A game silent user activation method and system based on artificial intelligence

Abstract

This invention discloses an artificial intelligence-based method and system for activating dormant game users. The method specifically includes: real-time acquisition of multi-dimensional behavioral flow data of users; constructing a spatiotemporal graph structure using this data; employing a graph neural network to perform node embedding learning on the spatiotemporal graph structure to capture the spatiotemporal correlation features of different dormant modes; and dynamically identifying the current dormant state of the target user using a temporal convolutional module with an adjustable receptive field; inputting the identified current dormant state into a two-layer reinforcement learning engine; accurately classifying the dormant mode using a deep Q-network at the bottom layer; and dynamically selecting the optimal intervention strategy from a strategy library using a proximal policy optimization network at the top layer, based on the classification results. This invention effectively improves the success rate of activating dormant users, enhances the user's gaming experience and retention rate, and has good adaptability to different game types. It also allows for in-depth analysis of the strategy decision-making process to optimize intervention strategies.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for activating silent game users based on artificial intelligence. Background Technology

[0002] In game operation, user churn is one of the key issues affecting the long-term development and revenue of a game. Many users often fall into a "silent state" at crucial points in the game, such as making payment decisions, failing levels, or participating in events, due to confusion, dissatisfaction, or difficulties. This silent state manifests as users ceasing operations but not exiting the game, placing them on the verge of potential churn. However, traditional customer service systems mainly rely on users proactively seeking help. For these silent users, the lack of effective identification mechanisms makes it difficult to promptly identify and address their issues, leading to a continuously rising churn rate and significant economic losses for game operators.

[0003] First, current technologies for identifying dormant users are limited and outdated. Customer service response is typically triggered only by explicit user actions, such as submitting a support ticket. However, many users, when in a dormant state, do not proactively express their needs through explicit actions like submitting support tickets; instead, they silently cease all activity. This means that current technologies cannot detect these latent dormant users in real time, preventing operators from intervening promptly in the early stages of user churn risk and missing the optimal opportunity for retention.

[0004] Secondly, current interaction methods for inactive users mainly rely on pushing generic messages, such as simply sending messages like "Need help?". This generic approach lacks specificity and fails to fully consider the user's current game context, behavioral situation, and personalized needs. Faced with such monotonous messages, users are easily inclined to ignore them, making it difficult to effectively attract their attention and, more importantly, to motivate them to re-engage with the game.

[0005] Finally, existing AI response strategies lack dynamic adjustment capabilities. When facing different types of inactive users, they cannot flexibly select appropriate language and intervention strategies based on key factors such as the user's willingness to pay, game history, and current level progress. For example, different language and guidance strategies should be used for users with a high willingness to pay but who are inactive due to level difficulty, and for users with a low willingness to pay but who are inactive due to monotonous gameplay. However, current technology cannot do this, resulting in poor intervention effects and an inability to effectively reduce user churn. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for activating dormant game users based on artificial intelligence, which effectively improves the success rate of activating dormant users, enhances the user's gaming experience and retention rate, and has good adaptability to different game types. It can also deeply analyze the strategy decision-making process to optimize the intervention strategy, thereby solving at least one of the above-mentioned problems in the prior art.

[0007] In a first aspect, the present invention provides a method for activating dormant game users based on artificial intelligence, the method specifically comprising: Real-time collection of multidimensional user behavior flow data; and construction of a spatiotemporal graph structure containing user operation sequences, virtual location trajectories, and social interactions using the multidimensional behavior flow data. Graph neural networks are used to learn node embeddings in spatiotemporal graph structures to capture the spatiotemporal correlation features of different silence modes, and a temporal convolution module with adjustable receptive field is combined to dynamically identify the current silence state of the target user. The identified current silence state is input into a two-layer reinforcement learning engine. The silence pattern is accurately classified through the deep Q-network at the bottom layer, and the optimal intervention strategy is dynamically selected from the policy library by the proximal policy optimization network at the top layer in combination with the classification results. Based on the selected intervention strategy, the AI ​​virtual characters in the game can perform personalized multimodal interactive outreach in a non-intrusive manner; Collect real-time user feedback data on this interaction, and continuously optimize the model parameters of the graph neural network and the two-layer reinforcement learning engine through online learning.

[0008] Secondly, the present invention provides an artificial intelligence-based system for activating dormant game users, the system specifically comprising: The data acquisition module is used to collect multi-dimensional behavioral flow data of users in real time, and to construct a spatiotemporal graph structure that includes user operation sequences, virtual location trajectories and social interactions through the multi-dimensional behavioral flow data; The state recognition module is used to learn node embeddings of spatiotemporal graph structures using graph neural networks to capture the spatiotemporal correlation features of different silence modes, and dynamically identify the current silence state of the target user by combining a temporal convolution module with an adjustable receptive field. The silence classification module is used to input the identified current silence state into the two-layer reinforcement learning engine. The silence pattern is accurately classified through the bottom-layer deep Q-network, and the optimal intervention strategy is dynamically selected from the policy library by the upper-layer proximal policy optimization network in combination with the classification results. The interactive outreach module is used to enable AI virtual characters in the game to perform personalized multimodal interactive outreach in a non-intrusive manner, based on the selected intervention strategy. The parameter optimization module is used to collect real-time feedback data from users regarding this interaction and continuously optimize the model parameters of the graph neural network and the two-layer reinforcement learning engine through online learning.

[0009] Thirdly, the present invention provides a computer device, comprising: a memory and a processor, and a computer program stored in the memory, wherein when the computer program is executed on the processor, it implements the AI-based method for activating silent game users as described in any of the above methods.

[0010] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the AI-based method for activating silent game users as described in any of the above methods.

[0011] Compared with the prior art, the present invention has at least one of the following technical effects: 1. This invention effectively improves the success rate of activating dormant users, enhances the user gaming experience and retention rate, and has good adaptability to different game types. It can also deeply analyze the strategy decision-making process to optimize intervention strategies.

[0012] 2. This invention overcomes the problem of delayed identification in existing technologies, and can detect users in the early stages of falling into a silent state in a timely manner, thus buying valuable time for subsequent intervention measures and effectively reducing the risk of user churn.

[0013] 3. This invention solves the problem of the single interaction method in the existing technology, greatly improves the user's acceptance and participation in the interaction, and effectively enhances the user's enthusiasm to re-engage in the game.

[0014] 4. This invention can flexibly select the most appropriate wording and intervention methods based on various user characteristics and behavioral patterns, such as payment intentions, game progress, and emotional state. It overcomes the problem of rigidity in existing technical strategies, significantly improves the activation effect on dormant users, reduces user churn rate, and brings better economic benefits to game operation.

[0015] 5. This invention collects real-time user feedback data on each interaction and continuously optimizes the model parameters of the graph neural network and two-layer reinforcement learning engine through online learning. This allows the entire system to continuously improve and optimize itself in response to changes in user behavior and the game environment, maintaining a high accuracy rate in identifying silent users and the effectiveness of intervention strategies, thus providing a strong guarantee for the long-term stable operation of the game.

[0016] 6. This invention constructs a spatiotemporal graph structure by collecting multidimensional data in real time, and combines graph neural networks and temporal convolution modules to accurately and dynamically identify the state of inactive game users, providing an accurate basis for subsequent activation intervention and reducing the risk of user churn.

[0017] 7. This invention utilizes graph neural networks to aggregate neighbor features to update user node representations. After constructing a temporal feature sequence, it is processed by a temporal convolution module, which can effectively capture the spatiotemporal correlation features of different silence modes and achieve accurate dynamic identification of the current silence state.

[0018] 8. This invention is based on a pluggable Transformer module of a two-layer reinforcement learning engine. By integrating team collaboration, plot context and player emotional experience information through different attention modules, the composite gameplay attention modules are weighted and summed to obtain an enhanced state vector, which enhances the adaptability of the strategy to different game types.

[0019] 9. This invention encapsulates the near-end policy optimization network to construct a background distribution dataset, uses a sliding time window to divide state-action pairs, calculates the SHAP contribution matrix and forms a heatmap, filters key features and identifies abrupt change moments, which helps to deeply understand the policy decision-making process and optimize intervention strategies.

[0020] 10. This invention generates interactive intent based on intervention strategies combined with game scenarios and user profiles. Figure 3 The tuple uses a large language model to generate personalized dialogue text, and schedules multimodal interactions based on AI role rules and game engine state to activate dormant users through personalized, non-intrusive multimodal interactions. Attached Figure Description

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

[0022] Figure 1 This is a flowchart illustrating a method for activating dormant game users based on artificial intelligence, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of an artificial intelligence-based game silent user activation system provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0023] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0024] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0025] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0026] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0027] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0028] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0029] In this application embodiment, the entity executing the process includes a terminal device. This terminal device includes, but is not limited to, devices capable of executing the methods disclosed in this application, such as servers, computers, smartphones, and tablets. Figure 1 A flowchart illustrating an embodiment of the artificial intelligence-based method for activating dormant game users disclosed in this invention is shown below in detail: S101 collects multi-dimensional behavior flow data of users in real time, and constructs a spatiotemporal graph structure containing user operation sequences, virtual location trajectories and social interactions through the multi-dimensional behavior flow data; S102 utilizes graph neural networks to learn node embeddings in spatiotemporal graph structures to capture spatiotemporal correlation features of different silence modes, and combines a temporal convolutional module with an adjustable receptive field to dynamically identify the current silence state of the target user. S103, the identified current silence state is input into the two-layer reinforcement learning engine, the silence mode is accurately classified through the bottom deep Q network, and the optimal intervention strategy is dynamically selected from the policy library by the upper proximal policy optimization network in combination with the classification results. S104, based on the selected intervention strategy, enable the AI ​​virtual character in the game to perform personalized multimodal interactive outreach in a non-intrusive manner; S105 collects real-time feedback data from users regarding this interaction and continuously optimizes the model parameters of the graph neural network and the two-layer reinforcement learning engine through online learning.

[0030] In this embodiment, during game operation, a data acquisition module deployed on the game server collects multi-dimensional behavioral flow data of users in real time. This data includes user operation sequences, such as clicks, swipes, and drags on the game interface; virtual location trajectories, i.e., the user's movement path information in the game's virtual scene; and social interaction data, including records of social behaviors such as chatting, teaming up, and trading between users and other players. After collecting this data, a data processing algorithm is used, with the user as a node and the user operation sequence, virtual location trajectory, and social interaction information as edge attributes, to construct a spatiotemporal graph structure containing the user operation sequence, virtual location trajectory, and social interaction. For example, in a role-playing game, when a user controls their character to move on the map, engage in combat, and communicate with other players, the system records these behaviors in real time and constructs a spatiotemporal graph structure according to the above rules. This structure can comprehensively reflect the user's behavioral patterns and social relationships in the game.

