Account grouping method and device, storage medium and electronic device

Abstract

The application discloses a kind of account grouping method and device, storage medium and electronic equipment.Therein, the method comprises: grouping account based on grouping strategy, obtain multiple account team sets, predict multiple account team sets by target secondary retention prediction model, determine the strategy secondary retention probability corresponding to each account team set, use genetic algorithm according to team secondary retention probability to multiple account team sets are iterated N rounds, determine target account team set based on the strategy secondary retention probability corresponding to each account team set generated by N rounds of iteration, realize the positive correlation reinforcement of team strategy and player retention, solve the core pain point that traditional artificial rule system is high in hysteresis and large in trial and error cost, provide quantifiable, iterative optimization paradigm for the intelligent matching system of multiplayer online game, the technical problem that the application solves is that account grouping efficiency and accuracy are low, which affects account retention rate.

Description

Technical Field

[0001] This application relates to the field of computers, and more specifically, to a method and apparatus for grouping accounts, a storage medium, and an electronic device. Background Technology

[0002] In multiplayer games, matchmaking mechanisms need to assign players to the same game and then further divide them into multiple teams. Different team combinations directly affect the player experience and subsequent activity levels. Related technologies primarily rely on manually preset rules for player team assignment, such as static grouping based on player profile tags: pre-defined tags for combat style, tactical preferences, etc., are used to divide teams. This method has significant limitations: First, manually formulating rules requires substantial resources for effectiveness verification and priority adjustment, especially in scenarios with multiple rules, where adjustment costs increase exponentially. Second, static rules cannot dynamically adapt to real-time player characteristics. When player behavior patterns or the game environment change, manual rules need repeated iterations and updates, leading to strategy lag. Furthermore, related technologies evaluate team assignment effectiveness through indirect indicators (such as tag matching), which lack a direct correlation with the core goal of player retention. It is difficult to verify the true impact of team assignment strategies on long-term activity levels. Therefore, related technologies suffer from low efficiency and accuracy in account grouping, resulting in technical problems that affect account retention rates.

[0003] There is currently no effective solution to the above problems. Summary of the Invention

[0004] This application provides a method and apparatus for grouping accounts, a storage medium, and an electronic device to at least solve the technical problem of low efficiency and accuracy in account grouping, which affects account retention rate.

[0005] According to one aspect of the embodiments of this application, a method for grouping accounts is provided, comprising: grouping accounts based on a grouping strategy to obtain multiple account team sets, wherein an account team in one of the multiple account team sets includes multiple accounts determined according to the grouping strategy; predicting the multiple account team sets using a target retention prediction model to determine a strategy retention probability corresponding to each account team set, wherein the target retention prediction model is used to predict the team retention probability of each account team in each account team set, and to determine the strategy retention probability based on the team retention probability, the target retention prediction model including using sample accounts... The model is obtained by training an initial second-time retention prediction model using a set of teams and sample labels, where the sample labels indicate the retention status of the sample account team set. A genetic algorithm is used to iterate through the multiple account team sets N times according to the team second-time retention probability. In the i-th iteration, the strategy second-time retention probability corresponding to each account team set generated in the i-th iteration is determined by the target second-time retention prediction model, where N and i are positive integers, and i ≤ N. Based on the strategy second-time retention probability corresponding to each account team set generated in the N-th iteration, a target account team set is determined, where the target account team set represents the account team set whose strategy second-time retention probability satisfies a preset condition.

[0006] According to another aspect of the embodiments of this application, an account grouping device is also provided, comprising: a grouping module, configured to group accounts based on a grouping strategy to obtain multiple account team sets, wherein an account team in one of the multiple account team sets includes multiple accounts determined according to the grouping strategy; and a prediction module, configured to predict the multiple account team sets using a target retention prediction model to determine the strategy retention probability corresponding to each account team set, wherein the target retention prediction model is used to predict the team retention probability of each account team in each account team set, and determine the strategy retention probability based on the team retention probability, the target retention prediction model including using sample accounts The model is obtained by training an initial second-time retention prediction model using a set of teams and sample labels, where the sample labels indicate the retention status of the sample account team set; the processing module is used to perform N rounds of iteration on the multiple account team sets according to the team second-time retention probability using a genetic algorithm, wherein, in the i-th round of iteration, the strategy second-time retention probability corresponding to each account team set generated in the i-th round of iteration is determined by the target second-time retention prediction model, where N and i are positive integers, i≤N; the generation module is used to determine a target account team set based on the strategy second-time retention probability corresponding to each account team set generated in the N rounds of iteration, wherein the target account team set represents the account team set whose strategy second-time retention probability satisfies a preset condition.

[0007] In an exemplary embodiment, the apparatus is configured to predict the multiple account team sets using a target retention prediction model in the following manner to determine the strategy retention probability corresponding to each account team set: predicting each account in each account team in the multiple account team sets using the target retention prediction model to determine the account retention probability corresponding to each account; determining the team retention probability corresponding to each account team based on the account retention probability; and determining the strategy retention probability corresponding to each account team set based on the team retention probability.

[0008] In an exemplary embodiment, the device is configured to predict the second-time retention probability of each account in each account team of the plurality of account teams by means of the target second-time retention prediction model in the following manner: obtaining the account features corresponding to each account in each account team; and determining the second-time retention probability of each account based on the account features.

[0009] In an exemplary embodiment, the device is configured to obtain account features corresponding to each account in each account team in the following manner: obtaining sequence features corresponding to each account, wherein the sequence features are used to represent the game data of the account within a preset time period; obtaining non-sequence features corresponding to each account, wherein the non-sequence features are used to represent the account parameters of the account; and determining the account features based on the sequence features and the non-sequence features.

[0010] In an exemplary embodiment, the device is configured to obtain the sequence features corresponding to each account in the following manner: obtain a feature matrix of size n·k generated by each account in the first n games within the preset time period, wherein each account in each game corresponds to k game features, and n and k are both positive integers; use a first network model to perform weighted fusion on the feature matrix to determine the sequence features, wherein the target retention prediction model includes the first network model.

[0011] In an exemplary embodiment, the device is configured to determine the account retention probability corresponding to each account based on the account features in the following manner: concatenating the sequence features and the non-sequence features to obtain the account features corresponding to each account; using a second network model to perform feature cross-referencing on the account features in the same account group to determine the account retention probability corresponding to each account, wherein the target retention prediction model includes the second network model.

[0012] In an exemplary embodiment, the apparatus is further configured to: determine the sample account team set and the sample label based on historical data; predict each sample account in each sample account team in the sample account team set using the initial second-time retention prediction model, and determine the sample account second-time retention probability corresponding to each sample account; determine the sample team second-time retention probability corresponding to each sample account team based on the sample account second-time retention probability; determine the sample strategy second-time retention probability corresponding to the sample account team set based on the sample team second-time retention probability; calculate a loss function based on the sample strategy second-time retention probability and the sample label, and adjust the model parameters of the initial second-time retention prediction model based on the loss function until the loss function meets a preset loss condition, and determine the initial second-time retention prediction model as the target second-time retention prediction model.

[0013] In an exemplary embodiment, the apparatus is configured to predict each sample account in each sample account team of the sample account team set using the initial retention prediction model in the following manner to determine the retention probability of the sample account corresponding to each sample account: predicting each sample account in each sample account team of the sample account team set using the initial retention prediction model to determine the initial retention probability of the sample account corresponding to each sample account; calibrating the initial retention probability of the sample account using a probability calibration function to determine the retention probability of the sample account corresponding to each sample account.

[0014] In an exemplary embodiment, the apparatus is used to perform N rounds of iteration on the plurality of account team sets using a genetic algorithm according to the team second retention probability in the following manner: during the i-th round of iteration, the team second retention probability corresponding to each account team in each account team set is obtained; the first account team and the second account team in each account team set whose team second retention probability satisfies the first sorting condition are swapped to determine the account team set used in the (i+1)-th round of iteration.

[0015] In an exemplary embodiment, the apparatus is used to determine the set of account teams used in the (i+1)th iteration by swapping elements of a first account team and a second account team in each account team set whose retention probabilities satisfy a first sorting condition: obtaining a first account in the first account team in each account team set whose retention probabilities satisfy a second sorting condition and a second account in the second account team in each account team set whose retention probabilities satisfy the second sorting condition; swapping the account teams to which the first account and the second account belong in each account team set to determine a third account team and a fourth account team; and using the third account team and the fourth account team to replace the first account team and the second account team, respectively, to determine the set of account teams used in the (i+1)th iteration.

