concatenated feature representation is {h1, d1}.

[0098] In step 3042, for each candidate interaction mode, prediction processing for multiple target tasks is performed based on the splicing features to obtain a second prediction probability for each target task.

[0099] Here, the second predicted probability represents the probability that the recommended object will perform the target task under the intervention of the candidate interaction mode.

[0100] For example, in this embodiment of the application, in the virtual game scene, the prediction processing is implemented through a neural network model. For each candidate interaction mode, the intervention target prediction processing module in the neural network model performs intervention target prediction based on the splicing features, and obtains the second prediction probability corresponding to each target task. Each interaction mode is regarded as an intervention, and the target task of intervention target prediction is the same as that of natural target prediction, including: short-term user retention rate, long-term user retention rate, and the probability of user continuing to play the game.

[0101] In some embodiments, step 3042 can be implemented by the following method: for each candidate interaction mode, the splicing features are nonlinearly transformed to obtain dimensionality-reduced splicing features; for each target task, the pre-configured function corresponding to the target task is called based on the dimensionality-reduced splicing features to perform mapping processing to obtain the second prediction probability corresponding to the target task.

[0102] For example, a non-linear transformation is performed on the spliced ​​features of each candidate interaction pattern. This non-linear transformation is achieved through an autoencoder. The autoencoder maps the original spliced ​​features to a low-dimensional space through its encoder part, and then reconstructs the spliced ​​feature data through its decoder part, resulting in dimensionality-reduced spliced ​​features. For each target task, a pre-configured function corresponding to the target task is called based on the dimensionality-reduced spliced ​​features for mapping processing. The pre-configured function is an activation function that maps the dimensionality-reduced spliced ​​features to values ​​between 0 and 1, obtaining the second predicted probability corresponding to the target task.

[0103] In step 3043, each second predicted probability is linearly transformed to obtain the second recommendation index corresponding to the candidate interaction mode.

[0104] For example, the linear transformation process is based on the weighted fusion of multiple target tasks. The weights are pre-set according to the importance of each target task. Each second prediction probability is weighted and fused based on the weight of the target task to obtain the second recommendation index corresponding to the candidate interaction mode.

[0105] In this embodiment, each candidate interaction mode is treated as an intervention. Different types of interventions are applied during multi-objective prediction processing. Prediction is performed under various intervention conditions based on the fusion of interaction mode features and object features, achieving more comprehensive interaction data analysis under different interventions. By assigning different weights to different target tasks and comprehensively considering multiple prediction objectives, multi-faceted data support is provided for recommending virtual scene interaction modes.

[0106] See also Figure 3A In step 305, a target interaction mode is selected from multiple candidate interaction modes based on the first recommendation index and the second recommendation index for each candidate interaction mode.

[0107] Here, the target interaction mode is used to perform recommendation processing on the recommended objects.

[0108] For example, the difference between the first recommendation metric and the second recommendation metric for each candidate interaction mode is calculated. The difference corresponding to each candidate interaction mode is used as the gain value of each candidate interaction mode. The candidate interaction mode with the highest gain value is selected as the target interaction mode, and the target interaction mode is recommended to the user's terminal device.

[0109] In some embodiments, step 305 can be implemented by the following method: for each candidate interaction mode, determining a first difference between a first recommendation index and a second recommendation index of the candidate interaction mode; sorting each candidate interaction mode in descending order according to each first difference to obtain a descending sorted list; selecting at least one candidate interaction mode at the head of the descending sorted list as the target interaction mode.

[0110] For example, for each candidate interaction mode, a first difference is calculated between the first recommendation metric and the second recommendation metric of the candidate interaction mode. The first difference is used as the gain value of each candidate interaction mode. Each candidate interaction mode is sorted in descending order according to the first difference to obtain a descending sorted list. The descending sort is arranged according to the size of the gain value. At least one candidate interaction mode at the head of the descending sorted list is selected as the target interaction mode. The candidate interaction mode at the head of the list has a higher recommendation effect.

[0111] For example, the first candidate interaction mode in the descending sorted list can be selected as the target interaction mode for recommendation; alternatively, the top five candidate interaction modes in the descending sorted list can be selected as the target interaction mode for recommendation; and yet another example is selecting the first candidate interaction mode in the descending sorted list whose difference is greater than a preset difference threshold as the target interaction mode for recommendation. The preset difference threshold is set according to the actual application scenario.

[0112] In some embodiments, feature extraction, natural target prediction, and intervention target prediction are implemented using a neural network model. In this embodiment, the neural network model includes a hybrid expert network and a gain model; see also... Figure 3D , Figure 3D This is a schematic diagram of the fourth process of the virtual scene interaction mode processing method provided in the embodiments of this application. Figure 3A Before step 301, it was also through Figure 3D Steps 3051 to 3054 in the process implement the training of the neural network model, which will be explained in detail below.

[0113] In step 3051, the training sample set is obtained.

[0114] Here, the training sample set includes sample pattern data of sample interaction patterns, sample object data, and the first and second actual indicators for each sample interaction pattern.

[0115] For example, a group of users to be recommended is randomly selected from the users in the virtual scene. Multiple candidate interaction pattern samples are randomly distributed to these users. The obtained sample pattern data, sample object data, and the first and second actual indicators of each sample interaction pattern are used as a training sample set. The obtained training sample set is then filtered and cleaned. The cleaned data is labeled according to the prediction target, including: short-term user retention rate, long-term user retention rate, and user probability of continuing to play. Training sample pairs (sample, label) are obtained. In this embodiment, the target label may also include, but is not limited to, the number of days the user has played the game, the number of games played in the current quarter, and the number of login days.