[0031] The constructed spatiotemporal graph structure is input into a pre-trained graph neural network model. The graph neural network learns node embeddings within the spatiotemporal graph structure, capturing the spatiotemporal correlation features of different silence patterns by analyzing the connections between nodes and the attributes of edges in the graph. For example, some users may become silent in a certain level due to excessive difficulty; their operation sequence might show repeated attempts in a certain area without success, the virtual location trajectory might linger in that area for a long time, and social interaction might decrease. The graph neural network can learn these feature patterns. Simultaneously, combined with a temporal convolutional module with an adjustable receptive field, this module can dynamically adjust the size of the receptive field based on the temporal characteristics of the data, further analyzing the temporal information in the spatiotemporal graph structure to dynamically identify the current silence state of the target user. For instance, when a user's operation frequency significantly decreases over a period of time, and there are no significant changes in virtual location and social interaction, the system determines that the user is in a silent state.

[0032] The identified current silence state is input into a two-layer reinforcement learning engine. This engine consists of a bottom-layer deep Q-network and an upper-layer proximal policy optimization network. The bottom-layer deep Q-network accurately classifies the silence pattern. Based on the input silence state information and previously learned experience, it determines the type of silence state, such as whether it's caused by level difficulty, monotonous gameplay, or difficulty making payment decisions. The upper-layer proximal policy optimization network combines the classification results from the bottom layer and dynamically selects the optimal intervention strategy from a pre-built strategy library. The strategy library contains various intervention strategies for different silence types. For example, for users silent due to level difficulty, it might provide level walkthrough hints or reduce the level difficulty; for users silent due to monotonous gameplay, it might recommend new gameplay or activities. For instance, when the deep Q-network determines that a user is silent due to level difficulty, the proximal policy optimization network will select an intervention strategy from the strategy library that provides level walkthrough hints.

[0033] Depending on the selected intervention strategy, the in-game AI virtual character performs personalized multimodal interactive outreach in a non-intrusive manner. Non-intrusive means that it will not excessively interfere with the user's normal gaming experience, such as interacting with the user through in-game chat windows or system prompts. Multimodal interactive outreach includes various forms such as text, voice, and images. For example, if the selected intervention strategy is to provide level walkthrough hints, the AI ​​virtual character can send detailed level walkthroughs in text form through the chat window, along with relevant image examples to help the user better understand; for some important hints, voice broadcasts can also be used to enhance the outreach effect.

[0034] After the AI ​​virtual character performs an interactive interaction, the system collects real-time feedback data from the user. This feedback data includes whether the user clicked on the prompt, whether they continued playing the game after receiving the prompt, and the length of time they continued playing. After collecting the feedback data, an online learning mechanism feeds this data into a graph neural network and a two-layer reinforcement learning engine, continuously optimizing their model parameters. For example, if the system finds that a certain intervention strategy elicits positive user feedback—that is, the user continues playing for a significantly longer period—it will adjust the model parameters to favor that intervention strategy in similar situations. Conversely, if the user feedback is negative, the system will reduce the probability of selecting that strategy, thereby gradually improving the intervention effect and reducing user churn.

[0035] For example, in role-playing games, multi-dimensional behavioral flow data of users is collected in real time, such as the user's movement trajectory in various map areas, combat operation sequences with different monsters, and social interaction information such as chat with other players, to construct a spatiotemporal graph structure. Graph neural networks and temporal convolutional modules are used to identify user inactivity states. For example, if a user stays in a high-level dungeon area for a long time and reduces combat operations, it is determined that the user is inactive due to the dungeon's difficulty. A two-layer reinforcement learning engine selects intervention strategies, such as providing the user with a video link to clear the dungeon or recommending teammates. The in-game AI virtual character reaches the user by popping up video links and team invitations in the chat window. User feedback is collected. If the user clicks the video link and continues to challenge the dungeon in a team, the model parameters are optimized to increase the probability of selecting such intervention strategies.

[0036] For example, in strategy games, multi-dimensional user behavior flow data is collected, including city-building operation sequences, resource acquisition trajectories, and alliance social interactions with other players, to construct a spatiotemporal graph structure. When a user is found to be inactive for an extended period during resource allocation or technology research, it is identified as silent due to strategic decision-making difficulties. A two-layer reinforcement learning engine selects intervention strategies, such as providing optimized resource allocation solutions or recommended technology research routes. An AI virtual character displays the recommended solutions through system prompts, accompanied by brief text explanations. Feedback data is collected based on whether the user adopts the recommended solutions and subsequent game development, optimizing model parameters and improving the activation effect on silent users in strategy games.

[0037] For example, in casual puzzle games, multidimensional user behavior flow data is collected, including game level challenge operation sequences, level selection trajectories, and interactions with friends, to construct a spatiotemporal graph structure. If a user stops operating after repeatedly failing at a high-difficulty level, it is identified as silence due to level difficulty. A two-layer reinforcement learning engine selects intervention strategies, such as providing hints on how to solve the level or unlocking similar easier levels for the user to practice. An AI virtual character provides voice prompts to inform the user of the solution or unlock information. Feedback data on whether users re-challenge levels and the results is collected to optimize model parameters and reduce user churn in casual puzzle games.

[0038] In some embodiments, step S101 above, which involves constructing a spatiotemporal graph structure containing user operation sequences, virtual location trajectories, and social interactions using multidimensional behavior flow data, specifically includes: Extract user operation sequences to characterize user operation logic, user location trajectories to reflect user movement patterns in the virtual environment, and social interaction data to reflect user interaction patterns with game characters from multidimensional behavior flow data. The nodes and edges of the spatiotemporal graph structure are set according to the user operation sequence, user location trajectory and social interaction data. The nodes include user nodes with user identifier as the core and function nodes with key function points in the game as the core. The edges are the spatiotemporal interaction relationships between nodes. A sliding time window is used to aggregate all active nodes and edges that meet the conditions within the current time window, generating a static spatiotemporal graph snapshot within the current time window; All time window static spatiotemporal graph snapshots are stored in a structured manner in the form of adjacency matrix and feature matrix to form a spatiotemporal graph structure.

[0039] In this embodiment, three key types of information are extracted from the user's multidimensional behavior flow data. First, there is the user operation sequence, which characterizes the user's operational logic. This sequence records various user actions on the game interface in detail. For example, in role-playing games, it shows the order and time intervals of actions such as character movement, attacking, and skill release; in strategy games, it shows the execution flow of actions such as building construction, resource allocation, and troop deployment. Second, there is the user location trajectory, which reflects the user's movement patterns in the virtual environment. Using the game map as a reference, it records the user's movement path in the virtual scene, including the areas traversed, the locations where the user stopped, and the movement speed. For example, in open-world games, it shows the user's route when exploring different areas. Finally, there is social interaction data, which reflects the user's interaction patterns with game characters. This includes chat content, team information, transaction records, and friend interactions between the user and other players. For example, it includes voice communication between the user and teammates in the game, and teamwork for completing tasks together.

[0040] After extracting the aforementioned key information, the nodes and edges of the spatiotemporal graph structure are constructed. Nodes are divided into two types: user nodes, centered on user identifiers, with each user corresponding to a unique user node carrying that user's unique identification information to distinguish different users; and function nodes, centered on key functional points in the game. These function points can be specific levels, important quest posting points, special item acquisition points, etc. For example, in an adventure game, a mysterious treasure cave or a challenging BOSS level can serve as function nodes. Edges represent the spatiotemporal interaction relationships between nodes, reflecting the interaction between users and function points, as well as between users themselves, at different times and spaces. For example, when a user enters a level, an edge is established between the user node and the corresponding function node for that level. The edge's attributes can record information such as the time the user entered the level and the duration of their stay within the level. When a user teams up with other players, an edge is established between the corresponding user nodes. The edge's attributes can include information such as the time the team formed the team and the tasks completed by the team.

[0041] To better capture the dynamic changes and temporal characteristics of user behavior in the game, a sliding time window method is used to aggregate nodes and edges. A suitable time window size is set, for example, every 10 minutes. Within the current time window, all active nodes are selected—those with recorded actions or interactions within that time window. Edges that meet certain criteria—representing valid interactions between nodes within that time window—are also identified. The selected active nodes and matching edges are aggregated to generate a static spatiotemporal graph snapshot of the current time window. This snapshot records the interactions between users and functionalities, and between users themselves, within that time window. For example, if some users enter a level or team up with other users to complete a task within a time window, this information will be aggregated into the static spatiotemporal graph snapshot for that time window.

[0042] All static spatiotemporal graph snapshots generated across time windows are stored in a structured manner for subsequent analysis and processing. This storage employs adjacency matrices and feature matrices. The adjacency matrix describes the connections between nodes, with elements indicating the existence and weight of edges between nodes. For example, if two user nodes have a team interaction, the corresponding element in the adjacency matrix can be set to a value representing the team strength. The feature matrix stores the feature information of each node. For user nodes, this information may include user level, payment history, and game behavior; for functional nodes, it may include level difficulty and task rewards. The combination of adjacency and feature matrices comprehensively records all information within the spatiotemporal graph structure, forming the final spatiotemporal graph structure and providing a foundation for subsequent node embedding learning and silent state recognition using graph neural networks.

[0043] For example, in role-playing games, the first step is to collect multi-dimensional behavioral flow data from users through a data acquisition module. From this data, user operation sequences are extracted, such as the order of skill releases in combat, and the order of movement and trading in towns; user location trajectories, such as exploration paths in different areas of the game map and movement routes in dungeons; and social interaction data, such as chat content, team information, and mentor-apprentice relationships between users and other players. Then, nodes and edges of a spatiotemporal graph structure are set up. User nodes are centered around the player's account ID, and functional nodes can include various dungeon entrances, the locations of important quest NPCs, and the trading market. Edge attributes record the interaction time and frequency between users and functional points, as well as the social interactions between users. A sliding time window is used, for example, every 15 minutes, aggregating active nodes and edges that meet certain conditions within the current window to generate a static spatiotemporal graph snapshot. Finally, snapshots of all windows are stored in the form of an adjacency matrix and a feature matrix. The adjacency matrix records node connections, and the feature matrix stores node feature information such as user level, dungeon completion status, and social relationships, forming a complete spatiotemporal graph structure.