[0016] In an exemplary embodiment, the device is used to group accounts based on a grouping strategy to obtain multiple sets of account teams: generating a first set of account teams using a random allocation rule; generating a second set of account teams using a tag matching rule; merging the first set of account teams and the second set of account teams to generate the multiple sets of account teams; the device is used to perform N rounds of iteration on the multiple sets of account teams using a genetic algorithm according to the team retention probability: during the i-th iteration, obtaining the team retention probability corresponding to each account team; sorting the account teams according to the team retention probability, and selecting the first P account teams and the last Q account teams, where P and Q are both positive integers; swapping the elements of the target account teams whose team retention probability satisfies the third sorting condition in the last Q account teams to determine the updated last Q account teams; combining the updated last Q account teams with the first P account teams to determine the set of account teams used in the (i+1)-th iteration, wherein the first P account teams and the last Q account teams belong to the same strategy.

[0017] In an exemplary embodiment, the apparatus is configured to determine a target account team set based on the strategy retention probability corresponding to each account team set generated by N iterations in the following manner: when the number of iterations reaches N, the account team set corresponding to the maximum value of the strategy retention probability in the Nth iteration is selected as the target account team set; when the number of iterations has not reached N, and the increase of the maximum value of the strategy retention probability in M ​​consecutive iterations is less than or equal to a first threshold, the iteration is terminated early and the account team set corresponding to the maximum value of the strategy retention probability in the current iteration is selected as the target account team set, where M is a positive integer less than N; when the number of iterations has not reached N, and the maximum value of the strategy retention probability is greater than or equal to a second threshold, the iteration is terminated early and the account team set corresponding to the maximum value of the strategy retention probability in the current iteration is selected as the target account team set.

[0018] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, wherein a computer program is stored in the computer program, and the computer program is configured to execute the aforementioned account grouping method when it is run.

[0019] According to another aspect of the embodiments of this application, a computer program product or computer program is provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the account grouping method described above.

[0020] According to another aspect of the embodiments of this application, an electronic device is also provided, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to execute the account grouping method described above through the computer program.

[0021] Through the embodiments of this application, a dynamic grouping mechanism based on multi-strategy initialization is adopted. By generating a set of account teams containing different team-up rules as the initial strategy pool for genetic algorithm iteration, a diverse search starting point is provided for subsequent optimization, achieving the goal of overcoming the limitations of manual rules and realizing the technical effect of automatic exploration of the strategy space. By constructing a deep neural network model that integrates sequential and non-sequential features, and using historical player team features and real retention labels to train the target second-day retention prediction model, the probability value output by the model is calibrated to achieve the goal of accurately quantifying the correlation between team-up strategies and retention, thereby realizing the direct coupling of strategy effect evaluation indicators and core business objectives. By introducing the evolutionary mechanism of the genetic algorithm, new strategies are dynamically generated using crossover and mutation operators. Through multiple rounds of iteration, a set of account teams with high strategy second-day retention probabilities is selected, achieving the goal of targeted optimization of the strategy space, thereby realizing the technical effect of efficiently converging to the Pareto optimal solution in an exponential combination of possibilities. By establishing a termination condition judgment mechanism based on strategy second-day retention probability, using a preset probability threshold or iteration round threshold as the stopping standard, and through the automatic selection mechanism of the globally optimal strategy, the goal of dynamically adapting to real-time player characteristics is achieved, thereby realizing a closed-loop optimization system without manual intervention. Ultimately, through the synergistic effect of the above-mentioned technical paths, a complete technical closed loop from strategy generation and effect evaluation to iterative optimization was formed, realizing the positive correlation between team strategy and player retention. This solved the core pain points of traditional manual rule systems, such as high lag and high trial-and-error costs, and provided a quantifiable and iterative optimization paradigm for intelligent matchmaking systems in multiplayer online games. In turn, it solved the technical problem of low efficiency and accuracy of account grouping, which affected account retention. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0023] Figure 1 This is a schematic diagram illustrating an application environment of an optional account grouping method according to an embodiment of this application;

[0024] Figure 2 This is a flowchart illustrating an optional account grouping method according to an embodiment of this application;

[0025] Figure 3 This is a schematic diagram of an optional account grouping method according to an embodiment of this application;

[0026] Figure 4 This is a schematic diagram of another optional account grouping method according to an embodiment of this application;

[0027] Figure 5 This is a schematic diagram of another optional account grouping method according to an embodiment of this application;

[0028] Figure 6 This is a schematic diagram of another optional account grouping method according to an embodiment of this application;

[0029] Figure 7 This is a schematic diagram of another optional account grouping method according to an embodiment of this application;

[0030] Figure 8 This is a schematic diagram of another optional account grouping method according to an embodiment of this application;

[0031] Figure 9 This is a schematic diagram of another optional account grouping method according to an embodiment of this application;

[0032] Figure 10 This is a schematic diagram of another optional account grouping method according to an embodiment of this application;

[0033] Figure 11 This is a schematic diagram of another optional account grouping method according to an embodiment of this application;

[0034] Figure 12 This is a schematic diagram of another optional account grouping method according to an embodiment of this application;

[0035] Figure 13 This is a schematic diagram of an optional account grouping device according to an embodiment of this application;

[0036] Figure 14 This is a schematic diagram of the structure of an optional account grouping product according to an embodiment of this application;

[0037] Figure 15 This is a schematic diagram of the structure of an optional electronic device according to an embodiment of this application. Detailed Implementation

[0038] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0039] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0040] To better understand the technical solutions provided in the embodiments of this application, the key terms involved in the embodiments of this application will be introduced first:

[0041] Genetic Algorithm (GA) is an intelligent optimization algorithm that simulates the process of biological evolution. By simulating mechanisms such as "natural selection" and "genetic mutation," it seeks optimal solutions to complex problems. It is particularly suitable for high-dimensional, nonlinear, and multi-objective optimization problems that are difficult to solve using traditional mathematical methods.

[0042] The present application will be described below with reference to embodiments:

[0043] According to one aspect of the embodiments of this application, a method for grouping accounts is provided. Optionally, in this embodiment, the above-mentioned account grouping method can be applied to, for example... Figure 1 The hardware environment shown consists of server 101 and terminal device 103. For example... Figure 1As shown, server 101 is connected to terminal device 103 via a network and can be used to provide services to terminal device or application 107 installed on terminal device. The application can be video application, instant messaging application, browser application, educational application, game application, etc. Database 105 can be set up on the server or independently of the server to provide data storage services for server 101, such as a game data storage server. The network mentioned above can include, but is not limited to, wired networks and wireless networks. The wired network includes local area networks, metropolitan area networks, and wide area networks. The wireless network includes Bluetooth, WIFI, and other networks that enable wireless communication. Terminal device 103 can be a terminal configured with an application, and can include, but is not limited to, at least one of the following: mobile phones (such as Android phones, iOS phones, etc.), laptops, tablets, handheld computers, MID (Mobile Internet Devices), PADs, desktop computers, smart TVs, smart voice interaction devices, smart home appliances, vehicle terminals, aircraft, virtual reality (VR) terminals, augmented reality (AR) terminals, mixed reality (MR) terminals, and other computer devices. The server mentioned above can be a single server, a server cluster composed of multiple servers, or a cloud server.

[0044] Combination Figure 1 As shown, the above account grouping method can be executed by an electronic device, which can be a terminal device or a server. The above account grouping method can be implemented by the terminal device or the server respectively, or by the terminal device and the server together.

[0045] Optionally, in the embodiments of this application, the specific application scenario may include, but is not limited to, the following: Figure 1 The following steps are shown:

[0046] S1: Grouping Strategy Generation and Initialization: The server first loads the player feature dataset for the current game, including structured data such as historical match records, social relationships, and equipment attributes, and then performs the following operations:

[0047] S1-1, Strategy rule parsing: Parses the pre-configured grouping strategy template and identifies the core parameters in the strategy (such as combat power balance weight and social relevance threshold);

[0048] S1-2, Strategy Instantiation: Perform multi-dimensional clustering of player accounts according to strategy rules, such as dividing players with a rank difference of less than 3 levels into a homogeneous group, or constructing a friend priority group based on social intimacy;

[0049] S1-3, Set Initialization: Generate multiple account team sets through random allocation or rule-driven generation. Each set contains at least two team division schemes (such as 4-person teams × 3 groups or 3-person teams × 4 groups) to ensure the strategy diversity of the initial set.

[0050] By combining rule parsing with parallel generation of multiple strategies, the limitations of a single grouping mode are overcome, providing a high-coverage candidate strategy pool for subsequent iterations.

[0051] S2: Second-Stage Retention Prediction Model Inference and Policy Evaluation:

[0052] The execution entity is the server, which calls a pre-trained target retention prediction model to perform batch predictions on the account team set:

[0053] S2-1, Feature Engineering Processing: Feature encoding is performed on each account team set, player attributes (such as role function and skill combination) are converted into numerical vectors, and serialized behavioral data (such as movement trajectory and resource consumption curve of the last 5 games) are extracted.