[0116] In step 3052, the initialized neural network model is invoked based on the training sample set to perform prediction processing, and a first prediction index and a second prediction index for the interaction mode of each sample are obtained.

[0117] For example, a neural network model initialized based on a training sample set is invoked for prediction processing. This prediction processing is implemented through a gain model within the neural network, which is a structural component of the neural network model. Figure 5A In this model, the gain model 504A includes a natural goal prediction module 5041A and an intervention goal prediction module 5042A. The gain model is used to estimate the incremental impact of an individual intervention, predicting the causal effect of a certain intervention on an individual's state or behavior. Based on the training sample set, the natural goal prediction module 5041A in the gain model 504A is used to predict the natural goal, obtaining a first prediction index. Similarly, based on the training sample set, the intervention goal prediction module 5042A in the gain model is used to predict the intervention goal, obtaining a second prediction index.

[0118] In step 3053, a first loss function of the neural network model is determined based on the difference between the first predicted index and the first actual index, and a second loss function of the neural network model is determined based on the difference between the second predicted index and the second actual index.

[0119] For example, in the case of natural target prediction processing, a first loss function of the neural network model is determined based on the difference between a first predicted indicator and a first actual indicator; in the case of intervention target prediction processing, a second loss function of the neural network model is determined based on the difference between a second predicted indicator and a second actual indicator.

[0120] In some embodiments, step 3053 can be implemented by the following method: for each target task, determining the first predicted probability contained in the first predicted indicator and the first actual probability contained in the first actual indicator; when the type of the target task is a preset type, determining the cross-entropy loss of the first predicted probability and the first actual probability of the target task; when the type of the target task is not a preset type, determining the mean squared error loss of the first predicted probability and the first actual probability of the target task; performing linear operation on the cross-entropy loss and mean squared error loss corresponding to each target task to obtain the first loss function of the neural network model.

[0121] For example, for each target task, the first predicted probability included in the first predicted metric and the first actual probability included in the first actual metric are determined. When the type of the target task is short-term user retention rate, the cross-entropy loss of the first predicted probability and the first actual probability of the target task is determined. When the type of the target task is long-term user retention rate or user continue playing probability, the mean squared error loss of the first predicted probability and the first actual probability of the target task is determined. The cross-entropy loss and mean squared error loss corresponding to each target task are weighted and fused among multiple targets to obtain the first loss function of the neural network model. The calculation principle of the second loss function is the same as that of the first loss function in step 3053, and will not be repeated here.

[0122] In step 3054, the parameters of the neural network model are updated based on the first loss function and the second loss function to obtain the trained neural network model.

[0123] For example, the first loss function and the second loss function are weighted and fused according to the weights among multiple objectives to obtain the total loss function of the neural network model. The model parameters are then updated according to the total loss function to obtain the trained neural network model.

[0124] In some embodiments, after step 3054, the trained neural network model is obtained by the following method: acquiring a test sample set, wherein the test sample set includes test mode data of test interaction modes, test object data, and actual difference values ​​corresponding to each test interaction mode, the actual difference value being the difference between the first actual index and the second actual index of the test interaction mode; calling the initialized neural network model based on the test interaction mode to perform prediction processing, obtaining a third prediction index and a fourth prediction index for each sample interaction mode; determining a second difference value between the third prediction index and the fourth prediction index; determining an evaluation index for the neural network model based on the second difference value and the actual difference value; and updating the parameters of the neural network model based on the first loss function in response to the evaluation index being less than a preset threshold, thereby obtaining the trained neural network model.

[0125] For example, a test sample set is obtained, with data types identical to those in the training samples. Based on the test interaction mode, an initialized neural network model is invoked for prediction processing, yielding a third predictive index for the natural target prediction and a fourth predictive index for the intervention target prediction for each sample interaction mode. A second difference between the third and fourth predictive indices is determined, serving as the gain value for the sample interaction mode during the test. Based on this second difference and the actual difference, an evaluation index for the neural network model is determined. The trained recommendation model is evaluated using the Area Under the Uplift Curve (AUUC) metric to assess whether the model meets the target threshold. The AUUC is used to plot an uplift curve based on the model output. The target threshold is determined by the slope of the curve, with the gain value corresponding to the point where the curve slope is maximum or close to maximum being used as the target threshold. This target threshold is then used as a preset threshold to judge the evaluation index of the neural network model.

[0126] For example, if the evaluation metric is less than a preset threshold, it indicates that the trained model has not met the evaluation criteria and the neural network model parameters need to be updated based on the first loss function to obtain the trained neural network model.

[0127] In some embodiments, the evaluation index is determined by the following method: constructing a curve based on the second difference and the actual difference for each sample interaction pattern, with the second difference as the x-axis and the actual difference as the y-axis; and determining the evaluation index based on the area under the curve.

[0128] For example, using the second difference as the x-axis and the actual difference as the y-axis, a curve is constructed based on the second difference and the actual difference of each sample interaction mode. The area under the curve is calculated, and the preset target threshold is determined by the slope of the constructed curve. The second difference corresponding to the point in the curve with the largest or closest to the largest slope is used as the evaluation index.