[0044] For example, in strategy games, multidimensional user behavior flow data is first collected to extract user operation sequences, such as the building upgrade order in city construction, the allocation order of resource collection, and the order of troop mobilization commands; user location trajectories, such as the user's movement paths to different resource points on the game map and the troop marching routes; and social interaction data, such as the user's alliance relationships with other players, trade records, and war assistance information. When setting nodes and edges, user nodes are centered around the player account, and functional nodes include resource points, military fortresses, and alliance headquarters in the game. Edges record the interaction information between users and functional points, as well as the social interaction information between users. A sliding time window, such as every 20 minutes, is used to aggregate nodes and edges to generate a static spatiotemporal graph snapshot. The snapshot is stored in the form of an adjacency matrix and a feature matrix. The adjacency matrix reflects the node connectivity, and the feature matrix stores node feature information such as the number of user resources, troop strength, and alliance status, constructing a spatiotemporal graph structure suitable for strategy games.

[0045] In some embodiments, in step S102 above, the step of using a graph neural network to learn node embeddings in the spatiotemporal graph structure to capture the spatiotemporal correlation features of different silence modes, and combining this with a temporal convolutional module with an adjustable receptive field to dynamically identify the current silence state of the target user, specifically includes: The spatiotemporal graph structure is input into the graph neural network, and the feature representation of the user node is updated by aggregating the feature information of the neighboring nodes to obtain the user node embedding vector that integrates spatial context information in each time slice. The embedding vectors of user nodes from multiple consecutive time slices are concatenated in chronological order to construct a temporal feature sequence that reflects the dynamic changes in user status. The temporal feature sequence is input into a temporal convolutional module with adjustable receptive field. By designing convolutional kernels with different dilation rates, the user's behavior patterns at different time granularities are dynamically captured, and the output features of multiple branches are fused to generate multi-scale fused features. Multi-scale fused features are input into a classifier containing a fully connected layer. By calculating the probability distribution of a preset silence pattern, the current silence state of the target user is identified.

[0046] In this embodiment, the constructed spatiotemporal graph structure is input into a graph neural network. The graph neural network possesses powerful feature extraction and aggregation capabilities, operating on each user node in the spatiotemporal graph structure. Specifically, for each user node, the graph neural network aggregates the feature information of its neighboring nodes. Neighboring nodes can be other user nodes that interact with the user in the game, or functional nodes related to the user's actions. Through this aggregation operation, the contextual information of the user in the game space can be fully considered, such as the user's social relationships with other players, and the user's interaction with game functionalities. After aggregating the feature information of neighboring nodes, the feature representation of the user node is updated, ultimately obtaining a user node embedding vector that incorporates spatial contextual information for each time slice. The user node embedding vector represents the user's game behavior and state characteristics within a specific time slice in a more abstract and efficient way. For example, in a multiplayer online role-playing game, if a user teams up with teammates to run a dungeon within a certain time slice, the graph neural network will aggregate the feature information of teammate nodes and dungeon functional nodes, update the feature representation of the user node, and generate an embedding vector reflecting its dungeon-running state within that time slice.

[0047] After obtaining the user node embedding vector for each time slice, these embedding vectors from multiple consecutive time slices are concatenated in chronological order. Since gaming is a dynamic process, user behavior and states change across different time slices. This concatenation operation constructs a temporal feature sequence reflecting the dynamic changes in user states. This temporal feature sequence contains information about user behavior trends and state changes over a period of time. For example, over a period of time, a user gradually progresses from the initial exploration stage to challenging higher-difficulty levels; by concatenating the embedding vectors from different time slices, the temporal feature sequence clearly presents this process.

[0048] The constructed temporal feature sequence is input into a temporal convolutional module with adjustable receptive fields. This module is a key component of this method, dynamically capturing user behavior patterns at different temporal granularities by designing convolutional kernels with varying dilation rates. Kernels with different dilation rates can observe user behavior from different time scales. For example, kernels with smaller dilation rates can capture rapid behavioral changes in the short term (e.g., within minutes), such as frequent skill releases; kernels with larger dilation rates can capture long-term behavioral trends (e.g., within hours or even days), such as gradual increases in game level. Through this multi-scale temporal granularity observation, the temporal convolutional module can comprehensively understand user behavior patterns. Simultaneously, this module fuses the output features from multiple branches, integrating the feature information captured at different time scales to generate multi-scale fused features. These multi-scale fused features more comprehensively and accurately reflect the user's behavioral characteristics across different time dimensions, providing rich information for subsequent silent state identification.

[0049] The generated multi-scale fusion features are input into a classifier containing fully connected layers. The classifier's role is to classify and determine the user's current state by calculating the probability distribution of preset silence patterns. These preset silence patterns are summarized based on a large amount of user behavior data and game operation experience, such as silence due to excessive level difficulty, monotonous gameplay, or confusion regarding payment decisions. The classifier calculates the probability of each silence pattern based on the degree of matching between the multi-scale fusion features and these preset silence patterns. Based on the probability, the silence pattern with the highest probability is selected as the target user's current silence state. For example, if the calculation shows that the probability of silence due to excessive level difficulty is the highest, then the user can be determined to be in a silence state caused by excessive level difficulty. In this way, the target user's current silence state can be accurately and dynamically identified, providing a basis for subsequent targeted intervention strategies.

[0050] For example, in competitive games, a spatiotemporal graph structure is first constructed, incorporating information such as player nodes and functional nodes of the game venue. A graph neural network aggregates the feature information of neighboring nodes to obtain the user node embedding vector for each time slice. For instance, in a team competitive game, when a player coordinates an attack with teammates in a certain time slice, the graph neural network aggregates the feature information of teammate nodes and functional nodes along the attack route, updating the feature representation of that player node. Then, the embedding vectors from multiple consecutive time slices are concatenated into a temporal feature sequence, reflecting the player's state changes during the game, such as the transition from a defensive to an offensive phase. Next, the temporal feature sequence is input into a temporal convolution module, where convolutional kernels with different dilation rates capture changes in the player's short-term operation frequency and long-term tactical execution. After generating multi-scale fusion features, this is input into a classifier for silent state identification. If a player suddenly stops operating at a crucial moment in the game, the classifier calculates based on the multi-scale fusion features and finds that the probability of silence due to excessive game pressure is high, thus identifying that the player is in a silent state due to excessive game pressure.

[0051] For example, in casual games, a spatiotemporal graph structure is first constructed, including player nodes, game item function nodes, etc. A graph neural network aggregates the feature information of neighboring nodes to obtain the user node embedding vector. For instance, during item collection, the graph neural network aggregates the feature information of item function nodes and other nearby player nodes. The embedding vectors from multiple time slices are concatenated to form a temporal feature sequence, reflecting the player's behavioral rhythm changes in the casual game. The temporal convolution module uses convolution kernels with different dilation rates to capture the player's item collection frequency in the short term and game progress in the long term. After generating multi-scale fused features, these are input into a classifier. If a player stops operating after repeatedly failing to complete a casual task, the classifier calculates that the highest probability of silence is due to a mismatch between the task difficulty and their own ability, thus identifying that the player is in a silent state due to the task's difficulty.

[0052] Furthermore, the step of inputting the spatiotemporal graph structure into a graph neural network and updating the feature representation of user nodes by aggregating the feature information of neighboring nodes to obtain the user node embedding vector fused with spatial context information for each time slice specifically includes: The spatiotemporal graph structure is input into the graph neural network. For each central user node and its neighboring nodes in the spatiotemporal graph structure, the original attention coefficient between them is calculated through the graph attention mechanism. The original attention coefficient is used to quantify the importance of the neighboring nodes to the central user node. The original attention coefficients of all neighboring nodes are normalized using softmax to obtain the normalized attention weights corresponding to each neighboring node. The normalized attention weights are used to reflect the degree of contribution of each neighboring node to the user node's behavioral intent at the current moment. The features of neighboring nodes are weighted and aggregated based on normalized attention weights, and the aggregated neighborhood features are fused with the features of the central user node itself. After passing through a nonlinear activation function, the updated user node feature representation of the current layer is obtained. By stacking multi-layer graph attention networks and combining them with a multi-head attention mechanism, the user node feature representation of each central user node is aggregated with spatial context information within the range of multiple-order neighbors, generating a user node embedding vector that integrates the game environment context for each time slice.

[0053] In this embodiment, the constructed spatiotemporal graph structure is input into a graph neural network. This spatiotemporal graph structure contains numerous central user nodes and their associated neighbor nodes. For each central user node and its neighbor nodes, a graph attention mechanism is used to calculate the initial attention coefficient between them. This initial attention coefficient plays a crucial role, quantifying the importance of neighbor nodes to the central user node. For example, in a multiplayer online role-playing game, the central user node represents the player currently performing game actions. Its neighbor nodes might be teammates who are in a party with that player, other player nodes within the same instance area, or nodes with whom the player has trade interactions. The initial attention coefficient calculated by the graph attention mechanism reflects the magnitude of the influence of these different neighbor nodes on the central player node's game behavior. For instance, teammates may have a greater influence on the central player node during team battles, resulting in a relatively high initial attention coefficient; while other player nodes that are far away and have no direct interaction may have a lower initial attention coefficient.

[0054] After calculating the original attention coefficients of all neighboring nodes, these coefficients need to be normalized using softmax. This step transforms the original attention coefficients into normalized attention weights for each neighboring node. Normalized attention weights clearly reflect the contribution of each neighboring node to the user node's behavioral intent at the current moment. After softmax normalization, the sum of the normalized attention weights of all neighboring nodes is 1. For example, in the aforementioned multiplayer online role-playing game, assuming a central player node has three neighboring nodes, after softmax normalization, one teammate node might have a normalized attention weight of 0.6, another teammate node 0.3, and the remaining unrelated players 0.1. This indicates that at the current moment, the first teammate node contributes the most to the central player node's behavioral intent, followed by the second teammate node, while the remaining unrelated players contribute the least.

[0055] Based on the normalized attention weights obtained in the previous step, the features of neighboring nodes are weighted and aggregated. Specifically, the features of each neighboring node are multiplied by its corresponding normalized attention weight, and then the weighted features of all neighboring nodes are summed to obtain the aggregated neighborhood features. This aggregated neighborhood feature integrates the significant influence of each neighboring node on the central user node. Next, we fuse the aggregated neighborhood features with the features of the central user node itself. This fusion process fully combines the influence of neighboring nodes and the characteristics of the central user node, making the updated features more comprehensive and accurate. Finally, the fused features are processed through a non-linear activation function. The role of the non-linear activation function is to introduce non-linear factors, enhance the expressive power of the model, and enable the model to better fit complex data relationships. After processing with the non-linear activation function, we obtain the updated user node feature representation of the current layer.