[0054] S2-2, Model Inference Execution: The encoded features are input into the target second-time retention prediction model. The second-time retention probability of each account team is calculated through the forward propagation of the deep neural network. The model uses an attention mechanism to capture the synergistic effect between players.

[0055] S2-3, Strategy retention probability aggregation: Calculate the weighted average of the retention probabilities of all teams within the same account team set (the weight can be configured as the number of team members or the proportion of combat power) to generate a strategy retention probability score matrix.

[0056] By using deep learning models to quantify the correlation between team allocation strategies and retention rates, a visual evaluation of the strategy's effectiveness can be achieved.

[0057] S3: Iterative optimization using genetic algorithm:

[0058] The server uses a genetic algorithm framework to perform multiple rounds of evolutionary optimization on the account team set:

[0059] S3-1, Selection Operation: Sort by strategy retention probability, use elite retention mechanism to select the top 20% of high-scoring strategies to enter the next generation, and eliminate the bottom 30% of inefficient strategies.

[0060] S3-2, Cross Operation: Pair up the retained strategies and generate new strategies by exchanging team members or reorganizing team rules. For example, swap the high-strength player groups in the two strategies to explore the possibility of cooperation.

[0061] S3-3, Mutation Operation: Introduce random perturbations to some strategies with a probability of 5%-15%, such as adjusting the team affiliation of individual players or fine-tuning the combat power balance threshold, to prevent the algorithm from getting stuck in local optima.

[0062] S3-4, Iterative State Detection: After the i-th iteration, calculate the distribution entropy of the policy's second-time retention probability in real time, and dynamically adjust the crossover probability and mutation intensity of subsequent iterations.

[0063] By employing a directed evolution mechanism, it efficiently searches for near-optimal solutions in an exponential policy space, balancing the needs of global exploration with local development.

[0064] S4: Target Account Team Set Decision and Output:

[0065] The server selects the target account team set from the N rounds of iterations based on the termination condition:

[0066] S4-1, Preset condition matching: Apply multi-level filtering rules, such as requiring the strategy's second-day retention probability to be ≥0.72 and the team's combat power difference coefficient to be ≤0.25, or select the candidate strategies with the highest second-day retention probability.

[0067] S4-2, Business rule adaptation: Optimize strategies based on real-time business needs, such as prioritizing strategies that include friend relationships during social activities, or strengthening the balance of combat power in competitive scenarios.

[0068] S4-3, Strategy Persistent Storage: Write the feature vectors and evaluation results of the target account team set into the strategy knowledge base for subsequent iterations of initialization parameter tuning or model training data augmentation.

[0069] Achieving closed-loop optimization and dynamic adaptation of team-based strategies significantly improves player retention and the intelligence level of the matchmaking system.

[0070] The entire process, from rule initialization in step S1 to strategy output in step S4, relies on player behavior data and model prediction results to reduce human intervention bias. Through the genetic algorithm iteration in step S3 and the strategy selection mechanism in step S4, the system can respond in real time to changes in player behavior patterns (such as tactical system changes caused by the launch of new characters). In steps S2 and S3, batch prediction and parallel computing technologies are used to ensure that the strategy generation time for tens of thousands of players is controlled within the second response range.

[0071] Optionally, in this embodiment, the above method can be applied to a game application, which can be a multiplayer online battle arena (MOBA) application or a single-player game (SPG) application. The types of the game applications may include, but are not limited to, at least one of the following: two-dimensional (2D) game applications, three-dimensional (3D) game applications, virtual reality (VR) game applications, augmented reality (AR) game applications, and mixed reality (MR) game applications. This is merely an example, and this embodiment does not impose any limitations on it.

[0072] In addition, the aforementioned shooting game application can be a third-person shooter (TPS) application, such as running the shooting game application from the perspective of a third-party character object other than the virtual character controlled by the current player, or it can be a first-person shooter (FPS) application, such as running the shooting game application from the perspective of the virtual character controlled by the current player.

[0073] It should be noted that this framework is applicable to multiplayer online game scenarios such as MOBA and FPS. Through an automated and intelligent team strategy generation mechanism, it systematically solves the problems of lag and trial-and-error costs in traditional manual rule systems.

[0074] The above is merely an example, and this embodiment does not impose any specific limitations.

[0075] Alternatively, as an alternative implementation method, such as Figure 2 As shown, the grouping methods for the above accounts include:

[0076] S202, Group the accounts based on the grouping strategy to obtain multiple account teams, wherein one account team in one of the multiple account teams includes multiple accounts determined according to the grouping strategy;

[0077] Optionally, in the embodiments of this application, the grouping strategy described above may include, but is not limited to, a set of logical rules or algorithms for dividing accounts into teams. Essentially, it defines the feature matching principles and combination constraints that must be followed during the account grouping process. For example, the grouping strategy may be constructed based on player attribute tags (such as combat style, equipment preference, historical win rate), social relationships (such as friend list, team records), or dynamic behavioral characteristics (such as recent activity level, in-game operation habits). In specific implementations, the grouping strategy can be a single rule (such as grouping by rank), a weighted combination of multiple rules (such as rank accounting for 60% and social affinity accounting for 40%), or an implicit rule generated based on a machine learning model (such as extracting player feature similarity through clustering algorithms). In typical application scenarios, the grouping strategy may include various types such as random allocation strategy, balanced combat power strategy, and social priority strategy, each strategy corresponding to different account combination logic.

[0078] It should be noted that the diversity of grouping strategies is not only reflected in the differences in rule types, but may also be expanded through the following dimensions: First, the strategy generation method can cover manually predefined rules (such as grouping by rank range), semi-automatic rules (such as manually setting weights and then dynamically adjusting them by the system), and fully automatic rules (such as strategies generated based on reinforcement learning); Second, the strategy optimization objectives can include single-objective optimization (such as maximizing the balance of team combat power) and multi-objective optimization (such as simultaneously optimizing combat power balance and social intimacy); Third, the scope of application of strategies can be divided into globally applicable strategies (applicable to all games), scenario-customized strategies (such as special grouping rules for peak nighttime hours), and player segmentation strategies (such as differentiated grouping for new players and experienced players).

[0079] Optionally, in this embodiment, the aforementioned account team set may include, but is not limited to, a set of candidate team assignment schemes generated through different grouping strategies, with each set representing a complete team configuration possibility. The account team set, as an intermediate data structure, stores multiple permutations and combinations of all accounts assigned to different teams in the same game. For example, in a game with 12 players, one account team set contains 3 teams of 4 players each, and another set also contains 3 teams of 4 players each. The generation of the account team set must satisfy game rule constraints (such as upper and lower limits on team size) and may inherit the feature matching priority defined in the grouping strategy. Its dynamic characteristics are reflected in the fact that the team combinations within the set are continuously updated according to the iterative process of the genetic algorithm to gradually approach the optimal team assignment scheme.

[0080] It should be noted that the method of generating the account team set can be adjusted according to actual needs: for example, in the initialization phase, Monte Carlo random sampling can be used to generate the candidate set, or historical best strategies can be reused as the initial set through transfer learning; during the iteration process, the set update can be based on the crossover and mutation mechanism of the classic genetic algorithm, or a hybrid algorithm such as simulated annealing and particle swarm optimization can be introduced; in addition, the set size (i.e. the number of candidate strategies) can be set to a fixed value (such as retaining the top 50 high-scoring strategies), or dynamically adjusted (such as gradually shrinking the search space according to the iteration rounds).

[0081] Optionally, in this embodiment, the aforementioned account team may include, but is not limited to, a game battle unit composed of multiple accounts, with each account team representing a group of players who can act independently in a match. As the basic unit for optimizing team composition strategies, the account team's composition must meet the requirements of a specific game mode (e.g., 2 to 5 players per team in some game modes). The principles for constructing account teams may involve balanced combat power (e.g., teams with similar total combat power), complementary roles (e.g., including healers, damage dealers, and tanks), or social connections (e.g., prioritizing team formation among members of the same guild). For example, in tactical competitive games, an account team may consist of one sniper, two assault soldiers, and one medic to meet tactical coordination requirements.

[0082] It should be noted that the composition dimensions of an account team can be further subdivided into: first, static attribute dimensions, including long-term stable features such as account level, equipment rating, and historical win rate; second, dynamic behavior dimensions, including recent combat data (such as resource collection efficiency) and operational habits (such as movement frequency and skill release accuracy); and third, social relationship dimensions, including friend intimacy, historical team formation success rate, and voice communication activity. These different dimensions can be used individually or in combination to define the internal synergy of an account team, and this application does not impose specific limitations on this.