[0129] In some embodiments, Figure 3A After step 305, the following processing is also performed: in response to the virtual object corresponding to the recommended object being in a virtual scene, the virtual object corresponding to the recommended object is controlled to perform interaction processing based on the target interaction mode; the interaction data of the virtual object corresponding to the recommended object in the virtual scene is recorded, wherein the interaction data is used to train the neural network model.

[0130] For example, after determining the target interaction mode based on the first recommendation index and the second recommendation index, in response to the virtual object corresponding to the user being in the virtual scene, the virtual object corresponding to the user is controlled to perform interaction processing based on the target interaction mode. The interaction of the virtual object is executed according to the interaction principles and logic in the target interaction mode, and the interaction data in the virtual scene is recorded. The interaction data is used to train the neural network model to improve the accuracy of the neural network model in determining the target interaction mode.

[0131] In this embodiment, by acquiring interaction pattern data and recommendation object data from multiple candidate interaction patterns, a variety of data foundations are provided for subsequent prediction and training of the neural network model. Input data is distributed to different expert networks, each of which can focus on processing a specific portion of the data, extracting features more accurately and reducing noise and errors in the feature extraction process. Natural target prediction and intervention target prediction are performed based on the extracted object features and interaction pattern features, treating each interaction pattern as an intervention, thus achieving multiple interventions in the prediction process. Each prediction process is performed with three objectives: short-term user retention rate, long-term user retention rate, and user probability of continuing to play. By simultaneously considering the weighted fusion of the prediction results of the three objectives, a first recommendation index and a second recommendation index are determined, achieving multi-objective, multi-intervention prediction processing. Using short-term user retention rate as the primary optimization objective and long-term user retention rate and user probability of continuing to play as auxiliary optimization objectives, the loss function of the neural network model is determined. The model is trained based on the loss function, and the area under the curve (AUC) is used as the evaluation criterion for the trained model, which improves the model's representation ability and prediction accuracy. Calculate the first difference between the first recommendation metric and the second recommendation metric, and select at least one interaction pattern from the top of the list arranged in descending order according to the first difference as the target interaction pattern for recommendation, thereby achieving accurate personalized recommendations and improving user retention rate.

[0132] The following will describe an exemplary application of the embodiments of this application in a real-world application scenario.

[0133] In the current gaming industry, to enhance user experience and increase user engagement, appropriate interactive content is typically intelligently selected and pushed to players within a virtual environment based on their behavior and preferences. However, current technologies often rely on manually defined rules to determine the timing and target audience for interactive content delivery. These rules are highly dependent on human experience and struggle to consider the impact of different interactive modes on different groups within the entire game market and among all player demographics. Recommendation models, which model the relevance between interactive modes and players, rank different interactive modes by relevance score and recommend the highest-scoring mode, fail to consider the differences between players sensitive to interactive modes and those who naturally convert. Furthermore, prediction methods based on single interventions or single objectives have limited optimization goals and cannot consider the multi-faceted impact of different interactive modes on players, resulting in poor personalized delivery.

[0134] This application embodiment constructs a multi-objective, multi-intervention gain model. Multiple candidate interaction modes and interaction mode datasets, along with recommended object data, are input into the gain model for feature extraction. Interaction mode features and object features are extracted from the network output. Natural target prediction and intervention target prediction are performed based on these interaction mode features and object features to obtain natural response values ​​and intervention response values. The difference between the natural response value and the intervention response value determines the gain value under multi-intervention conditions. Based on the gain value of each candidate interaction mode, a target interaction mode is determined for player recommendations.

[0135] The following explanation is in conjunction with the accompanying drawings. Figure 4 , Figure 4 This is a fifth flowchart illustrating the processing method for the virtual scene interaction mode provided in this application embodiment. The executing entity can be a terminal device, a server, or a combination of both. This application embodiment uses... Figure 1 Taking the server as the execution subject as an example, it will be combined with Figure 4 The steps shown are explained in detail.

[0136] In step 401, interaction pattern data and recommendation object data of multiple candidate interaction patterns are obtained.

[0137] For example, users experience different candidate interaction modes in a virtual scene. These interaction modes are designed for virtual objects within the scene, guiding their behavior and dialogue. They include storylines, virtual object settings, and the rules and logic governing how virtual objects react based on the virtual scene and behavior. Interaction modes allow virtual objects to dynamically communicate and interact, constructing a user experience that combines plot and virtual object choices based on the preset storyline and real-time context. Recommendation targets are users experiencing the virtual scene. Each user has an account, and interaction mode data from multiple different candidate interaction modes is obtained for user recommendations. User account-related data is also obtained as recommendation target data. Recommendation target data includes: general basic information (e.g., frequently used game virtual objects, game virtual object usage rate, game virtual object win rate, player's match data in the last 14 days, ranking, game duration, etc.), account level, gender, age, and interaction data of the virtual objects corresponding to the account in the virtual scene. Interaction mode data includes: storyline data, virtual object settings, interaction rules and logical feedback, and interaction mode types.

[0138] In step 402, feature extraction is performed on the interaction pattern data and the recommended object data to obtain the interaction pattern features of each candidate interaction pattern and the object features of the recommended object.

[0139] For example, feature extraction is performed on both the interaction pattern data and the recommendation object data. Feature extraction can be achieved through the extraction network in a hybrid expert network. The interaction pattern data and recommendation object data are taken as input. A hybrid expert network is a sparse-gated deep learning model, mainly composed of a set of expert networks and a gating network. The basic idea of ​​a hybrid expert network is to divide the input data into multiple regions according to the task type and assign one or more expert networks to the data in each region. Each expert network can focus on processing this part of the input data, thereby improving the overall performance of the model. The extraction network is a network layer used to extract useful information from the input data, helping to identify and process key information in player features. Multiple layers can be stacked. Through the extraction network, the interaction pattern features of each candidate interaction pattern and the object features of the recommendation object are obtained.