[0056] To enable the user node feature representation of each central user node to aggregate spatial context information from multiple levels of neighbors, a stacked multi-layer graph attention network combined with a multi-head attention mechanism is employed. The stacked multi-layer graph attention network allows information to be transferred and fused between different layers, with each layer capturing neighbor node information of different orders. For example, the first layer might primarily capture first-order neighbor node information for the central user node, while the second layer could further capture second-order neighbor node information, and so on. Through multi-layer stacking, the central user node can obtain a wider range of neighbor node information. The multi-head attention mechanism calculates attention coefficients and aggregates neighbor node features from multiple different perspectives. Each head can focus on different feature subspaces, thereby extracting information from different aspects. By combining the multi-head attention mechanism, the generated user node feature representation becomes richer and more comprehensive. Finally, after multi-layer stacking and the combination of the multi-head attention mechanism, we generate a user node embedding vector that integrates the game environment context for each time slice. This embedding vector efficiently and comprehensively represents the user's game behavior and state features within a specific time slice, taking into account the influence of the surrounding game environment and neighbor nodes.

[0057] For example, in strategy games, players need to perform complex operations such as resource management, army command, and territory acquisition. First, a spatiotemporal graph structure is constructed, incorporating player nodes, resource nodes, and army nodes. When this spatiotemporal graph structure is input into a graph neural network, for each central player node, its neighboring nodes may include player nodes in adjacent territories and nearby resource nodes. The initial attention coefficients are calculated using a graph attention mechanism. For instance, a stronger player node in an adjacent territory might have a higher initial attention coefficient for the central player node because it may pose a threat to the central player's territory security; conversely, a node that is far away and has fewer resources might have a lower initial attention coefficient. After softmax normalization to obtain normalized attention weights, the features of neighboring nodes are weighted and aggregated. For example, the army deployment features of a stronger neighboring player, after weighted aggregation, have a greater impact on the central player node. The aggregated neighborhood features are then fused with the central player node's own resource management features and processed through a non-linear activation function to obtain the current layer's feature representation. By stacking multi-layer graph attention networks and combining them with a multi-head attention mechanism, the central player node can comprehensively consider various information within the range of multiple levels of neighbors, such as the strength of players at different distances and the distribution of resources, to generate a user node embedding vector that incorporates the game environment context. This vector can accurately reflect the player's strategic state in the current game environment.

[0058] For example, in casual puzzle games, such as match-3 games, the spatiotemporal graph structure includes player nodes, item nodes, and level nodes. After inputting this spatiotemporal graph structure into a graph neural network, for the central player node, neighboring nodes might be usable item nodes or adjacent level nodes. When calculating the original attention coefficients, item nodes with better performance may have higher original attention coefficients for the central player node because they help the player complete levels more easily. After normalization to obtain normalized attention weights, the item node features are weighted and aggregated. The aggregated neighborhood features are then fused with the central player node's own elimination operation features, and processed through a non-linear activation function to obtain the current layer's feature representation. Through multi-layer stacking and multi-head attention mechanisms, the central player node can integrate information from multiple levels of its neighbors, such as the effects of different items and the difficulty differences of adjacent levels, to generate a user node embedding vector that incorporates the game environment context. This vector reflects the player's operational state and needs in the current casual puzzle game.

[0059] Furthermore, the step of inputting the temporal feature sequence into a temporal convolutional module with adjustable receptive field capability, dynamically capturing user behavior patterns at different temporal granularities by designing convolutional kernels with different dilation rates, and fusing the output features of multiple branches to generate multi-scale fused features, specifically includes: The temporal feature sequence is input into multiple dilated convolutional branches in parallel, and a different dilation rate is configured for each dilated convolutional branch to achieve feature extraction of the input sequence at different temporal granularities. The feature maps output by each dilated convolution branch are unified and activated in the channel dimension to obtain multiple intermediate feature maps with the same number of channels. Multiple intermediate feature maps are stitched together along the channel dimension and input into the fusion layer for cross-scale information interaction, generating multi-scale fusion features that integrate local details and global trends.

[0060] In this embodiment, the constructed temporal feature sequence is input in parallel into multiple dilated convolutional branches. Each dilated convolutional branch is carefully configured with a different dilation rate. The dilation rate is a key parameter in dilated convolution, determining the sampling interval of the convolutional kernel. Different dilation rates allow the convolutional kernel to focus on the features of the input sequence at different time scales. For example, in a dilated convolutional branch with a shorter time granularity, the dilation rate is smaller, and the convolutional kernel samples the input sequence more densely, thereby capturing rapid changes and local details in user behavior within a short period, such as frequent user actions or sudden pauses in operation. In a dilated convolutional branch with a longer time granularity, the dilation rate is larger, and the sampling interval of the convolutional kernel is larger, enabling it to focus on the overall trend and global features of user behavior over a longer period, such as changes in overall user activity over a period of time or the continuous use of a certain game function. Through this parallel input and different dilation rate settings, each dilated convolutional branch can extract features from the input temporal feature sequence at different time granularities.

[0061] After extracting features from the input sequence, each dilated convolution branch outputs its own feature map. Due to the different configurations of the dilated convolution branches, their output feature maps may differ in the channel dimension. To facilitate subsequent feature fusion, we need to unify these feature maps in the channel dimension. This process can be achieved through specific convolution operations or parameter adjustments to ensure that the feature maps output by each dilated convolution branch have the same number of channels. After unifying the channel number, activation processing is applied to each feature map. The purpose of activation processing is to introduce non-linearity, enhance the model's expressive power, and enable the model to better fit complex data relationships. Commonly used activation functions, such as ReLU (Revised Linear Unit), can set negative values ​​in the feature map to zero, retain positive values, and enhance their feature representation, thereby making the feature map more discriminative and representative. After activation processing, we obtain multiple intermediate feature maps with the same number of channels.

[0062] After obtaining multiple intermediate feature maps, they are concatenated along the channel dimension. Channel concatenation is a simple and effective feature fusion method that integrates features extracted from different time granularities by different dilated convolution branches. For example, local detail features extracted by short-time granularity branches and global trend features extracted by long-time granularity branches are concatenated to form a feature set containing rich information. However, channel concatenation alone is insufficient to fully explore the correlation and complementarity between features at different scales. Therefore, the concatenated features are input into a fusion layer for cross-scale information interaction. The fusion layer can employ specific neural network structures, such as fully connected layers or more complex convolutional neural network structures. In the fusion layer, features at different scales can communicate and influence each other, and the model can learn how to comprehensively judge user behavior patterns based on features at different scales. For example, through cross-scale information interaction, the model can understand the relationship between frequent user operations in a short period of time and overall activity over a long period of time, thereby more accurately grasping user behavior characteristics. After processing by the fusion layer, a multi-scale fused feature integrating local details and global trends is finally generated. This multi-scale fusion feature can comprehensively and accurately reflect users' behavioral patterns at different time granularities, providing strong support for the subsequent accurate identification of users' silent states.

[0063] For example, in competitive games, players have high operation frequency and fast reaction speed, resulting in a fast-paced game. First, a temporal feature sequence is constructed, recording various player operation information during the game, such as the time and frequency of movement, attack, and skill release. This temporal feature sequence is then input in parallel into multiple dilated convolution branches. For dilated convolution branches with short temporal granularity, a smaller dilation rate is set to capture local details such as continuous attack operations or rapid changes in movement trajectory within a short period. For dilated convolution branches with long temporal granularity, a larger dilation rate is set to focus on the overall performance trend of the player in a complete game, such as aggressive early-game offense, stable mid-game defense, and decisive late-game sprint. After each dilated convolution branch outputs feature maps, the number of channels is unified and activation processing is performed to obtain intermediate feature maps. These intermediate feature maps are concatenated along the channel dimension and input into a fusion layer. The fusion layer learns the relationship between the player's high-frequency operations within a short period and the overall performance trend of the game, generating a multi-scale fusion feature that integrates local details and global trends. This multi-scale fusion feature can help accurately determine whether a player has entered a silent state due to competition pressure, operational difficulties, or other reasons.

[0064] For example, in role-playing games, players focus more on story experience, task completion, and character development, resulting in a relatively slow game pace. The constructed temporal feature sequence includes player action information at different task stages, interaction information with NPCs (non-player characters), and character attribute changes. When the temporal feature sequence is input into the dilated convolution branch, the short-time granularity branch captures the player's action details when completing a specific task, such as frequently clicking on task items or engaging in dialogue with specific NPCs within a short period. The long-time granularity branch focuses on the player's behavioral trends throughout the game's story progression, such as preferences for different storylines and long-term planning for character development. After channel unification and activation processing to obtain intermediate feature maps, they are then stitched together and fused. The fusion layer integrates the player's action details in specific tasks and behavioral trends throughout the story progression to generate multi-scale fused features. This feature can be used to determine whether the player has entered a silent state due to story confusion, excessive task difficulty, or character development bottlenecks, allowing for timely activation measures.

[0065] In some embodiments, in step S103 above, the step of inputting the identified current silence state into a two-layer reinforcement learning engine, accurately classifying the silence pattern through a bottom-layer deep Q-network, and dynamically selecting the optimal intervention strategy from the policy base through an upper-layer proximal policy optimization network based on the classification results, specifically includes: The identified current silence state is constructed as a state space, which is then input into the deep Q-network at the bottom layer of the two-layer reinforcement learning engine. The Q-values ​​of each intervention strategy are output through the multi-layer fully connected layers and competitive architecture of the deep Q-network, and the silence mode probability distribution is obtained by accurate classification through the softmax classification layer. The state space, the Q-values ​​of each intervention strategy, and the probability distribution of the silence mode are concatenated to form an enhanced state vector; The enhanced state vector is input into the proximal policy optimization network of the upper layer of the two-layer reinforcement learning engine. The proximal policy optimization network outputs the probability distribution of the intervention policy and selects the optimal intervention action through sampling or greedy policy.

[0066] In this embodiment, the current silent state of the identified target user is constructed as a state space. This state space is a comprehensive dataset containing various key information related to the user's current silent state, such as the user's game scenario (whether it's a payment decision scenario, a level failure scenario, or an activity participation scenario), the user's previous behavioral patterns (operation frequency, social interaction, etc.), current game progress (level progress, task completion status, etc.), and other factors that may affect the user's silence. The constructed state space is input into the deep Q-network at the bottom layer of the two-layer reinforcement learning engine. The deep Q-network is a multi-layered neural network that performs complex feature extraction and transformation on the input state space through multiple fully connected layers. After the multiple fully connected layers, a competitive architecture is used to further process the data, which can highlight the differences and importance between different features. After these processing steps, the deep Q-network outputs the Q-values ​​of each intervention strategy. The Q-value is an indicator that measures the expected return obtained by taking a certain intervention strategy in a specific state. After obtaining the Q-values ​​of each intervention strategy, these Q-values ​​are then processed through a softmax classification layer. The softmax classification layer can convert Q-values ​​into probability distributions, thereby accurately classifying the silence mode and clarifying which specific silence mode type the current user's silence state belongs to, such as silence caused by level difficulty, silence caused by monotonous gameplay, or silence caused by confusion about payment, etc.