[0083] S204, predict the retention probability of multiple account teams by using the target retention prediction model, and determine the strategy retention probability corresponding to each account team set. The target retention prediction model is used to predict the team retention probability of each account team in each account team set, and determine the strategy retention probability based on the team retention probability. The target retention prediction model includes a model trained by using sample account team sets and sample labels to train the initial retention prediction model. The sample labels are used to indicate the retention status of the sample account team sets.

[0084] Optionally, in the embodiments of this application, the above-mentioned target retention prediction model may include, but is not limited to, a prediction system based on deep learning. This system establishes a mapping relationship between team-based strategies and player retention by analyzing the multi-dimensional characteristics and historical behavioral data of players within an account team. This model typically consists of a feature encoding layer, a temporal processing module, and a probability prediction module. For example, it may use a Transformer structure to process players' serialized game data (such as the number of defeats, movement trajectories, and resource consumption in the last 10 games), and fuse static attributes (such as rank and equipment rating) through a fully connected network to finally output the next-day retention probability for each account within the team. During training, the model optimizes the cross-entropy loss between the predicted value and the actual retention label through backpropagation, and combines ordinal-preserving regression to calibrate the probability distribution bias. In typical application scenarios, the target retention prediction model can adapt to the feature systems of different game types. For example, in tactical competitive games, it focuses on combat behavior analysis, while in social games, it focuses on player interaction feature modeling.

[0085] It should be noted that the dimensions of the target retention prediction model can be expanded according to actual needs: First, the input feature dimension can include basic attributes (such as account level, equipment rating), behavior sequences (such as heatmaps of movement trajectories in the last 20 games), and social graphs (such as team relationship network embedding representations); Second, the model structure dimension can use temporal convolutional networks to handle short-term behavior fluctuations, graph neural networks to mine collaborative relationships between players, and attention mechanisms to capture key feature interactions; Third, the training strategy dimension can adopt course learning to gradually transition from simple scenarios to complex scenarios, adversarial training to enhance model robustness, and federated learning to achieve cross-server data joint modeling. This application does not make specific limitations on these aspects.

[0086] Optionally, in this embodiment, the aforementioned strategy retention probability may include, but is not limited to, a quantitative indicator reflecting the overall retention effect of the team strategy, which is obtained by aggregating the retention probabilities of all teams within the same account's team set. For example, the strategy retention probability may be the arithmetic mean of the retention probabilities of all teams within the set, reflecting the overall expected value of player retention under this strategy; or a weighted average method may be used, assigning different weights based on the number of team members or combat power to optimize the impact of small-scale teams on strategy evaluation. In complex scenarios, the calculation of the strategy retention probability can quantify the marginal contribution of a single player in different team combinations, thereby more accurately evaluating the group effect of the team strategy.

[0087] It should be noted that the calculation method for the strategy retention probability can be adjusted according to business objectives: for example, in scenarios emphasizing team synergy, collaborative filtering algorithms can be used to uncover common characteristics of historically high retention teams, serving as a benchmark for strategy evaluation; in scenarios pursuing global optimization, Nash equilibrium theory can be introduced to calculate the optimal solution under multi-strategy games; in scenarios requiring rapid iteration, Monte Carlo simulation can be used to generate virtual strategy retention probabilities to accelerate the evaluation process. Combinations of different calculation methods can form a hierarchical evaluation system to meet diverse business needs.

[0088] The aforementioned team retention probability can include, but is not limited to, a predicted value representing the retention tendency of players within a single team. Its calculation process comprises three stages: feature extraction, model inference, and probability calibration. For example, for a four-player team, the model jointly encodes each player's historical match data (such as average survival time in the last five matches and resource gathering efficiency) with current team composition characteristics (such as role synergy and rank differences), calculates the retention probability of each player using a deep neural network, and then takes the geometric mean of the retention probabilities of all players in the team as the team's second-time retention probability. In practical applications, the calculation of the team's second-time retention probability may incorporate a time decay factor, giving higher weight to recent behavioral data to improve the timeliness of the prediction results.

[0089] Optionally, in this embodiment, the aforementioned sample account team set may include, but is not limited to, a historical team data set used to train the target retention prediction model. This set covers team combinations generated under different periods and matching rules, along with their actual retention results. For example, the sample set may contain team records from 100,000 matches over the past 30 days, with each record containing feature vectors of team members (such as combat style tags and social affinity scores) and corresponding next-day login status tags. The data preprocessing stage may employ a sliding window sampling strategy to ensure the training set covers diverse match scenarios during weekends / weekdays and peak / off-peak periods.

[0090] It should be noted that there are several possibilities for expanding the criteria for constructing the sample account team set: First, the time span can be short-term data (such as the last 7 days) to capture player behavior trends, or long-term data (such as quarterly data) to extract periodic patterns; second, the data sources can integrate cross-server battle records, test server experimental data, and simulator-generated data; third, the sampling strategy can adopt stratified sampling to ensure a balanced proportion of players in each rank, cluster sampling to cover typical player groups, and active learning to select high-information samples. This application does not make specific limitations on these aspects.

[0091] Optionally, in the embodiments of this application, the aforementioned sample labels may include, but are not limited to, binary or multi-class classification labels used for supervised model training, and their definition depends on the retention period setting of the business scenario. For example, in the basic implementation, the sample label may be defined as a binary variable (0 indicates no login the next day, 1 indicates login). In the extended implementation, multi-period labels such as 7-day retention and 30-day retention can be introduced to construct a multi-task learning framework. The label generation process may include a data cleaning step, such as removing abnormal login records caused by external factors such as server maintenance, or deduplicating short-term repeated login behaviors to improve the reliability of the label data.

[0092] It should be noted that the labeling system for the sample tags can be enriched through the following dimensions: First, at the time granularity level, minute-level activity (such as behavior density within 1 hour after login), daily activity (such as daily game time), and weekly activity (such as consecutive login days) can be defined; Second, at the behavior depth level, basic login behavior (such as entering the game lobby), core behavior (such as completing a full game), and value-added behavior (such as purchasing virtual items) can be distinguished; Third, at the correlation analysis level, composite retention indicators can be constructed by combining external data (such as social media interaction frequency and customer service consultation records), and this application does not make specific limitations on this.

[0093] Optionally, in the embodiments of this application, the initial second-time retention prediction model may include, but is not limited to, an untrained or pre-trained basic model architecture, which serves as the starting point for optimizing the target second-time retention prediction model. The initial model may use a classic machine learning model (such as a gradient boosting decision tree) as a baseline, or reuse retention prediction model parameters trained in other game scenarios based on transfer learning. For example, in the cold start phase, the initial model may load parameters from a player behavior analysis module trained in a MOBA game, fine-tuning to adapt to differences in game feature distribution. Model initialization strategies may include different methods such as random parameter initialization and Xavier uniform distribution initialization to control the convergence speed and stability of the training process.

[0094] S206. Use a genetic algorithm to iterate through multiple account team sets for N rounds according to the team second-time retention probability. In the i-th round of iteration, the strategy second-time retention probability corresponding to each account team set generated in the i-th round of iteration is determined by the target second-time retention prediction model. Here, N and i are positive integers, and i≤N.

[0095] Optionally, in the embodiments of this application, the genetic algorithm described above may include, but is not limited to, an optimization algorithm that simulates the process of biological evolution, which iteratively optimizes candidate solutions through operations such as selection, crossover, and mutation. In the scenario of game team grouping, the chromosome encoding of the genetic algorithm can be represented as a group of teaming strategies (such as the permutation and combination of player teams), and the fitness function is defined as the second-time retention probability of the strategy. High-fitness strategies are selected through roulette wheel selection or tournament selection, new strategies are generated through single-point crossover or multi-point crossover operations, and mutation is achieved through random replacement or heuristic adjustment. For example, in the iterative process containing 10 candidate strategies, the algorithm may retain the top 3 strategies with high second-time retention probabilities, perform crossover and recombination on the remaining 7 strategies (such as exchanging some team members in two strategies), and randomly adjust the team affiliation of 1 to 2 players in the recombined strategy, thereby generating a new generation of strategy set. The core advantage of the genetic algorithm lies in its global search capability, which can avoid getting trapped in local optima and is suitable for high-dimensional combinatorial optimization problems of teaming strategies.

[0096] It should be noted that the optimization mechanism of the genetic algorithm can be extended in the following dimensions: First, the selection strategy can adopt an elite retention strategy (directly retaining the top K optimal strategies), a steady-state selection strategy (only replacing some low-fitness strategies), or a fitness ratio selection (dynamically adjusting the selection ratio according to the iteration rounds); Second, the crossover method can include uniform crossover (exchanging each gene position according to probability), sequential crossover (retaining part of the arrangement order of the parent strategy), or heuristic crossover based on matching quality (prioritizing the exchange of players with complementary features); Third, the mutation operation can be designed as random player replacement, directional adjustment based on historical win rates, or generating perturbation directions through reinforcement learning strategies. This application does not make specific limitations on these aspects.