[0140] In some embodiments, see Figure 5B , Figure 5BThis is a second structural diagram of the neural network model provided in this application embodiment. The input 5011 of the recommendation model includes interaction mode data and recommendation object data. After receiving the input information data, the recommendation model calls the extraction network 502 to perform feature extraction on the input information. The extraction network 502 includes multiple expert networks and an expert shared network. Experts are pre-trained sub-networks (neural networks or layers) that are specifically designed to handle specific data or tasks. In this application embodiment, the extraction network 502 includes three expert networks: expert network 5021, expert network 5022, and expert network 5023. Each expert network is used to process specific data. For example, expert network 5021 is used to process the general basic information of the recommendation object, expert network 5022 is used to process the remaining information of the recommendation object, and expert network 5023 is used to process the interaction mode data. The shared expert network 5024 is used to identify and process information used in multiple situations, compressing common information that can be used in multiple situations into the shared expert network to reduce parameter redundancy between other expert networks. After each expert network performs feature extraction, it corresponds to a gating structure. Expert network 5021 corresponds to gating 50211, expert network 5022 corresponds to gating 50221, expert network 5023 corresponds to gating 50231, and the shared expert network 5024 corresponds to gating 50241. Each gating network is used to apply weights to different features. The gating network is a selector that routes the input data to the relevant experts. It receives a single data element as input and then outputs a weight, which represents the contribution of each expert network to processing the input data.

[0141] Feature extraction is performed on the input data in parallel by expert networks in extraction network 502. Each expert network consists of multiple fully connected layers and attention modules. Each expert network encodes the input into another encoded value. The encoded values ​​obtained by the four expert networks are weighted and fused into a combined encoded value. The output structures of the four expert networks are identical. A weight assignment is learned through a gating network, where the weights represent the relevance of the input data to each expert, resulting in interaction pattern features and object features with specific weight assignments. The interaction pattern features and object features obtained by extraction network 502 are input into multi-level extraction network 503, which performs feature fusion processing to obtain a higher-level fused feature representation. Multi-level extraction network 503 also includes three expert networks: expert network 5031 (corresponding to gating 50311), expert network 5032 (corresponding to gating 50321), expert network 5033 (corresponding to gating 50331), and shared expert network 5034 (corresponding to gating 50341), which have the same function as the expert networks and shared expert networks in extraction network 502.

[0142] See also Figure 4 In step 403, natural target prediction is performed based on object features to obtain natural response values.

[0143] For example, based on object features obtained through expert network processing, natural target prediction is performed without any intervention to obtain natural response values. Natural target prediction includes: short-term user retention rate as the primary prediction target, and long-term user retention rate and the probability of continued user play as auxiliary prediction targets. Short-term user retention rate is assigned a weight of 1, and long-term user retention rate and the probability of continued user play are assigned weights of 0.02. Linear operations are performed on the response values ​​obtained from each target prediction based on these weights to obtain the final natural response value predicted under multi-target conditions. Figure 5B In this system, the prediction of natural response values ​​under multiple objectives is achieved through the natural objective prediction module 504B.

[0144] In step 404, intervention target prediction is performed for each interaction mode feature and object feature to obtain the intervention response value corresponding to each candidate interaction mode.

[0145] For example, intervention target prediction is performed for each interaction pattern feature and object feature. Each interaction pattern is considered as an intervention, and the intervention response value corresponding to each candidate interaction pattern is obtained. The intervention target prediction has the same objective as the natural target prediction, with short-term user retention rate as the primary prediction objective and long-term user retention rate and the probability of user continuing to play as secondary prediction objectives. The short-term user retention rate is assigned a weight of 1, and the long-term user retention rate and the probability of user continuing to play are assigned weights of 0.02. Based on the weights, the response values ​​obtained from each target prediction are processed linearly to obtain the multi-objective predicted intervention response value for each interaction pattern. Figure 5B In the example of intervention target prediction module 505B and intervention target prediction module 506B, the number of intervention target prediction modules in the actual scenario corresponds to the number of candidate interaction modes. By applying interventions from candidate interaction mode 1 to candidate interaction mode n respectively, intervention target prediction under multiple objectives can be achieved.

[0146] See also Figure 4 In step 405, the difference between the natural response value and the intervention response value is used as the gain value of each candidate interaction mode, and the candidate interaction mode with the highest gain value is selected as the target interaction mode.

[0147] For example, the difference between the natural response value and the intervention response value is used as the gain value for each candidate interaction mode. The response value is a number between 0 and 1. For instance, after a player finishes the current game, the probability of continuing the game without intervention is 0.5 (the natural response value). After intervention, the probability of continuing the game is 0.6 (the intervention response value). In this case, the gain is 0.6 - 0.5 = 0.1, which is the actual predicted gain value. The candidate interaction mode with the highest gain value is selected as the target interaction mode and recommended to users to increase user retention.

[0148] In some embodiments, see Figure 6 , Figure 6 This is a schematic diagram of the sixth step in the processing method of virtual scene interaction mode provided in this application embodiment; before realizing feature extraction and target prediction of interaction mode data and object data through the recommendation model, it also involves... Figure 6 Steps 601 to 606 in the process are used to train the recommendation model, which will be explained in detail below.