[0067] After accurately classifying the silence patterns and obtaining the Q-values ​​and silence pattern probability distributions for each intervention strategy, this information is concatenated with the initially constructed state space. The state space provides basic information about the user's current silence state, the Q-values ​​of each intervention strategy reflect the expected returns of different strategies in the current state, and the silence pattern probability distribution clarifies the specific type of silence the user is currently in. Concatenating these three parts forms an enhanced state vector. This enhanced state vector contains richer and more comprehensive information than the original state space; it not only considers the user's current state but also incorporates the expected effects of different intervention strategies and the specific silence pattern of the user, providing a more sufficient basis for subsequently selecting the optimal intervention strategy.

[0068] The resulting augmented state vector is input into the proximal policy optimization network (PPR) of the two-layer reinforcement learning engine. The PPR is an advanced reinforcement learning algorithm network capable of in-depth analysis and processing of the input augmented state vector. Through complex internal calculations and parameter adjustments, the PPR outputs a probability distribution of intervention strategies. This probability distribution represents the likelihood of adopting each intervention strategy in the current augmented state. After obtaining the probability distribution of intervention strategies, two methods can be used to select the optimal intervention action. One is a sampling strategy, which randomly selects an intervention strategy as the optimal intervention action based on the probability distribution. This method has a certain degree of randomness and can explore the effects of different intervention strategies. The other is a greedy strategy, which selects the intervention strategy with the highest probability value in the probability distribution as the optimal intervention action. This method focuses more on selecting the intervention strategy with the highest expected return in the current state. Either of these methods can dynamically select the optimal intervention strategy best suited to the current user's dormant state from the strategy library. This allows the AI ​​virtual character in the game to then use this strategy to provide personalized interactive engagement with the user, activating dormant users and reducing user churn.

[0069] For example, suppose the preceding steps identify a user who has fallen into a silent state due to excessive level difficulty. First, construct the user's state space, which includes information about the current level (level number, remaining lives, etc.), the user's actions within that level (attack frequency, movement trajectory, etc.), and the user's historical game data (past level completion history, skill level, etc.). Input the state space into a deep Q-network, which outputs the Q-values ​​of various intervention strategies, such as providing extra items, reducing level difficulty, or giving hints. A softmax classification layer is used to obtain the probability distribution of the silence mode, clarifying that the user falls into the category of silence due to level difficulty. Then, the state space, the Q-values ​​of each intervention strategy, and the silence mode probability distribution are concatenated into an enhanced state vector. This enhanced state vector is input into a proximal policy optimization network. If a greedy strategy is used, the network might choose to give hints, because for users with some skill who are stuck on level difficulty, providing hints can help them better cope with the challenge, and this strategy likely has the highest probability value in the probability distribution. This allows the network to select the optimal intervention strategy best suited for the user from the policy library.

[0070] For example, suppose another user is identified as becoming inactive due to monotonous gameplay. The user's state space is constructed by including the frequency and duration of their participation in various gameplay modules, their feedback on different modules, and their social interactions within the game. This information is then input into a deep Q-network to obtain the Q-values ​​of various intervention strategies, such as recommending new gameplay, hosting special events, or increasing gameplay rewards. Classification determines that this user belongs to the type of inactivity caused by monotonous gameplay. These are concatenated to form an enhanced state vector, which is then input into a proximal policy optimization network. If a sampling strategy is used, hosting special events might be randomly selected as an intervention strategy. This is because hosting special events can bring novelty and new gaming experiences to users inactive due to monotonous gameplay. This approach explores the effects of different intervention strategies on activating this user, allowing for further optimization of intervention strategy selection.

[0071] Furthermore, the two-layer reinforcement learning engine also includes a pluggable Transformer module designed for different game types, the pluggable Transformer module including a first attention module, a second attention module, a third attention module and a composite gameplay attention module; The first attention module is used to generate a first enhanced state vector that integrates team coordination information by reorganizing teammate-related features into a team coordination sequence and using a multi-head self-attention mechanism to calculate the coordination attention weight between the player and teammates. The second attention module is used to extract the behavioral feature sequences of players at key plot nodes and encode the temporal dependencies between plot nodes using a causal mask self-attention mechanism, thereby generating a second enhanced state vector containing plot context semantics. The third attention module is used to extract micro-behavioral features related to the flow state and mine emotional patterns using a local sensitive hash attention mechanism to generate a third enhanced state vector that represents the player's emotional experience. The composite gameplay attention module is used to obtain a fourth enhanced state vector by assigning learnable fusion weight parameters to each activated dedicated module and weighting and summing the outputs of each dedicated module according to the fusion weight parameters. The dedicated modules include a first attention module, a second attention module, and a third attention module.

[0072] In this embodiment, competitive games are used as an example. In competitive games, teamwork, control of key plot points, and player emotional state play a crucial role in the game experience and user retention.

[0073] In team-based competitive gameplay, such as Multiplayer Online Battle Arena (MOBA) or Team Deathmatch, teamwork is a key factor in achieving victory. The first attention module collects relevant feature data about teammates, including their chosen hero or character type, skill level, equipment, current health, and mana. These features are then reorganized into a teamwork sequence according to certain rules, such as the teammates' positioning in the game or their role function (e.g., tank, DPS, support). A multi-headed self-attention mechanism is then used to calculate the collaborative attention weights between players and teammates. This mechanism analyzes the collaborative relationship between players and teammates from multiple perspectives, such as the impact of teammates' skill timing on players, and the degree of coordination between teammates' movement trajectories and players. After calculating the influence weight of each teammate on the player, a first enhanced state vector integrating teamwork information is generated. This vector accurately reflects the player's collaborative performance within the team, such as the synergy between players and teammates and whether there are areas requiring coordination, providing crucial information for subsequent intervention strategies. For example, if a low collaborative attention weight between a player and teammates is detected, it may indicate a problem with teamwork, allowing for the selection of intervention strategies that include sending teamwork suggestions.

[0074] Competitive games also feature key plot points, such as important tournament finals and crucial character duels. The second attention module is responsible for extracting the behavioral feature sequences of players at these key plot points. For example, in tournament finals, it records the player's operation frequency, target selection, and skill usage order. A causal masking self-attention mechanism is used to encode the temporal dependencies between plot points. This mechanism ensures that when calculating attention weights, only information from the current and previous moments is considered, thus better capturing the causal relationships and temporal logic between plot points. For example, analyzing how a player's early actions affect the later match results generates a second enhanced state vector containing plot context semantics. This vector reflects the player's behavioral patterns and decision-making logic in key plot points, helping to understand the player's state in the current game situation. If a player hesitates at a key plot point, it may indicate a lack of understanding of the situation; in this case, intervention strategies providing situation analysis suggestions can be selected.

[0075] Competitive games often place significant pressure on players, and their emotional state greatly impacts game performance and retention. The third attention module extracts micro-behavioral features related to the flow state, such as player response time, error rate, and mouse movement speed. These micro-behavioral features reflect the player's emotional state; for example, a longer response time may indicate tension or confusion, while an increased error rate may indicate frustration. A Local Sensitive Hashing (LSH) attention mechanism is employed to mine emotional patterns. LSH can quickly identify behavioral patterns with similar emotional characteristics, and analyzing these patterns generates a third enhanced state vector representing the player's emotional experience. This vector intuitively reflects the player's emotional state in the game, providing a reference for selecting appropriate intervention strategies. If the player's emotional experience vector indicates a state of high tension, intervention strategies such as sending relaxation prompts or adjusting the game difficulty can be selected.

[0076] In comprehensive competitive gameplay, multiple factors are involved, including teamwork, key storylines, and player emotions. The composite gameplay attention module activates the first, second, and third attention modules. Learnable fusion weight parameters are assigned to each activated module, dynamically adjusting based on the game's specific circumstances and player performance. For example, in key team-based competitive storylines, if teamwork significantly impacts the outcome, the fusion weight parameter of the first attention module will be relatively large; if crucial storyline decisions determine the course of the match, the weight parameter of the second attention module will increase; and if player emotions significantly influence actions, the weight parameter of the third attention module will adjust accordingly. The outputs of each module are weighted and summed according to the fusion weight parameters to obtain a fourth enhanced state vector. This vector integrates information from teamwork, storyline context, and player emotions, providing a more comprehensive reflection of the player's state in comprehensive competitive gameplay. Based on the fourth enhanced state vector, the optimal intervention strategy is selected from the strategy library, such as simultaneously providing teamwork suggestions, situation analysis suggestions, and emotion regulation suggestions, to better activate dormant users.

[0077] Taking role-playing games as an example, these games emphasize plot development, character development, and player interaction with the game world. The pluggable Transformer module plays an important role in different gameplay styles.

[0078] In team-based dungeons in role-playing games, players need to cooperate with teammates to complete tasks. The first attention module collects teammate-related characteristics, such as character class, level, equipment attributes, and skill combinations. These characteristics are then reorganized into a team coordination sequence, for example, arranged according to the complementarity of character classes, with tanks in the front, DPS in the middle, and supports in the back. A multi-head self-attention mechanism is used to calculate the coordination attention weights between players and teammates, analyzing the level of support different class teammates provide to players in dungeons, such as whether the tank's taunt skill effectively protects the player, and whether the support's healing is timely. A first enhanced state vector, incorporating team coordination information, is generated, which helps understand the players' coordination in team dungeons. If problems are found in the coordination between players and teammates, such as the tank failing to attract monster aggro in time, resulting in the player taking excessive damage, an intervention strategy of sending a team coordination reminder can be selected.

[0079] The main storyline is a crucial factor in attracting players to role-playing games. The second attention module extracts behavioral feature sequences of players at key story nodes, such as the duration of dialogues with important NPCs, chosen quest branches, and exploration time within story scenes. A causal masking self-attention mechanism is employed to encode the temporal dependencies between story nodes, analyzing the player's decision-making logic and points of interest at different story stages. For example, if a player spends a considerable amount of time choosing a particular story branch, it may indicate a strong interest in that branch's plot. A second enhanced state vector containing contextual semantics is generated, which helps understand the player's comprehension and engagement with the story. If a player stops interacting at a key story node, it may indicate confusion about the story's development; in such cases, intervention strategies such as providing plot hints or guidance can be implemented.