[0097] Optionally, in the embodiments of this application, the aforementioned N rounds of iteration may include, but are not limited to, the evolutionary cycle control parameter of the genetic algorithm, which defines the total number of rounds of policy optimization. The value of N can be dynamically adjusted according to computing resources, real-time requirements, or convergence status. For example, it can be set to 5 rounds in a real-time matching scenario to achieve fast response, and to 100 rounds in an offline optimization scenario to fully explore the policy space. Each round of iteration includes four stages: policy evaluation (calculating fitness through the target second-retention prediction model), policy selection (retaining high-fitness policies), policy recombination (cross-generating new policies), and policy mutation (introducing random perturbations). In typical applications, the termination condition for N rounds of iteration can be set to the variance of the policy second-retention probability being less than a threshold (e.g., fluctuation less than 1% for 3 consecutive rounds), or reaching a preset maximum number of iterations, to ensure that the algorithm outputs a stable solution within a reasonable time.

[0098] It should be noted that there are several possibilities for setting the termination conditions for N rounds of iteration: First, convergence can be determined based on the mean change rate of the policy retention probability (e.g., the increase is less than 0.5% for three consecutive rounds), the diversity index of the policy set (e.g., the difference between policies is less than a threshold), or the stability of the Pareto front (e.g., adding a new policy no longer significantly changes the optimal solution set); Second, resource constraints may include the maximum computation time (e.g., no more than 500 milliseconds), the upper limit of memory usage (e.g., the size of the policy set does not exceed 10,000), or the processor load threshold (e.g., CPU utilization does not exceed 80%); Third, dynamic policy adjustment can be combined with early stopping mechanisms (terminating early when an overfitting trend is detected), elastic scaling mechanisms (adjusting the number of iterations according to real-time computing resources), or phased optimization (dividing N rounds into multiple sub-stages and setting termination conditions for each). This application does not make specific limitations on these.

[0099] Optionally, in the embodiments of this application, the aforementioned i-th iteration may include, but is not limited to, a specific evolutionary stage in the genetic algorithm optimization process, which characterizes the algorithm's evolutionary state in the time dimension. In the i-th iteration, the algorithm dynamically adjusts the selection pressure (e.g., gradually increasing the screening threshold) or mutation probability (e.g., reducing the perturbation intensity as the iteration progresses) based on the historical optimization results of the previous i-1 rounds. For example, in the initial iteration stage (i = 1 to 3), the algorithm may adopt a high mutation rate (e.g., 20%) to expand the search range; in the intermediate stage (i = 4 to 7), a gradual crossover strategy is adopted to balance exploration and development; in the later stage (i ≥ 8), the focus is on local fine-tuning. The intermediate results of the i-th iteration (e.g., the distribution trend of the strategy's second-retention probability) can be used to detect the algorithm's convergence state and provide a basis for dynamically adjusting the iteration parameters.

[0100] It should be noted that the optimization of the i-th iteration process can be enhanced in the following dimensions: First, dynamic tuning of iteration parameters may include exponential decay of crossover probability with each round, adaptive adjustment of mutation probability based on population diversity, or segmented setting of selection pressure based on historical convergence speed; Second, parallel processing may employ multi-threaded synchronous evaluation of strategy retention probability, distributed computing framework for sharding strategy set processing, or GPU acceleration of model inference process; Third, utilization of historical information may involve introducing a memory pool to store excellent strategies from previous generations, predicting optimal parameter configuration through meta-learning, or constructing a strategy evolution trajectory map to guide the search direction. This application does not impose specific limitations on these aspects.

[0101] Furthermore, the iterative update mechanism for the account team set can be further expanded: First, diversity maintenance can be achieved by introducing niche techniques (to limit the over-breeding of similar strategies), periodically injecting random new strategies, or adopting a diversity-weighted fitness function; Second, strategy generation can be combined with local search algorithms (such as simulated annealing) to deeply optimize high-potential strategies, or by integrating external knowledge (such as player feedback data) to guide the direction of strategy evolution; Third, multi-objective optimization can construct a Pareto optimal solution set (simultaneously optimizing second-retention probability, matching time, and server load), or by converting multiple indicators into a comprehensive fitness function through a weighted summation method. This application does not make specific limitations on this.

[0102] S208, determine the target account team set based on the policy retention probability corresponding to each account team set generated by N rounds of iteration, where the target account team set represents the account team set whose policy retention probability satisfies the preset condition.

[0103] Optionally, in this embodiment, the aforementioned preset conditions may include, but are not limited to, decision rules for filtering target account teams, which define a threshold or ranking rule for the strategy's second-time retention probability. The preset conditions can be set as static thresholds (e.g., strategy second-time retention probability ≥ 0.8), dynamic thresholds (e.g., top 5% quantile), or combinations of multiple conditions (e.g., second-time retention probability ≥ 0.7 and team strength difference ≤ 10%). For example, in a real-time matching scenario, the preset conditions may require the strategy's second-time retention probability to be in the top 10% quantile of the current iteration round, and the strategy generation time to be no more than 50 milliseconds. The complexity of the preset conditions can be adjusted according to business needs, for example, by introducing fuzzy logic to handle uncertainty, or by combining multi-objective optimization to generate a Pareto optimal solution set.

[0104] It should be noted that the rules for setting preset conditions can be enriched through the following dimensions: First, the threshold type may include absolute thresholds (such as second-day retention probability ≥ 0.75), relative thresholds (such as 10% higher than the historical average), or mixed thresholds (such as simultaneously meeting absolute value and ranking requirements); Second, the condition logic may include single-condition judgment, multiple conditions and / or logical combinations, or hierarchical condition filtering based on decision trees; Third, the dynamic adjustment mechanism may combine time decay factors (such as relaxing the threshold during peak periods), player grouping strategies (such as using a lower threshold for new players), or real-time load balancing (such as prioritizing strategies with low computational time when server pressure is high), and this application does not make specific limitations on this.

[0105] Optionally, in this embodiment, the target account team set may include, but is not limited to, the optimal team strategy set selected after N rounds of iteration, representing the player grouping scheme with the best comprehensive evaluation result within the current iteration cycle. The generation of the target account team set must meet preset constraints and may retain multiple candidate strategies to cope with dynamic changes in real-time matching. For example, in the iteration results containing 100 account team sets, the target set may consist of the top 5 sets with the highest retention probability for each strategy, with each set corresponding to a different team focus (such as high retention potential or high balanced combat power). In practical applications, the selection of the target set may be further combined with business rule filtering (such as excluding strategies containing suspected cheaters) or operational strategy preferences (such as prioritizing the promotion of new player-guided grouping schemes).

[0106] It should be noted that there are several optimization paths for the generation logic of the target account team set: First, in single-objective optimization scenarios, the set with the highest second-day retention probability can be directly selected; second, in multi-objective optimization scenarios, Pareto front analysis can be used to simultaneously optimize metrics such as second-day retention probability, matching time, and social satisfaction; third, in dynamic scenarios, a candidate pool caching mechanism can be built to dynamically adjust the target set based on real-time player behavior data (such as temporary team-up requests). For example, during sudden hot events, the selection of the target set can temporarily incorporate event task completion rate prediction metrics to adapt to changes in operational needs; this application does not impose specific limitations on this.

[0107] Furthermore, the application dimensions of the target account team set can be further expanded: First, the strategy persistent storage can archive the historical best set for use in the initial strategy generation or model training data enhancement during the cold start phase; Second, strategy interpretability analysis can assist operators in understanding the strategy logic through feature importance attribution (such as the distribution pattern of player activity time in high-order retention strategies) or rule extraction (such as decision tree visualization of key screening conditions); Third, strategy gray release can use A / B testing to compare the retention differences between new and old strategies, or gradually expand the coverage of high-scoring strategies in batches, which is not specifically limited in this application.

[0108] In one exemplary embodiment, Figure 3 This is a schematic diagram of an optional account grouping method according to an embodiment of this application, such as... Figure 3 As shown, taking a team matchmaking scenario in a multiplayer online battle arena (MOBA) game as an example, the process includes, but is not limited to, the following:

[0109] S1: The server retrieves player account data from the current matchmaking pool and generates an initial team composition strategy set based on character specialization characteristics.

[0110] Analyze players' historical match data to identify their win rates and usage frequency for their core roles (such as tanks, damage dealers, and supports) and construct a character ability matrix;

[0111] Generate initial strategies that include role balance constraints (such as each team must contain at least 1 tank, 1 support, and 2 DPS), while allowing the random generation of some non-standard strategies to increase diversity;

[0112] Eliminate invalid strategies that violate character ratio rules (such as having more than 3 DPS characters in a single team) or map-specific restrictions (such as disabling dual support lineups).