[0149] In step 601, training data is acquired.

[0150] For example, a group of users to be recommended is randomly selected from the users in the virtual scene, and multiple candidate interaction modes are randomly distributed to the users to be recommended. The interaction data of users in different interaction modes, as well as the interaction mode data and user account data, are used as training data.

[0151] In step 602, the training data is cleaned.

[0152] For example, the obtained training data is filtered and cleaned, and the cleaned data is labeled according to the prediction target. The target labels include: user short-term retention rate, user long-term retention rate, and user probability of continuing to play the game, to obtain training sample pairs of (sample, label). In this embodiment of the application, the target labels may also include, but are not limited to, user game days, user current quarterly number of games, login days, etc.

[0153] In step 603, the training dataset and the test dataset are divided.

[0154] For example, in the (sample, label) dataset after adding target labels, 10% is randomly divided as the test dataset and the remaining 90% is used as the training dataset.

[0155] In step 604, a recommendation model is constructed by inputting the training dataset and the test dataset into the model for optimization.

[0156] For example, a recommendation model is constructed, including an extraction network composed of multiple expert networks and a shared expert network for user feature extraction processing, and a gain network. The gain network includes a natural prediction target module and an intervention target prediction module, which are the same in structure and principle as the recommendation model shown in Figure 5. The training dataset and the test dataset are input into the model to calculate the gain value. In this embodiment, in addition to the interactive mode as intervention, it may also include, but is not limited to, game type (e.g., leading, lagging, absolute advantage, comeback, even game type, and 5v5, 1v9 game mode) and game style (e.g., aggressive, conservative, stable style). The implementation principle is the same as steps 401 to 404, and will not be repeated here.

[0157] For example, after predicting the natural response value and intervention response value for each target, a loss function is determined for training the recommendation model. The first sub-loss function is the cross-entropy loss between the predicted target label and the actual target label, targeting short-term user retention. The second sub-loss function is the mean squared error loss between the predicted target label and the actual label, targeting long-term user retention. The third sub-loss function is the mean squared error loss between the predicted target label and the actual label, targeting the probability of the user continuing to play. Based on the weights of the short-term user retention rate, long-term user retention rate, and probability of the user continuing to play, the first, second, and third sub-loss functions are linearly added to obtain the total loss function. The recommendation model is then trained and optimized based on the total loss function. In this embodiment, the optimization algorithm for the recommendation model can also be a neural network optimization algorithm such as stochastic gradient descent, adaptive moment estimation, or adaptive adjustment. No limitation is placed on the optimization algorithm for the model here.

[0158] In step 605, it is determined whether the evaluation result of the trained model using test data meets the target threshold.

[0159] For example, the trained recommendation model is evaluated using the area under the curve (AUC) as an evaluation metric. The predicted gain value is evaluated to determine whether it meets the target threshold. The AUC is used to plot the boost curve based on the model output. The target threshold is determined by the slope of the curve. The boost score corresponding to the point where the slope is maximum or close to maximum is used as the target threshold. The X-axis of the curve represents the potential gain of an individual after intervention relative to the individual without intervention, as predicted by the model. The Y-axis represents the gain value, which is the average performance difference between the individuals who received intervention and those who did not, as actually observed.

[0160] If the judgment result in step 605 is yes, then step 606 is executed to store the trained model in the server.

[0161] For example, a recommendation model that meets the evaluation results is deployed to a server. The server can call the trained model to perform recommendation processing. At this time, the trained recommendation model can accurately recommend the user's target interaction pattern and perform recommendation processing to the object to be recommended through the trained recommendation pattern.

[0162] If the result of the judgment in step 605 is negative, continue to execute step 604.

[0163] The virtual scene interaction mode processing method provided in this application embodiment has the following beneficial effects:

[0164] By acquiring interaction pattern data and recommendation object data from multiple candidate interaction modes, a diverse data foundation is provided for subsequent model prediction and training. Feature extraction is performed on the interaction pattern data and recommendation object data using multiple expert networks to obtain corresponding features, enabling multi-dimensional representation of the features used for prediction and facilitating centralized processing of subsequent multi-objective and multi-intervention predictions. Based on the extracted object features and interaction pattern features, natural target prediction and intervention target prediction are performed separately, treating each interaction mode as an intervention to achieve multiple interventions in the prediction process. For each target prediction, predictions are made based on three objectives: short-term user retention rate, long-term user retention rate, and user probability of continuing to play. The final natural response value and intervention response value are determined by weighted fusion of the prediction results for these three objectives, achieving multi-objective and multi-intervention prediction processing. Simultaneously, short-term user retention rate is the primary optimization objective, while long-term user retention rate and user probability of continuing to play are secondary optimization objectives. Training and evaluating the recommendation model improves its representational ability and prediction accuracy. The difference between the natural response value and the intervention response value is used as the gain value. The interaction mode with the highest gain value is selected as the target interaction mode for recommendation. In the process of selecting the target interaction mode, the influence of multiple indicators of the recommended user in different intervention scenarios under various objectives is comprehensively considered to achieve accurate personalized recommendations and improve user retention.