[0080] Character development is one of the core gameplay elements of role-playing games, and the player's emotional state affects the progress and experience of character development. The third attention module extracts micro-behavioral features related to the flow state, such as the player's reaction when leveling up, the effort put into acquiring equipment, and the time spent on the development interface. A locality-sensitive hashing attention mechanism is used to mine emotional patterns and determine the player's satisfaction and enthusiasm for character development. For example, if a player acts quickly and frequently checks character attributes when leveling up, it may indicate enthusiasm for character development; if a player frequently abandons quests while acquiring equipment, it may indicate boredom with the current development method. A third enhanced state vector is generated to represent the player's emotional experience, reflecting their emotional state regarding character development. If a player is found to be in a low mood, intervention strategies such as providing new character development suggestions or adjusting the difficulty of development can be implemented.

[0081] The integrated role-playing gameplay incorporates multiple elements such as team dungeons, main storyline, and character development. The composite gameplay attention module activates the first, second, and third attention modules, assigning them learnable fusion weight parameters. These weight parameters are dynamically adjusted based on the game's progress and player performance. For example, during main storyline quests in team dungeons, if team coordination significantly impacts quest completion, the weight parameter of the first attention module increases; if storyline decisions are crucial, the weight parameter of the second attention module increases; and if the player experiences significant emotional fluctuations related to character development, the weight parameter of the third attention module is adjusted accordingly. The outputs of each dedicated module are weighted and summed according to the fusion weight parameters to obtain a fourth enhanced state vector. This vector integrates information from team coordination, storyline context, and player emotions, comprehensively reflecting the player's state within the integrated role-playing gameplay. Based on this fourth enhanced state vector, the optimal intervention strategy is selected, such as simultaneously providing team coordination suggestions, storyline guidance, and character development adjustment suggestions, to effectively activate dormant users.

[0082] Furthermore, the method also includes: The near-end policy optimization network is encapsulated as a policy prediction function, and a background distribution dataset is constructed by sampling historical interaction data; A sliding time window mechanism is used to divide the state-action pairs of the continuous decision steps of the two-layer reinforcement learning engine into multiple analysis windows; At each time point within each analysis window, the SHAP contribution matrix of each augmented state vector to the optimal intervention action is calculated using a model-independent KernelExplainer combined with the background distribution dataset. The SHAP contribution matrices of all analysis windows are summarized, and the contribution value of each enhanced state vector in each analysis window is mapped to color intensity using a two-dimensional heatmap to form a feature importance heatmap. The cumulative absolute contribution value of each enhanced state vector is calculated based on the feature importance heatmap to screen key features, and the abrupt change moment of the contribution value is identified by the change point detection algorithm.

[0083] In this embodiment, the near-end policy optimization network is encapsulated as a policy prediction function. This function takes the augmented state vector as input and outputs the probability distribution of each intervention action. For example, when a user is in a silent state after failing a level, the augmented state vector may contain features such as level difficulty, the user's historical completion time, and current health. The policy prediction function calculates the probability of providing hints, reducing difficulty, or skipping the level based on these features. Specifically, the parameters of the trained near-end policy optimization network (including its policy network and value network) are stored in a fixed format. For example, the weight matrices of each layer of neurons in the policy network (such as the weights from the input layer to the hidden layer and the weights from the hidden layer to the output layer) and the bias vector, along with the parameters of the value network (such as the linear layer parameters for state value estimation), are saved to a structured file (such as JSON or HDF5 format). Subsequently, a policy prediction function is constructed based on the fixed parameters. This function takes the augmented state vector (composed of the state space, Q-value, and the probability distribution of the silent mode) as input and calculates the probability distribution of the intervention policy through forward propagation.

[0084] In specific implementation, the policy prediction function performs the following operations in sequence: (1) Normalize the input augmented state vector (e.g., scale it to the [0,1] interval) to match the data distribution during training; (2) Input the normalized vector into the policy network, and generate the initial probability of each intervention policy by passing through a fully connected layer, an activation function (e.g., ReLU), and a Softmax layer in sequence; (3) Correct the initial probability by combining the output of the value network (state value estimation) (e.g., by using the advantage function for weighting), and finally output the corrected probability distribution of the intervention policy. For example, if the augmented state vector is [level difficulty = high, user payment tendency = medium, Q value_strategy push = 0.8], the policy prediction function may output a probability of 0.7 for "pushing level strategy" and a probability of 0.3 for "sending general encouragement message".

[0085] Simultaneously, a background distribution dataset is constructed by sampling from historical interaction data. Sampling must cover different game scenarios (such as payment decisions and activity participation), user types (high willingness to pay, low activity), and types of inactive states (entering silence due to confusion, dissatisfaction, or difficulty) to ensure the diversity and representativeness of the dataset. For example, interaction records of 1000 users in an inactive state from the past month can be randomly selected, and their enhanced state vectors and corresponding optimal intervention actions can be extracted as samples.

[0086] A sliding time window mechanism is employed to divide the consecutive decision steps of the two-layer reinforcement learning engine into multiple analysis windows. The window size is set to a fixed number of decision steps (e.g., one window for every 10 consecutive decision steps), and the sliding step size is half the window size (e.g., sliding 5 decision steps at a time) to balance analysis granularity and computational efficiency. For example, if the two-layer reinforcement learning engine generates 100 decision steps in one hour, it can be divided into 11 analysis windows according to the above parameters (window 1 covers decision steps 1-10, window 2 covers decision steps 6-15, and so on). Each window contains a set of consecutive state-action pairs for subsequent local contribution analysis.

[0087] Within each analysis window, the model-independent KernelExplainer method is used at each time step to calculate the SHAP contribution matrix of the augmented state vector to the optimal intervention action, in conjunction with the background distribution dataset. KernelExplainer quantifies the contribution of each feature to the final decision by simulating changes in model output when features are missing. For example, in the 7th decision step of analysis window 3, the augmented state vector contains three features: "Level Difficulty = High", "User's Historical Failure Count = 3", and "Current Remaining Time = 30 seconds". If the optimal intervention action is "Provide Level Strategy Hints", then the contribution value of each feature to selecting this action is calculated (e.g., Level Difficulty Contribution +0.4, Historical Failure Count Contribution +0.3, Remaining Time Contribution -0.1), forming the SHAP contribution matrix for that time step. This process is repeated until all time steps within the window are covered.

[0088] After summarizing the SHAP contribution matrices of all analysis windows, the importance of features is visualized using a two-dimensional heatmap. The horizontal axis represents the analysis window number, and the vertical axis represents the feature dimension of the enhanced state vector (such as level difficulty, historical failure count, etc.). The color intensity corresponds to the absolute value of the contribution. For example, if "level difficulty" has a generally high contribution value and a dark color in windows 5-8, it indicates that this feature has a significant impact on the choice of intervention action in these windows; if "current remaining time" has a light color in all windows, it indicates that its contribution is low. The heatmap can intuitively show the trend of feature importance over time, helping analysts quickly locate key features.

[0089] Based on the feature importance heatmap, the cumulative absolute contribution value of each enhanced state vector across all analysis windows is calculated (e.g., summing the absolute values ​​of a feature's contribution values ​​across all windows). Features with cumulative values ​​higher than a threshold are selected as key features. For example, if the threshold is set to the average of all feature cumulative values, and the cumulative value of "Level Difficulty" is 0.8, which is higher than the average of 0.5, it is determined to be a key feature. Simultaneously, a change point detection algorithm (such as the CUSUM algorithm) is used to identify abrupt changes in contribution values, i.e., detecting whether there are points in the contribution value sequence that significantly deviate from historical trends. For example, if the contribution value of "User's Historical Failure Count" suddenly jumps from 0.2 to 0.7 in window 10 and remains high in subsequent windows, then window 10 is determined to be a point of abrupt change, possibly corresponding to events such as game difficulty adjustments or changes in user behavior patterns.

[0090] This embodiment realizes interpretability analysis of the decision-making process of a two-layer reinforcement learning engine, which can accurately identify key features and moments of abrupt changes that affect the choice of intervention strategies, providing data support for optimizing model parameters and adjusting intervention strategies, thereby improving the activation effect of dormant users and reducing user churn rate.

[0091] In some embodiments, step S104 above, which involves enabling the AI ​​virtual character in the game to perform personalized multimodal interactive outreach in a non-intrusive manner according to the selected intervention strategy, specifically includes: Based on the selected intervention strategy, and combined with the current game scene context and user profile characteristics, an interaction intent is generated that includes the interaction subject, interaction modality, and interaction content template. Figure 3 tuple; The interactive content template is combined with the user's historical dialogue style characteristics, current silence mode type, and real-time in-game status to form prompt words, which are then input into a large language model that has been fine-tuned for the game domain. Personalized dialogue text that conforms to the game's world view and character design is output in real time through an autoregressive generation method. Based on the interaction subject, the preset appearance rules of each AI virtual character in the game world are obtained, and the spatiotemporal triggering conditions are determined by real-time monitoring of the game engine status. When an AI virtual character appears that meets the time-space triggering conditions and conforms to the preset appearance rules, personalized dialogue text is coordinated and dispatched to reach multiple interaction channels according to the interaction modality. The interaction channels include text interaction, voice interaction, action interaction and special effects interaction of the AI ​​virtual character.

[0092] In this embodiment, based on the intervention strategy selected by the two-layer reinforcement learning engine, the system first combines the current game scenario context (such as the user's current level, activity stage, and time point) with user profile features (such as payment level, historical activity level, and preferred game type) for analysis. For example, if a pre-defined strategy is "to provide strategy hints to high-paying users who are silent due to level difficulty," the system will extract scenario context such as the current level number, the user's historical completion time, and the reasons for the three most recent failures, as well as user profile features such as VIP level and the number of times they have paid in the past 30 days. This information is then mapped to interaction intent through a pre-defined rule engine or a lightweight decision tree model. Figure 3 A tuple contains the interaction subject, interaction modality, and interaction content template. This triple clarifies the core elements to be reached, providing a foundation for subsequent script generation and role allocation.

[0093] The interactive content template is combined with the user's historical dialogue style characteristics (such as the user's frequently used words and tone preferences when replying to NPCs), the current silence mode type (such as silence due to difficulty / silence due to monotonous gameplay), and the in-game real-time status (such as the user's remaining health and currently held items) to form a prompt word. For example, if the user frequently uses words such as "help" or "how to pass" in their historical dialogue, the current silence mode is "silence due to difficulty," and the real-time status is "remaining health 1," then the prompt word could be constructed as "Generate a strategy guide to help the user pass level 5, with a friendly and encouraging tone, including the suggestion 'use healing potions when health is low,' which fits the calm and composed persona of 'Mentor Irene.'" This prompt word is then input into a large language model fine-tuned for the game domain, such as a game-specific model based on a general LLM secondary training. The model outputs personalized dialogue text word by word through an autoregressive generation method that conforms to the game's world view (such as fantasy style) and the character's persona (such as the mentor's authority and care).