[0113] S2: Use the pre-trained lineup second-retention prediction model to evaluate the policy's effectiveness:

[0114] Extract player character proficiency (e.g., number of games played with a certain hero), team collaboration metrics (e.g., support response speed), and lineup counter relationships (e.g., win rate of common enemy lineups) to construct a multi-dimensional feature vector;

[0115] By inputting team composition features into a temporal convolutional network, the system predicts the 7-day retention tendency of players under a specific team composition and outputs the team composition strength score and retention probability.

[0116] Strategies are prioritized based on a weighted average of lineup strength and retention probability (e.g., strength weight 60%, retention weight 40%).

[0117] S3: Employing an improved genetic algorithm for multi-dimensional strategy evolution:

[0118] The strategy of retaining the top 10% of high-scoring solutions directly into the next generation avoids the loss of high-quality solutions;

[0119] Gene exchange between strategies with complementary roles (such as recombining the tank positions of a high-survivability lineup with the output positions of a high-burst lineup);

[0120] Based on historical performance data of the lineup, low-win-rate characters are replaced in a targeted manner (such as replacing support heroes with a win rate of less than 45% with high-win-rate candidates of the same role).

[0121] Multi-objective fitness calculation: Simultaneously optimize lineup strength, character diversity, and estimated matching time to construct a Pareto optimal solution set.

[0122] S4: Select the final squad deployment plan based on the real-time battlefield environment:

[0123] For different map modes (such as tower defense mode and capture the stronghold mode), corresponding optimization conditions are loaded. For example, in tower defense mode, the weight of the wave clearing speed indicator is increased.

[0124] Provided that the retention rate of the strategy is met, the team allocation scheme with the least impact on server load should be selected first (such as avoiding the allocation of too many high-latency players to the same physical node);

[0125] An implicit feedback mechanism is introduced to automatically increase the priority of relevant strategies for players who have actively liked similar lineups.

[0126] This embodiment addresses the issue of fixed team compositions in traditional matchmaking by combining role-based constraints with random exploration during the strategy generation phase. In the model evaluation phase, a multi-objective weighted mechanism quantifies the relationship between team strength and long-term retention, overcoming the limitations of a single win rate indicator. During the decision-making and deployment phase, game environment adaptation and load balancing mechanisms achieve a dual improvement in strategy stability and system resource efficiency. Simultaneously, by reducing team composition conflict complaint rates, it optimizes the player community ecosystem, providing a reusable intelligent matchmaking paradigm for the long-term operation of competitive games.

[0127] In one exemplary embodiment, Figure 4 This is a schematic diagram of another optional account grouping method according to an embodiment of this application, such as... Figure 4 As shown, taking the scenario of matching interest-based communities in a multiplayer online social casual game as an example, the process includes, but is not limited to, the following:

[0128] S1: Social attribute grouping strategy generation and initialization. The server obtains the social behavior data of currently active players and generates an initial set of community strategies based on interest tags and interaction relationships.

[0129] Analyze player history (such as participation in themed events and virtual item collection records) and extract interest tags (such as music creation, virtual gardening, and story puzzle solving);

[0130] Generate strategies that include homogeneous interest strategies (such as teaming up players who collect the same virtual furniture) and complementary interest strategies (such as mixing music creators and dance enthusiasts), while also introducing random perturbation strategies.

[0131] Filter player groups with negative interaction records (such as blocking player groups that have been reported by each other) to ensure a positive community atmosphere.

[0132] S2: Predicting the correlation between community activity and retention, and evaluating the effectiveness of the strategy by calling a multimodal social second-day retention prediction model:

[0133] By integrating shallow player attributes (such as login time patterns and gift-giving frequency) with deep behavioral sequences (such as voice chat duration trends and collaborative task completion rates), a spatiotemporal attention feature vector is constructed.

[0134] Based on potential social relationship chains, predict players' 30-day retention tendency and increase in interaction activity in specific communities;

[0135] The overall score of the community strategy is calculated, including the second-day retention rate (weight 50%), the average number of daily interaction events (weight 30%), and the interest diversity index (weight 20%).

[0136] S3: Iterative optimization using a social genetic algorithm, employing a social relationship-enhancing genetic algorithm for dynamic optimization:

[0137] Select high-retention community strategies from the strategy pool and exchange their interest subgroups (e.g., recombining the music creation group in strategy A with the virtual gardening group in strategy B);

[0138] High-influence players are randomly inserted into low-activity communities with a probability of probability, or highly dependent small groups are split up to stimulate cross-group interaction;

[0139] Simultaneously optimize the second-day retention probability, community topic generation capabilities (such as predicting the rate of new topic generation), and cross-group collaboration potential (such as calculating the cross-complementary value of interests).

[0140] S4: Adaptive community strategy decision-making and deployment, selecting the final community solution based on real-time social trending topics:

[0141] Identify recent popular event themes (such as holiday-limited tasks) and prioritize community strategies that include relevant interest tags;

[0142] While meeting the second-time retention threshold, players with high interaction needs (such as those who frequently initiate chats) are evenly distributed to different communities to avoid information overload in a single group;

[0143] The maximum number of people in a community will be dynamically adjusted based on server resources (e.g., a small group of 6 people will be used during peak activity periods, and a large group of 12 people will be used during off-peak periods).

[0144] In this embodiment, during the strategy generation stage, the homogenization bottleneck of communities under artificial rules is overcome by using a hybrid initialization of interest graphs and random strategies, thereby improving the efficiency of cold start of interest. During the model evaluation stage, the periodic characteristics of social behavior are captured by using spatiotemporal attention mechanisms, and implicit relationship chains are mined by combining graph networks, thereby reducing the error rate of second-time retention prediction and significantly improving players' sense of belonging and enthusiasm for content co-creation, providing a technical foundation for the sustainable operation of the virtual social ecosystem.

[0145] In one exemplary embodiment, Figure 5 This is a schematic diagram illustrating another optional account grouping method according to an embodiment of this application, such as... Figure 5 As shown, taking a matchmaking scenario in a tactical cooperative team shooter game as an example, the process includes, but is not limited to, the following:

[0146] S1: Tactical role grouping strategy generation and initialization. The server collects player combat data in the matchmaking queue in real time and generates an initial team strategy set based on role functions and tactical history.

[0147] Analyze players' historical battle records to identify their proficiency index and virtual item mastery level for their core combat roles (such as assault, sniper, and medic), and construct an adaptation matrix;

[0148] Generate strategies that include standard tactical configurations (such as each team consisting of 1 medic, 2 assault soldiers, and 1 sniper), while also introducing experimental combinations (such as a high-survivability lineup with two medics or an all-assault lineup with all assault soldiers).

[0149] Eliminate strategies that violate map mechanics (such as teams lacking bomb disposal experts in bomb defusal mode) or combinations with unbalanced virtual item counter relationships (such as a team-wide melee virtual item configuration).

[0150] S2: Predicting the correlation between tactical coordination effectiveness and retention, and using a tactical second-retention prediction model to evaluate the long-term value of squad strategies:

[0151] Extract player micro-operation data (such as headshot rate and healing response delay), team macro-indicators (such as stronghold capture efficiency and team battle coordination score), and social interaction density (such as voice command frequency) to construct a three-dimensional feature tensor.

[0152] By inputting features into a spatiotemporal convolutional network, the system predicts the 14-day retention probability of players under specific tactical combinations and simultaneously outputs a team tactical strength score (such as positional warfare defense capability and raid success rate).

[0153] The weights are dynamically adjusted according to the game's operational phase (e.g., tactical innovation is emphasized in the new season, and retention stability is emphasized in the stable period), to generate a comprehensive strategy priority list.

[0154] S3: Iterative optimization of the tactical genetic algorithm, using a tactical-enhanced genetic algorithm for strategy evolution:

[0155] Select historical high-win-rate tactical templates (such as crossfire formations and diamond advance strategies) as a high-quality gene pool to improve the quality of initial iterations;

[0156] Exchange roles between complementary strategies (e.g., refactor the medics of strategy A and the assault team of strategy B into a rapid rescue unit);

[0157] Based on the characteristics of the terrain (such as narrow alleys or open plains), randomly adjust the character's positioning logic, or replace the configuration of virtual props to adapt to the terrain's counter-relationships.

[0158] Balance tactical intensity, role diversity, and real-time matching to construct a non-dominated solution set for subsequent decision-making.

[0159] S4: Real-time tactical environment decision-making and deployment, combined with dynamic battlefield data to output final unit plans:

[0160] Load the optimal tactical library for the current map (e.g., prioritize highly mobile lineups in capture mode, and focus on defensive fortification construction capabilities in defense mode);

[0161] Dynamically detect player device latency and voice communication quality, and prioritize assigning low-latency players in the same region to the same team;

[0162] Based on historical data, predict players' tactical execution tendencies (such as aggressive advances or conservative defense) to avoid pairing players with conflicting strategies with each other.