[0165] The following description continues to illustrate the exemplary structure of the virtual scene interaction mode processing device 455 provided in the embodiments of this application as a software module. In some embodiments, such as Figure 2As shown, the software modules in the virtual scene interaction mode processing device 455 stored in the memory 450 may include: a feature extraction module 4551, used to acquire interaction mode data and recommendation object data of multiple candidate interaction modes, wherein the candidate interaction modes are interaction modes between virtual objects in the virtual scene; performing feature extraction on the interaction mode data and recommendation object data respectively to obtain the interaction mode features of each candidate interaction mode and the object features of the recommendation object; a recommendation module 4552, performing natural target prediction for multiple target tasks based on the object features, determining a first recommendation index based on the first prediction probability of each target task obtained from the prediction; performing intervention target prediction for multiple target tasks for each interaction mode feature and object feature respectively to obtain a second recommendation index corresponding to each candidate interaction mode; and selecting a target interaction mode from multiple candidate interaction modes based on the first recommendation index and the second recommendation index of each candidate interaction mode, wherein the target interaction mode is used for recommendation processing of the recommendation object.

[0166] In some embodiments, the feature extraction module 4551 is further configured to: encode the interaction pattern data for each candidate interaction pattern to obtain a first encoding vector; perform different transformations on the first encoding vector to obtain multiple second encoding vectors; fuse the multiple second encoding vectors to obtain the interaction pattern feature of the candidate interaction pattern; encode the recommended object data to obtain a third encoding vector; perform different transformations on the third encoding vector to obtain multiple fourth encoding vectors; and fuse the multiple fourth encoding vectors to obtain the object feature of the recommended object.

[0167] In some embodiments, the recommendation module 4552 is further configured to perform a nonlinear transformation on the object features to obtain dimensionality-reduced object features; for each target task, to call a pre-configured function corresponding to the target task based on the dimensionality-reduced object features to perform mapping processing to obtain a first prediction probability corresponding to the target task, wherein the first prediction probability represents the probability that the recommended object will perform the target task; and to perform a linear transformation on each first prediction probability to obtain a first recommendation index.

[0168] In some embodiments, the recommendation module 4552 is further configured to concatenate each interaction mode feature with the object feature to obtain a concatenated feature corresponding to each candidate interaction mode; for each candidate interaction mode, perform prediction processing on multiple target tasks based on the concatenated features to obtain a second prediction probability for each target task, wherein the second prediction probability represents the probability that the recommended object will perform the target task under the intervention of the candidate interaction mode; and perform linear transformation processing on each second prediction probability to obtain a second recommendation index corresponding to the candidate interaction mode.

[0169] In some embodiments, the recommendation module 4552 is further configured to perform a nonlinear transformation on the splicing features for each candidate interaction mode to obtain dimensionality-reduced splicing features; and for each target task, to call the pre-configured function corresponding to the target task based on the dimensionality-reduced splicing features to perform mapping processing to obtain the second predicted probability corresponding to the target task.

[0170] In some embodiments, the recommendation module 4552 is further configured to, for each candidate interaction mode, determine a first difference between a first recommendation index and a second recommendation index of the candidate interaction mode; sort each candidate interaction mode in descending order according to each first difference to obtain a descending sorted list; and select at least one candidate interaction mode at the head of the descending sorted list as the target interaction mode.

[0171] In some embodiments, after selecting a target interaction mode from multiple candidate interaction modes based on a first recommendation index and a second recommendation index for each candidate interaction mode, the recommendation module 4552 is further configured to, in response to the virtual object corresponding to the recommendation object being in a virtual scene, control the virtual object corresponding to the recommendation object to perform interaction processing based on the target interaction mode; and record the interaction data of the virtual object corresponding to the recommendation object in the virtual scene, wherein the interaction data is used to train a neural network model.

[0172] In some embodiments, feature extraction, natural target prediction, and intervention target prediction are implemented through a neural network model. Before acquiring interaction pattern data and recommendation object data for multiple candidate interaction patterns, the model training module 4553 is further configured to acquire a training sample set, wherein the training sample set includes sample pattern data, sample object data, and a first actual indicator and a second actual indicator for each sample interaction pattern. Based on the training sample set, the initialized neural network model is invoked for prediction processing to obtain a first prediction indicator and a second prediction indicator for each sample interaction pattern. Based on the difference between the first prediction indicator and the first actual indicator, a first loss function of the neural network model is determined, and based on the difference between the second prediction indicator and the second actual indicator, a second loss function of the neural network model is determined. Based on the first loss function and the second loss function, the parameters of the neural network model are updated to obtain the trained neural network model.

[0173] In some embodiments, the model training module 4553 is further configured to, for each target task, determine the first predicted probability contained in the first predicted indicator and the first actual probability contained in the first actual indicator; when the type of the target task is a preset type, determine the cross-entropy loss of the first predicted probability and the first actual probability of the target task; when the type of the target task is not a preset type, determine the mean squared error loss of the first predicted probability and the first actual probability of the target task; and perform linear operations on the cross-entropy loss and the mean squared error loss corresponding to each target task to obtain the first loss function of the neural network model.

[0174] In some embodiments, after updating the neural network model parameters based on the first loss function and the second loss function to obtain the trained neural network model, the model training module 4553 is further configured to obtain a test sample set, wherein the test sample set includes test mode data of the test interaction mode, test object data, and actual difference corresponding to each test interaction mode, the actual difference being the difference between the first actual index and the second actual index of the test interaction mode; calling the initialized neural network model based on the test interaction mode to perform prediction processing, obtaining a third prediction index and a fourth prediction index for each sample interaction mode; determining a second difference between the third prediction index and the fourth prediction index; determining an evaluation index of the neural network model based on the second difference and the actual difference; and updating the neural network model parameters based on the first loss function in response to the evaluation index being less than a preset threshold, thereby obtaining the trained neural network model.