[0094] The system obtains the preset appearance rules of the interactive subject in the game world, including the range of levels allowed to appear (e.g., only appearing in levels 1-10), time restrictions (e.g., active from 10:00 to 22:00 daily), and triggering prerequisites (e.g., after the user fails 3 times consecutively). Simultaneously, the system monitors the game engine status in real time (e.g., the user's current level number, number of failures, and timestamp) to determine if the spatiotemporal triggering conditions are met. For example, if the user is currently on level 5 and has failed 2 times consecutively, and the time is 15:00, while the preset rule for "Mentor Irene" is "appears in levels 1-10, daily from 10:00 to 22:00, and after ≥3 consecutive failures," then the spatiotemporal triggering conditions are not fully met (insufficient number of failures). If the user fails 3 times consecutively, the conditions are met, and the system proceeds to the next step of character scheduling.

[0095] When an AI virtual character appears, meeting the time-space trigger conditions and pre-defined appearance rules, the system coordinates personalized dialogue text across multiple interaction channels based on interaction modalities (such as text + voice + action). Specifically, the text interaction channel displays the generated dialogue text in the NPC's dialogue box; the voice interaction channel uses speech synthesis technology to convert the text into speech that matches the character's personality and plays it through the in-game audio system; the action interaction channel controls the NPC to perform pre-defined actions; and the special effects interaction channel adds visual effects around the NPC to attract the user's attention. For example, when "Mentor Irene" appears, a dialogue box pops up on the game screen displaying strategy text, while the NPC waves and explains in voice, and a glowing arrow is generated on the ground pointing to the potion's location. This multi-modal collaboration enhances the reach and increases the user's willingness to re-engage in the game.

[0096] This embodiment achieves fully automated execution of the entire process from intervention strategies to personalized multimodal interactions, ensuring both the accuracy and non-intrusiveness of the outreach, while also enhancing the perceived value to users through multimodal collaboration, thereby effectively activating dormant users and reducing churn rate.

[0097] In some embodiments, step S105 above, which involves collecting real-time feedback data from the user regarding the interaction and continuously optimizing the model parameters of the graph neural network and the two-layer reinforcement learning engine through online learning, specifically includes: During the interactive interaction process of the AI ​​virtual character, multi-dimensional feedback data of the user is collected in real time. The multi-dimensional feedback data includes explicit behavioral feedback, implicit behavioral feedback and physiological signal feedback. Multi-dimensional feedback data is associated with and encapsulated into structured feedback event tuples with access context information and stored in a multi-task experience replay pool; A periodic online update mechanism is set up for graph neural networks. Based on the multi-task experience replay pool, the importance matrix of historical parameters is calculated using the elastic weight solidification technique. A regularization term is introduced into the cross-entropy classification loss to constrain the magnitude of changes to historical important parameters. A semi-online update strategy is adopted for the two-layer reinforcement learning engine. By setting different update frequencies for the deep Q network and the proximal policy optimization network, and combining a pruning mechanism and a dynamically adjusted ε-greedy exploration rate, batch data is sampled from the multi-task experience replay pool to update the parameters of the deep Q network and the proximal policy optimization network respectively.

[0098] In this embodiment, during the personalized multimodal interaction of the AI ​​virtual character, the system collects multi-dimensional feedback data from the user through collaboration between the game client and the server. Explicit behavioral feedback includes the user's direct operational response to the interactive content, such as clicking the "Accept Help" button in the NPC dialogue box, closing the prompt window, or initiating a dialogue. Implicit behavioral feedback indirectly infers the user's attitude by analyzing changes in their game behavior, such as a shorter time interval between continuing to operate after receiving the interaction, an increased level completion rate, and an increased frequency of social interaction. Physiological signal feedback obtains the user's physiological reactions through wearable devices or game peripherals (such as controller vibration feedback or eye-tracking devices), such as changes in heart rate, gaze duration, and fluctuations in operational intensity. For example, when the AI ​​virtual character pushes a level guide, the system records whether the user clicks "View Details" (explicit feedback) and monitors whether the number of subsequent attempts to complete the level decreases (implicit feedback). If the user is wearing a smart bracelet, it can also collect whether their heart rate drops from a tense state to a stable state, such as a decrease from 90 beats / min to 70 beats / min (physiological signal feedback).

[0099] The collected multi-dimensional feedback data is associated with and encapsulated into structured feedback event tuples based on the access context information. The access context information includes the timestamp of the interaction, the game scene (such as level number, activity type), the type of AI virtual character (such as mentor NPC or friend NPC), the access modality (text / voice / action), and the intervention strategy number (such as "high-paying user level strategy guidance"). For example, if a user receives a strategy guide push by "Mentor Eileen" via text + voice at 20:30 when the user is silent due to the difficulty of level 5, and the user clicks the "View Details" button, and the subsequent level completion time is shortened from 10 minutes to 5 minutes, then the encapsulated event tuple is: [Timestamp: 20xx-xx-xx 20:30:00, Scene: Level 5 - Difficulty Silence, Character: Mentor Eileen, Modality: Text + Voice, Strategy Number: S001, Explicit Feedback: Click to View Details, Implicit Feedback: Completion Time -50%, Physiological Signal: Heart Rate -20 beats / min]. All event tuples are stored in the multi-task experience replay pool in chronological order. This replay pool supports retrieval by category, role, strategy, and other dimensions, providing a data foundation for subsequent model optimization.

[0100] A periodic online update mechanism (e.g., every 30 minutes) is implemented for the graph neural network to balance model real-time performance and stability. During updates, the system extracts feedback event tuples relevant to the current batch of users from the multi-task experience replay pool and uses an elastic weighting solidification technique to calculate the importance matrix of historical parameters. This technique assigns weights to each parameter by analyzing its contribution to historical tasks (e.g., the impact of a node embedding parameter on the accuracy of identifying silent states), with higher weights indicating greater importance to historical tasks. During the update process, a regularization term is introduced into the cross-entropy classification loss function to constrain the magnitude of changes to historically important parameters. For example, if a node embedding parameter contributed 80% to the accuracy of identifying the "silent due to difficulty" state in past tasks, the adjustment of this parameter during updates is limited to within ±10% of the original value, preventing the model from forgetting key historical knowledge due to new data impact. In this way, the graph neural network can both absorb new user feedback data and retain its ability to identify historical silent patterns.

[0101] A semi-online update strategy is adopted for the two-layer reinforcement learning engine, balancing strategy exploration and utilization through differentiated update frequencies. The bottom-layer Deep Q-Network (DQN) is responsible for classifying silent patterns, with a high parameter update frequency (e.g., every 10 minutes) to quickly adapt to newly emerging silent patterns. The upper-layer Proximal Policy Optimization Network (PPO) is responsible for intervening in policy selection, with a low parameter update frequency (e.g., every hour) to ensure the stability of policy selection. During updates, the system samples batch data from the multi-task experience replay pool and employs a clipped DQN mechanism to limit the magnitude of each update (e.g., no more than 5% of the current value) to avoid policy oscillations. A dynamically adjusted ε-greedy exploration rate is introduced for the PPO. The initial ε value is set to 0.3 (30% probability of randomly selecting a policy). As feedback data accumulates, if a policy receives positive feedback for 5 consecutive times (e.g., game time increases after the user accepts the touch), the ε value gradually decreases to 0.1 (10% probability of randomly selecting), and vice versa, to enhance the utilization of effective policies. By using semi-online updates, the two-layer reinforcement learning engine can optimize classification and policy selection in a timely manner while avoiding performance fluctuations caused by frequent updates.

[0102] This embodiment realizes a closed-loop process from multi-dimensional feedback collection to dynamic optimization of model parameters, enabling graph neural networks and two-layer reinforcement learning engines to continuously absorb real-time user feedback, improve the accuracy of silent state recognition and the effectiveness of intervention strategies, and ultimately reduce user churn rate.

[0103] Reference Figure 2 An embodiment of the present invention provides a game dormant user activation system 2 based on artificial intelligence, wherein system 2 specifically includes: The data acquisition module 201 is used to collect multi-dimensional behavior flow data of users in real time, and to construct a spatiotemporal graph structure containing user operation sequences, virtual location trajectories and social interactions through the multi-dimensional behavior flow data; The state recognition module 202 is used to learn node embedding of spatiotemporal graph structure using graph neural network to capture spatiotemporal correlation features of different silence modes, and dynamically identify the current silence state of the target user by combining a temporal convolution module with adjustable receptive field. The silence classification module 203 is used to input the identified current silence state into the two-layer reinforcement learning engine. The silence pattern is accurately classified through the bottom-layer deep Q network, and the optimal intervention strategy is dynamically selected from the policy library by combining the classification results through the upper-layer proximal policy optimization network. The interactive outreach module 204 is used to enable the AI ​​virtual character in the game to perform personalized multimodal interactive outreach in a non-intrusive manner according to the selected intervention strategy; The parameter optimization module 205 is used to collect real-time feedback data from users regarding this interaction and to continuously optimize the model parameters of the graph neural network and the two-layer reinforcement learning engine through online learning.

[0104] It is understandable that, such as Figure 1 The content shown in the AI-based game silent user activation method embodiment is applicable to this AI-based game silent user activation system embodiment. The specific functions implemented by this AI-based game silent user activation system embodiment are the same as those shown in the example. Figure 1 The illustrated method for activating dormant game users based on artificial intelligence is the same as the one shown, and achieves the same beneficial effects. Figure 1 The beneficial effects achieved by the AI-based game silent user activation method embodiment shown are the same.

[0105] It should be noted that the information interaction and execution process between the above systems are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0106] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0107] Reference Figure 3 The present invention also provides a computer device 3, including: a memory 302 and a processor 301, and a computer program 303 stored on the memory 302. When the computer program 303 is executed on the processor 301, it implements the game silent user activation method based on artificial intelligence as described in any of the above methods.

[0108] The computer device 3 may be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device 3 may include, but is not limited to, a processor 301 and a memory 302. Those skilled in the art will understand that... Figure 3 The computer device 3 is merely an example and does not constitute a limitation on the computer device 3. It may include more or fewer components than shown in the figure, or combine certain components, or different components, such as input / output devices, network access devices, etc.

[0109] The processor 301 can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), 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.