[0163] In this embodiment, during the strategy generation stage, a hybrid strategy combining functional balance constraints and experimental combinations is used to overcome the rigidity of lineups in traditional matchmaking. During the model evaluation stage, micro-operations and macro-tactical indicators are integrated, and spatiotemporal convolution is used to capture dynamic battlefield relationships, thereby improving the accuracy of second-retention prediction, significantly enhancing the smoothness of team battle coordination and players' tactical identification, and providing a continuously optimized squad infrastructure for high-intensity competitive scenarios.

[0164] Through the embodiments of this application, a dynamic grouping mechanism based on multi-strategy initialization is adopted. By generating a set of account teams containing different team-up rules as the initial strategy pool for genetic algorithm iteration, a diverse search starting point is provided for subsequent optimization, achieving the goal of overcoming the limitations of manual rules and realizing the technical effect of automatic exploration of the strategy space. By constructing a deep neural network model that integrates sequential and non-sequential features, and using historical player team features and real retention labels to train the target second-day retention prediction model, the probability value output by the model is calibrated to achieve the goal of accurately quantifying the correlation between team-up strategies and retention, thereby realizing the direct coupling of strategy effect evaluation indicators and core business objectives. By introducing the evolutionary mechanism of the genetic algorithm, new strategies are dynamically generated using crossover and mutation operators. Through multiple rounds of iteration, a set of account teams with high strategy second-day retention probabilities is selected, achieving the goal of targeted optimization of the strategy space, thereby realizing the technical effect of efficiently converging to the Pareto optimal solution in an exponential combination of possibilities. By establishing a termination condition judgment mechanism based on strategy second-day retention probability, using a preset probability threshold or iteration round threshold as the stopping standard, and through the automatic selection mechanism of the globally optimal strategy, the goal of dynamically adapting to real-time player characteristics is achieved, thereby realizing a closed-loop optimization system without manual intervention. Ultimately, through the synergistic effect of the above-mentioned technical paths, a complete technical closed loop from strategy generation and effect evaluation to iterative optimization was formed, realizing the positive correlation between team strategy and player retention. This solved the core pain points of traditional manual rule systems, such as high lag and high trial-and-error costs, and provided a quantifiable and iterative optimization paradigm for intelligent matchmaking systems in multiplayer online games. In turn, it solved the technical problem of low efficiency and accuracy of account grouping, which affected account retention.

[0165] As an alternative solution, Figure 6 This is a schematic diagram illustrating another optional account grouping method according to an embodiment of this application, such as... Figure 6 As shown, a target retention prediction model is used to predict the retention probability of multiple account teams, determining the policy retention probability for each account team team, including:

[0166] S602, predict the second-time retention probability of each account in each account team in a set of multiple account teams by using the target retention prediction model;

[0167] S604, determine the team retention probability for each account team based on the account retention probability;

[0168] S606 determines the strategy retention probability for each account's team set based on the team retention probability.

[0169] Optionally, in this embodiment, the aforementioned account retention probability may include, but is not limited to, a quantitative indicator characterizing the likelihood of a single player's retention under a specific team configuration. Its value ranges from 0 to 1, reflecting the probability that a player will continue playing due to the current team strategy. For example, for a player skilled in sniping, if assigned to an assault team lacking frontline cover, the model may output a lower account retention probability based on churn records in similar scenarios in their historical data. The calculation of the account retention probability needs to consider individual player differences; for example, social players are more sensitive to the quality of team communication than lone wolf players.

[0170] It should be noted that there are multiple optimization paths for generating the account retention probability: First, the probability aggregation method can use fuzzy logic to handle player preference conflicts (such as wanting to team up with friends but rejecting a specific tactical style), or introduce the Shapley value from game theory to allocate player contribution; Second, the time dimension can be subdivided into immediate retention (login behavior within 1 hour after the end of the game), short-term retention (within 24 hours), and long-term retention (within 7 days) to build a multi-task prediction model; Third, the application scenario can be extended to asymmetric competitive games (such as 1v4 chase mode), and special rules can be adapted by adjusting feature weights. This application does not make specific limitations in this regard.

[0171] Optionally, in this embodiment, the team's second-time retention probability may include, but is not limited to, an aggregated indicator reflecting the retention potential of the entire team, which is calculated by applying specific rules to the second-time retention probabilities of all accounts within the team. For example, a geometric mean can be used to calculate the team's second-time retention probability (to avoid interference from extreme values), or a weighted average method can be used to assign a higher weight to the team leader's account. In complex scenarios, the calculation of the team's second-time retention probability may introduce a synergistic effect factor, such as amplifying the overall probability when there are player combinations in the team with a historical team win rate exceeding a threshold.

[0172] It should be noted that the calculation rules for the team's second-time retention probability can be enhanced in the following dimensions: First, the calculation method can introduce a team role integrity factor (such as reducing the probability when a healing role is missing), or dynamically adjust the weight according to the map type (such as increasing the weight of resource allocation efficiency in survival mode); Second, the adjustment factor can include a social intimacy coefficient (increased probability of teaming up with friends) and a historical combat win rate coefficient (strengthened probability of counter lineups); Third, the evaluation dimensions can add implicit factors such as intra-team communication quality indicators (such as voice command response delay) and tactical execution consistency (such as skill release coordination), which are not specifically limited in this application.

[0173] Optionally, in this embodiment, the aforementioned strategy retention probability may include, but is not limited to, a core indicator for evaluating the overall effectiveness of a team strategy, which is generated by aggregating the retention probabilities of all teams within the same account's team set. For example, the strategy retention probability may be the maximum retention probability of teams within the set (reflecting optimal potential), or the average of the top 25% of teams may be obtained using quantile statistics (balancing stability and upper limit exploration). In dynamic matching scenarios, the strategy retention probability may incorporate a time decay factor to prioritize the evaluation of the short-term effects of newly generated strategies.

[0174] It should be noted that the optimization direction of the strategy retention probability includes multiple possibilities for expansion: First, the aggregation method can adopt the sliding window mean (only evaluating the strategies generated in the most recent N iterations), or combine the strategy generation timestamp for freshness weighting; Second, the application scenario can be adapted to large-scale team battle modes (such as 50v50 faction battles), through a hierarchical aggregation strategy (first calculating the team probability and then summing the faction probability); Third, the dynamic adjustment mechanism can introduce real-time load feedback (such as prioritizing low computational consumption strategies when the server pressure is high), or player behavior prediction intervention (such as adjusting the threshold according to the retention fluctuation during peak periods). This application does not make specific limitations on this.

[0175] Through the embodiments of this application, the second-day retention probability of an account can be predicted first, then the second-day retention probability of a team can be predicted based on the second-day retention probability of the account, and finally the second-day retention probability of the strategy can be determined based on the second-day retention probability of the team, so as to map account characteristics and retention, and the output result is the grouping strategy with the optimal account retention.

[0176] As an alternative solution, such as Figure 6 As shown, the target retention prediction model is used to predict the retention probability of each account in each account group within a set of multiple account groups, determining the retention probability of each account. This includes:

[0177] Obtain the account characteristics corresponding to each account in each account team;

[0178] The retention probability of each account is determined based on account characteristics.

[0179] Optionally, in this embodiment, the aforementioned account features may include, but are not limited to, a multi-dimensional dataset describing player account attributes and behavioral patterns. This dataset is extracted using structured or unstructured data processing techniques to quantify a player's potential retention tendency under specific team configurations. For example, account features may include static attributes (such as account level, equipment rating, and historical win rate), dynamic behavioral sequences (such as heatmaps of movement trajectories in the last 10 games and skill release interval distribution), and social association indicators (such as historical team formation success rate and voice interaction frequency). In specific implementations, account features may undergo standardization (such as Z-score normalization), temporal feature engineering (such as sliding window statistics of resource acquisition efficiency changes in the last 5 games), or embedding representation learning (such as compressing high-dimensional behavioral data through an autoencoder) to adapt to the model's input requirements.

[0180] It should be noted that the dimensions for constructing account features can be extended to the following directions: First, feature types can include environmental context features (such as the game time being during holidays or weekdays), device performance features (such as the stability of the player's terminal frame rate), and psychological profile features (such as the player's decision-making style inferred from behavioral data); Second, feature preprocessing methods can employ wavelet transform to extract the periodic patterns of behavioral sequences, graph convolutional networks to mine community structures in social relationship networks, and adversarial generative networks to enhance the representational ability of sparse data; Third, feature source dimensions can integrate general features across game modes (such as the distribution of account activity time), features customized for specific gameplay (such as the survival time in the poison circle in team competitive mode), and features from third-party data interfaces (such as the analysis of operating habits driven by peripheral devices). This application does not impose specific limitations on these aspects.