[0175] In some embodiments, the model training module 4553 is further configured to construct a curve based on the second difference and the actual difference of each sample interaction mode, using the second difference as the abscissa and the actual difference as the ordinate; and to determine the evaluation index based on the area under the curve.

[0176] This application provides a computer program product, which includes a computer program or computer-executable instructions stored in a computer-readable storage medium. The processor of an electronic device reads the computer-executable instructions from the computer-readable storage medium and executes the computer-executable instructions, causing the electronic device to perform the virtual scene interaction mode processing method described above in this application embodiment.

[0177] This application provides a computer-readable storage medium storing computer-executable instructions or a computer program. When the computer-executable instructions or the computer program are executed by a processor, the processor will execute the processing method of the virtual scene interaction mode provided in this application embodiment. For example, ... Figure 3A The method for handling virtual scene interaction modes is shown.

[0178] In some embodiments, the computer-readable storage medium may be a memory such as RAM, ROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a variety of devices including one or any combination of the above-mentioned memories.

[0179] In some embodiments, computer-executable instructions may take the form of programs, software, software modules, scripts, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as stand-alone programs or as modules, components, subroutines, or other units suitable for use in a computing environment.

[0180] As an example, computer-executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple co-located files (e.g., files that store one or more modules, subroutines, or code sections).

[0181] As an example, computer-executable instructions can be deployed to execute on a single electronic device, or on multiple electronic devices located at one location, or on multiple electronic devices distributed across multiple locations and interconnected via a communication network.

[0182] In summary, this application's embodiments extract features from interaction pattern data and recommended object data of multiple candidate interaction patterns, thereby more accurately identifying and understanding user preferences and needs. Based on object features, natural target prediction is performed to obtain a first recommendation index, and intervention target prediction is performed for each interaction pattern feature and object feature to obtain a second recommendation index corresponding to each candidate interaction pattern, evaluating the impact of different interaction patterns on the recommended object. Based on the first recommendation index and the second recommendation index for each candidate interaction pattern, the target interaction pattern with the greatest appeal to the recommended object is selected for recommendation, improving user engagement and retention rates, and enhancing the accuracy of interaction pattern recommendations.

[0183] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. A method for processing virtual scene interaction modes, characterized in that, The method includes: Acquire interaction pattern data and recommendation object data for multiple candidate interaction patterns, wherein the candidate interaction patterns are interaction patterns between virtual objects in the virtual scene; Feature extraction is performed on the interaction pattern data and the recommendation object data respectively to obtain the interaction pattern features of each candidate interaction pattern and the object features of the recommendation object; Based on the object features, natural target prediction is performed for various target tasks, and a first recommendation index is determined based on the first prediction probability of each target task obtained from the prediction. For each interaction mode feature and object feature, the intervention target prediction of the multiple target tasks is performed to obtain the second recommendation index corresponding to each candidate interaction mode; Based on the first recommendation index and the second recommendation index for each of the candidate interaction modes, a target interaction mode is selected from the plurality of candidate interaction modes, wherein the target interaction mode is used to perform recommendation processing on the recommended object.

2. The method according to claim 1, characterized in that, The step of extracting features from the interaction pattern data and the recommended object data to obtain the interaction pattern features and the object features of each candidate interaction pattern includes: For each candidate interaction pattern, the interaction pattern data is encoded to obtain a first encoding vector; By performing different transformation processes on the first encoded vector, multiple second encoded vectors are obtained; The multiple second encoding vectors are fused to obtain the interaction pattern features of the candidate interaction pattern; The recommended object data is encoded to obtain a third encoding vector; By performing different transformation processes on the third encoding vector, multiple fourth encoding vectors can be obtained; The multiple fourth encoding vectors are fused to obtain the object features of the recommended object.

3. The method according to claim 1, characterized in that, The step of predicting natural targets for multiple target tasks based on the object features, and determining a first recommendation index based on a first prediction probability for each target task obtained from the prediction, includes: The object features are subjected to a nonlinear transformation to obtain dimensionality-reduced object features; For each target task, a pre-configured function corresponding to the target task is called based on the dimensionality-reduced object features to perform mapping processing, thereby obtaining a first predicted probability corresponding to the target task, wherein the first predicted probability represents the probability that the recommended object performs the target task; A linear transformation is performed on each of the first predicted probabilities to obtain the first recommendation index.

4. The method according to claim 1, characterized in that, The step involves predicting the intervention targets for each of the interaction pattern features and object features for the various target tasks, resulting in a second recommendation index corresponding to each candidate interaction pattern, including: Each interaction mode feature is concatenated with the object feature to obtain the concatenated feature corresponding to each candidate interaction mode; For each of the candidate interaction patterns, prediction processing of the multiple target tasks is performed based on the splicing features to obtain a second prediction probability for each of the target tasks, wherein the second prediction probability represents the probability that the recommended object performs the target task under the intervention of the candidate interaction pattern; A linear transformation is performed on each of the second predicted probabilities to obtain the second recommendation index corresponding to the candidate interaction mode.