[0110] In some embodiments, the memory 302 may be an internal storage unit of the computer device 3, such as a hard disk or memory of the computer device 3. In other embodiments, the memory 302 may be an external storage device of the computer device 3, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 3. Furthermore, the memory 302 may include both internal and external storage units of the computer device 3. The memory 302 is used to store the operating system, applications, boot loader, data, and other programs, such as the program code of the computer program. The memory 302 can also be used to temporarily store data that has been output or will be output.

[0111] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the AI-based method for activating silent game users as described in any of the above methods.

[0112] In this embodiment, if the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0113] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0114] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0115] In the embodiments disclosed in this application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0116] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

Claims

1. A method for activating dormant game users based on artificial intelligence, characterized in that, The method specifically includes: Real-time collection of multidimensional user behavior flow data; and construction of a spatiotemporal graph structure containing user operation sequences, virtual location trajectories, and social interactions using the multidimensional behavior flow data. Graph neural networks are used to learn node embeddings in spatiotemporal graph structures to capture the spatiotemporal correlation features of different silence modes, and a temporal convolution module with adjustable receptive field is combined to dynamically identify the current silence state of the target user. The identified current silence state is input into a two-layer reinforcement learning engine. The silence pattern is accurately classified through the deep Q-network at the bottom layer, and the optimal intervention strategy is dynamically selected from the policy library by the proximal policy optimization network at the top layer in combination with the classification results. Based on the selected intervention strategy, the AI ​​virtual characters in the game can perform personalized multimodal interactive outreach in a non-intrusive manner; Collect real-time user feedback data on this interaction, and continuously optimize the model parameters of the graph neural network and the two-layer reinforcement learning engine through online learning.

2. The method according to claim 1, characterized in that, The construction of a spatiotemporal graph structure containing user operation sequences, virtual location trajectories, and social interactions using multidimensional behavior flow data specifically includes: Extract user operation sequences to characterize user operation logic, user location trajectories to reflect user movement patterns in the virtual environment, and social interaction data to reflect user interaction patterns with game characters from multidimensional behavior flow data. The nodes and edges of the spatiotemporal graph structure are set according to the user operation sequence, user location trajectory and social interaction data. The nodes include user nodes with user identifier as the core and function nodes with key function points in the game as the core. The edges are the spatiotemporal interaction relationships between nodes. A sliding time window is used to aggregate all active nodes and edges that meet the conditions within the current time window, generating a static spatiotemporal graph snapshot within the current time window; All time window static spatiotemporal graph snapshots are stored in a structured manner in the form of adjacency matrix and feature matrix to form a spatiotemporal graph structure.

3. The method according to claim 1, characterized in that, The method utilizes graph neural networks to learn node embeddings in the spatiotemporal graph structure to capture the spatiotemporal correlation features of different silence modes, and combines this with a temporal convolutional module with an adjustable receptive field to dynamically identify the current silence state of the target user. Specifically, this includes: The spatiotemporal graph structure is input into the graph neural network, and the feature representation of the user node is updated by aggregating the feature information of the neighboring nodes to obtain the user node embedding vector that integrates spatial context information in each time slice. The embedding vectors of user nodes from multiple consecutive time slices are concatenated in chronological order to construct a temporal feature sequence that reflects the dynamic changes in user status. The temporal feature sequence is input into a temporal convolutional module with adjustable receptive field. By designing convolutional kernels with different dilation rates, the user's behavior patterns at different time granularities are dynamically captured, and the output features of multiple branches are fused to generate multi-scale fused features. Multi-scale fused features are input into a classifier containing a fully connected layer. By calculating the probability distribution of a preset silence pattern, the current silence state of the target user is identified.

4. The method according to claim 3, characterized in that, The process of inputting the spatiotemporal graph structure into a graph neural network and updating the feature representation of the user node by aggregating the feature information of neighboring nodes to obtain the user node embedding vector fused with spatial context information for each time slice specifically includes: The spatiotemporal graph structure is input into the graph neural network. For each central user node and its neighboring nodes in the spatiotemporal graph structure, the original attention coefficient between them is calculated through the graph attention mechanism. The original attention coefficient is used to quantify the importance of the neighboring nodes to the central user node. The original attention coefficients of all neighboring nodes are normalized using softmax to obtain the normalized attention weights corresponding to each neighboring node. The normalized attention weights are used to reflect the degree of contribution of each neighboring node to the user node's behavioral intent at the current moment. The features of neighboring nodes are weighted and aggregated based on normalized attention weights, and the aggregated neighborhood features are fused with the features of the central user node itself. After passing through a nonlinear activation function, the updated user node feature representation of the current layer is obtained. By stacking multi-layer graph attention networks and combining them with a multi-head attention mechanism, the user node feature representation of each central user node is aggregated with spatial context information within the range of multiple-order neighbors, generating a user node embedding vector that integrates the game environment context for each time slice.

5. The method according to claim 3, characterized in that, The process involves inputting temporal feature sequences into a temporal convolutional module with adjustable receptive fields. By designing convolutional kernels with different dilation rates, it dynamically captures user behavior patterns at different temporal granularities and fuses the output features of multiple branches to generate multi-scale fused features. Specifically, this includes: The temporal feature sequence is input into multiple dilated convolutional branches in parallel, and a different dilation rate is configured for each dilated convolutional branch to achieve feature extraction of the input sequence at different temporal granularities. The feature maps output by each dilated convolution branch are unified and activated in the channel dimension to obtain multiple intermediate feature maps with the same number of channels. Multiple intermediate feature maps are stitched together along the channel dimension and input into the fusion layer for cross-scale information interaction, generating multi-scale fusion features that integrate local details and global trends.

6. The method according to claim 1, characterized in that, The identified current silence state is input into a two-layer reinforcement learning engine. The underlying deep Q-network accurately classifies the silence pattern, and the upper-layer proximal policy optimization network dynamically selects the optimal intervention policy from the policy base based on the classification results. Specifically, this includes: The identified current silence state is constructed as a state space, which is then input into the deep Q-network at the bottom layer of the two-layer reinforcement learning engine. The Q-values ​​of each intervention strategy are output through the multi-layer fully connected layers and competitive architecture of the deep Q-network, and the silence mode probability distribution is obtained by accurate classification through the softmax classification layer. The state space, the Q-values ​​of each intervention strategy, and the probability distribution of the silence mode are concatenated to form an enhanced state vector; The enhanced state vector is input into the proximal policy optimization network of the upper layer of the two-layer reinforcement learning engine. The proximal policy optimization network outputs the probability distribution of the intervention policy and selects the optimal intervention action through sampling or greedy policy.

7. The method according to claim 6, characterized in that, The two-layer reinforcement learning engine also includes a pluggable Transformer module designed for different game types. The pluggable Transformer module includes a first attention module, a second attention module, a third attention module, and a composite gameplay attention module. The first attention module is used to generate a first enhanced state vector that integrates team coordination information by reorganizing teammate-related features into a team coordination sequence and using a multi-head self-attention mechanism to calculate the coordination attention weight between the player and teammates. The second attention module is used to extract the behavioral feature sequences of players at key plot nodes and encode the temporal dependencies between plot nodes using a causal mask self-attention mechanism, thereby generating a second enhanced state vector containing plot context semantics. The third attention module is used to extract micro-behavioral features related to the flow state and mine emotional patterns using a local sensitive hash attention mechanism to generate a third enhanced state vector that represents the player's emotional experience. The composite gameplay attention module is used to obtain a fourth enhanced state vector by assigning learnable fusion weight parameters to each activated dedicated module and weighting and summing the outputs of each dedicated module according to the fusion weight parameters. The dedicated modules include a first attention module, a second attention module, and a third attention module.

8. The method according to claim 6, characterized in that, The method further includes: The near-end policy optimization network is encapsulated as a policy prediction function, and a background distribution dataset is constructed by sampling historical interaction data; A sliding time window mechanism is used to divide the state-action pairs of the continuous decision steps of the two-layer reinforcement learning engine into multiple analysis windows; At each time point within each analysis window, the SHAP contribution matrix of each augmented state vector to the optimal intervention action is calculated using a model-independent KernelExplainer combined with the background distribution dataset. The SHAP contribution matrices of all analysis windows are summarized, and the contribution value of each enhanced state vector in each analysis window is mapped to color intensity using a two-dimensional heatmap to form a feature importance heatmap. The cumulative absolute contribution value of each enhanced state vector is calculated based on the feature importance heatmap to screen key features, and the abrupt change moment of the contribution value is identified by the change point detection algorithm.

9. The method according to any one of claims 1 to 8, characterized in that, The aforementioned method, based on the selected intervention strategy, enables the AI ​​virtual characters within the game to perform personalized multimodal interactive outreach in a non-intrusive manner, specifically including: Based on the selected intervention strategy, the current game scene context and user profile characteristics are analyzed to generate an interaction intent triplet containing the interaction subject, interaction modality, and interaction content template. The interactive content template is combined with the user's historical dialogue style characteristics, current silence mode type, and real-time in-game status to form prompt words, which are then input into a large language model that has been fine-tuned for the game domain. Personalized dialogue text that conforms to the game's world view and character design is output in real time through an autoregressive generation method. Based on the interaction subject, the preset appearance rules of each AI virtual character in the game world are obtained, and the spatiotemporal triggering conditions are determined by real-time monitoring of the game engine status. When an AI virtual character appears that meets the time-space triggering conditions and conforms to the preset appearance rules, personalized dialogue text is coordinated and dispatched to reach multiple interaction channels according to the interaction modality. The interaction channels include text interaction, voice interaction, action interaction and special effects interaction of the AI ​​virtual character.

10. A game dormant user activation system based on artificial intelligence, characterized in that, The system specifically includes: The data acquisition module is used to collect multi-dimensional behavioral flow data of users in real time, and to construct a spatiotemporal graph structure that includes user operation sequences, virtual location trajectories and social interactions through the multi-dimensional behavioral flow data; The state recognition module is used to learn node embeddings of spatiotemporal graph structures using graph neural networks to capture the spatiotemporal correlation features of different silence modes, and dynamically identify the current silence state of the target user by combining a temporal convolution module with an adjustable receptive field. The silence classification module is used to input the identified current silence state into the two-layer reinforcement learning engine. The silence pattern is accurately classified through the bottom-layer deep Q-network, and the optimal intervention strategy is dynamically selected from the policy library by the upper-layer proximal policy optimization network in combination with the classification results. The interactive outreach module is used to enable AI virtual characters in the game to perform personalized multimodal interactive outreach in a non-intrusive manner, based on the selected intervention strategy. The parameter optimization module is used to collect real-time feedback data from users regarding this interaction and continuously optimize the model parameters of the graph neural network and the two-layer reinforcement learning engine through online learning.

Images