5. The method according to claim 4, characterized in that, For each candidate interaction pattern, based on the splicing features, prediction processing for multiple target tasks is performed to obtain a second prediction probability for each target task, including: For each of the candidate interaction patterns, the splicing features are nonlinearly transformed to obtain dimensionality-reduced splicing features; For each target task, a pre-configured function corresponding to the target task is called based on the dimensionality-reduced concatenation features to perform mapping processing, thereby obtaining the second predicted probability corresponding to the target task.

6. The method according to claim 1, characterized in that, The step of selecting a target interaction mode from the plurality of candidate interaction modes based on the first recommendation index and the second recommendation index for each of the candidate interaction modes includes: For each candidate interaction pattern, a first difference is determined between the first recommendation index and the second recommendation index of the candidate interaction pattern; Each candidate interaction pattern is sorted in descending order based on each of the first differences to obtain a descending sorted list; At least one of the candidate interaction modes at the head of the descending sorted list is selected as the target interaction mode.

7. The method according to any one of claims 1 to 6, characterized in that, The feature extraction, the natural target prediction, and the intervention target prediction are implemented through a neural network model. Before acquiring interaction pattern data and recommendation object data for multiple candidate interaction patterns, the method further includes: Obtain a training sample set, wherein the training sample set includes sample pattern data of sample interaction modes, sample object data, and the first actual indicator and the second actual indicator of each sample interaction mode. Based on the training sample set, the initialized neural network model is invoked to perform prediction processing, and a first prediction index and a second prediction index for each sample interaction mode are obtained. Based on the difference between the first predicted index and the first actual index, a first loss function of the neural network model is determined, and based on the difference between the second predicted index and the second actual index, a second loss function of the neural network model is determined. The parameters of the neural network model are updated based on the first loss function and the second loss function to obtain the trained neural network model.

8. The method according to claim 7, characterized in that, Determining the first loss function of the neural network model based on the difference between the first predicted indicator and the first actual indicator includes: For each target task, determine the first predicted probability contained in the first predicted indicator and the first actual probability contained in the first actual indicator; When the type of the target task is a preset type, determine the first predicted probability of the target task and the cross-entropy loss of the first actual probability; When the type of the target task is not a preset type, determine the mean squared error loss of the first predicted probability and the first actual probability of the target task. A linear operation is performed on the cross-entropy loss and mean squared error loss corresponding to each target task to obtain the first loss function of the neural network model.

9. The method according to claim 7, characterized in that, After updating the neural network model parameters based on the first loss function and the second loss function to obtain the trained neural network model, the method further includes: Obtain a test sample set, wherein the test sample set includes test mode data of test interaction modes, test object data, and actual difference value corresponding to each test interaction mode, wherein the actual difference value is the difference between the first actual indicator and the second actual indicator of the test interaction mode; Based on the test interaction mode, the initialized neural network model is invoked to perform prediction processing, and a third prediction index and a fourth prediction index for each sample interaction mode are obtained. Determine the second difference between the third predictive indicator and the fourth predictive indicator; The evaluation index of the neural network model is determined based on the second difference and the actual difference; In response to the evaluation index being less than a preset threshold, the parameters of the neural network model are updated based on the first loss function to obtain the trained neural network model.

10. The method according to claim 9, characterized in that, The step of determining the evaluation index of the neural network model based on the second difference and the actual difference includes: A curve is constructed based on the second difference and the actual difference for each sample interaction mode, with the second difference as the x-axis and the actual difference as the y-axis. The evaluation index is determined based on the area under the curve.

11. The method according to claim 7, characterized in that, After selecting the target interaction mode from the plurality of candidate interaction modes based on the first recommendation index and the second recommendation index for each of the candidate interaction modes, the method further includes: In response to the virtual object corresponding to the recommended object being in a virtual scene, the virtual object corresponding to the recommended object is controlled to perform interactive processing based on the target interaction mode; The interaction data of the virtual object corresponding to the recommended object in the virtual scene is recorded, wherein the interaction data is used to train the neural network model.

12. A processing device for a virtual scene interaction mode, characterized in that, The device includes: The feature extraction module is used to acquire interaction pattern data and recommendation object data of multiple candidate interaction patterns, wherein the candidate interaction patterns are interaction patterns between virtual objects in the virtual scene; and to extract features from the interaction pattern data and the recommendation object data respectively to obtain the interaction pattern features of each candidate interaction pattern and the object features of the recommendation object. The recommendation module is used to predict natural targets for multiple target tasks based on the object features, and determine a first recommendation index based on a first prediction probability for each target task obtained from the prediction; to predict intervention targets for the multiple target tasks for each interaction pattern feature and the object features respectively, and obtain a second recommendation index corresponding to each candidate interaction pattern; and to select a target interaction pattern from the multiple candidate interaction patterns based on the first recommendation index and the second recommendation index for each candidate interaction pattern, wherein the target interaction pattern is used to perform recommendation processing on the recommended object.

13. An electronic device, characterized in that, The electronic device includes: Memory is used to store executable instructions or computer programs. A processor, when executing computer-executable instructions or computer programs stored in the memory, implements the processing method of the virtual scene interaction mode as described in any one of claims 1 to 11.

14. A computer-readable storage medium storing computer-executable instructions or a computer program, characterized in that, When the computer-executable instructions or computer program are executed by a processor, they implement the processing method of the virtual scene interaction mode as described in any one of claims 1 to 11.

15. A computer program product comprising computer-executable instructions or a computer program, characterized in that, When the computer-executable instructions or computer program are executed by a processor, the processing method of the virtual scene interaction mode as described in any one of claims 1 to 11 is implemented.

Images