[0181] Through the embodiments of this application, fine-grained modeling of account features can accurately capture the implicit relationship between player behavior patterns and team configuration, providing high-information-density input data for second-retention prediction; the dynamic weight allocation mechanism enables the model to adapt to the differences in feature importance in different game scenarios, improving the scenario adaptability of the prediction results; diversified feature processing and model architecture selection provide a scalable technical foundation for strategy optimization in complex game environments, avoiding the generalization bottleneck caused by a single model architecture.

[0182] As an alternative solution, such as Figure 6 As shown, obtain the account characteristics corresponding to each account in each account team, including:

[0183] Obtain the sequence features corresponding to each account, where the sequence features are used to represent the account's game data within a preset time period;

[0184] Obtain the non-sequence features corresponding to each account, where the non-sequence features are used to represent the account parameters;

[0185] Account characteristics are determined based on sequence features and non-sequence features.

[0186] Optionally, in this embodiment, the aforementioned sequence features may include, but are not limited to, a time-series data set reflecting the dynamic behavior changes of a player's account within a specific time window. This data, recorded through structured time series, details of gameplay behavior and is used to capture the correlation between short-term player behavior patterns and long-term trends. For example, the sequence features may include time-series correlated indicators such as average survival time per game, kill / assist time distribution curves, and resource acquisition efficiency change rates from the last 20 games. In specific implementations, the sequence features may undergo sliding window statistics (e.g., calculating the variance of movement speed in the last 5 games), time-series interpolation (handling missing game counts due to disconnections), or Fourier transform to extract periodic patterns (e.g., identifying daily periodic fluctuations during active player periods) to adapt to the input requirements of different models.

[0187] It should be noted that the data collection and processing methods for sequence features can be extended to the following dimensions: First, the data collection period can be set to a fixed window (such as 50 games in the last 24 hours), a dynamic window (automatically adjusting the number of games based on player activity), or an event-triggered window (such as collecting only game data after participating in a specific mode); Second, the preprocessing methods can employ wavelet denoising to eliminate abnormal fluctuations, dynamic time warping to align sequences of different game lengths, or adversarial generative networks to enhance the richness of small sample sequences; Third, the feature types can include operational behavior sequences (such as skill release interval heatmaps), social interaction sequences (such as team invitation response time distributions), or economic behavior sequences (such as virtual resource consumption rate curves), and this application does not specifically limit these.

[0188] Optionally, in this embodiment, the aforementioned non-sequential features may include, but are not limited to, a set of static parameters describing the inherent attributes and stable state of a player account, reflecting the account's long-term accumulated attributes or configuration preferences. For example, non-sequential features may include time-independent indicators such as account level, rank rating, total equipment power, character skill tree configuration ratio, and highest historical season ranking. In practical applications, non-sequential features may undergo standardization (e.g., mapping rank rating to the 0-1 range), discretization (e.g., dividing equipment power into 5 level ranges), or correlation encoding (e.g., converting character skill tree configuration into a binary bitmap) to improve the model's adaptability to feature distribution.

[0189] It should be noted that the parameter system of non-sequential features can be enriched through the following dimensions: First, the parameter types can cover combat preference categories (such as virtual item usage tendency ratings), social relationship categories (such as friend intimacy rankings), and growth system categories (such as achievement task completion progress); Second, the parameter sources can integrate native in-game data (such as character attribute panels), third-party plugin data (such as peripheral device operation habit analysis), or player-customized configurations (such as interface layout preferences); Third, the parameter update mechanism can adopt real-time updates (such as rank ratings dynamically changing with match results), periodic snapshots (such as recording equipment library combat power at 0:00 every day), or event-triggered updates (such as resetting skill tree configuration when acquiring a new character), and this application does not make specific limitations on this.

[0190] Optionally, in this embodiment, the aforementioned account features may include, but are not limited to, composite feature vectors generated by fusing sequential and non-sequential features, which comprehensively represent the dynamic behavior patterns and static attribute configurations of a player's account. The construction process of account features involves feature concatenation, dimensionality reduction, or cross-modal fusion techniques. For example, sequential features may be extracted using an LSTM network to extract temporal latent representations, and then weighted and concatenated with the embedding vectors of non-sequential features. In typical application scenarios, account features may employ a multi-head attention mechanism to dynamically adjust the weight ratio of sequential and non-sequential features, or utilize graph neural networks to mine potential correlations between features (such as the synergistic relationship between rank rating and recent kill efficiency).

[0191] It should be noted that there are multiple technical paths for account feature fusion: First, the construction method can adopt early fusion (directly concatenating sequential and non-sequential data at the feature input layer), late fusion (extracting features through independent sub-networks and then aggregating them), or hybrid fusion (alternating fusion in a multi-layer network structure); Second, feature interaction can be designed as a cross-attention mechanism (allowing sequential features to focus on relevant non-sequential dimensions), a gated fusion network (dynamically controlling the proportion of feature information flow), or tensor decomposition (mining higher-order feature interactions); Third, feature optimization strategies can introduce contrastive learning to enhance the feature clustering of similar accounts, knowledge distillation to compress feature dimensions to reduce computational overhead, or self-supervised pre-training to mine the intrinsic correlation of features. This application does not make specific limitations on these.

[0192] Through the embodiments of this application, the refined extraction of sequential features can effectively capture the short-term fluctuations and long-term evolution patterns of player behavior, enhancing the model's sensitivity to changes in player status; the stable attribute description of non-sequential features provides the model with basic account profile support, avoiding prediction bias caused by short-term behavioral fluctuations; the multimodal fusion mechanism of account features, through dynamic weight allocation and high-order interaction modeling, realizes the coordinated expression of static attributes and dynamic behaviors, providing high-information-density feature inputs for subsequent team allocation strategy optimization.

[0193] As an optional approach, the sequence features corresponding to each account are obtained, including:

[0194] Obtain the feature matrix of size n·k generated by each account in the first n games within a preset time period, where each account in each game corresponds to k game features, and n and k are both positive integers;

[0195] The first network model is used to perform weighted fusion of the feature matrix to determine the sequence features, wherein the target second-stay prediction model includes the first network model.

[0196] Optionally, in this embodiment, the aforementioned preset time period may include, but is not limited to, a dynamic time window setting for collecting player match data, which flexibly defines the observation range for account behavior analysis according to business needs. For example, the preset time period may be set to a fixed duration (such as the last 7 days or 30 days), a dynamically adjusted activity cycle (such as automatically calculating the time span covering 80% of match behavior based on the average number of matches played per day), or an event triggering cycle (such as a specific observation period after a character version update). In specific implementations, the preset time period may be combined with a sliding window mechanism to update data (such as adding data for the current day and removing the earliest data each day), or a time decay factor may be used (such as giving 70% weight to data from the last 3 days and decreasing the weight to earlier data) to enhance the ability to capture changes in player behavior trends.

[0197] It should be noted that the rules for setting the preset time period can be expanded in the following dimensions: First, the time granularity can be subdivided into minute level (such as analyzing the peak activity within a single day), hour level (such as distinguishing between daytime and nighttime behavior patterns), or week level (such as capturing the weekend effect); Second, the dynamic adjustment mechanism can be based on player activity (such as shortening the period for highly active players to capture recent changes), game version iteration rhythm (such as resetting the period after the start of a new season), or abnormal behavior detection (such as extending the backtracking period when cheating behavior is detected); Third, boundary condition processing can adopt overlapping windows (such as sliding for 1 hour each day to maintain data continuity) or asymmetric windows (such as tracing back 3 days and reserving 1 day for the future), and this application does not make specific limitations on this.

[0198] Optionally, in this embodiment, the aforementioned n·k feature matrix may include, but is not limited to, a two-dimensional data structure composed of temporal game features, where n represents the number of games collected within a preset time period, and k represents the number of feature dimensions extracted from each game. For example, in a scenario where n=50 and k=8, the feature matrix contains 50 rows (corresponding to 50 games) and 8 columns (such as kills per game, survival time, resource contribution value, skill release frequency, movement distance, healing received, economic gain, and team ranking). The construction of this matrix must satisfy temporal integrity (e.g., arranged in the order of game occurrence time) and feature consistency (e.g., missing values ​​are filled using interpolation or mean) to ensure the stability of subsequent model processing.

[0199] It should be noted that there are several possibilities for constructing the n·k feature matrix: First, the feature types can cover combat behavior (such as kill streaks, distribution of hit parts), economic operations (such as resource conversion efficiency, equipment upgrade paths), and social interactions (such as command response delay, rescue trigger frequency); Second, the matrix generation rules can set dynamic dimensions (such as allowing k values ​​to change in different game modes), sparse compression (such as using CSR format to store matrices with a high proportion of zero values), or block processing (such as splitting a 50×8 matrix into 5 10×8 sub-matrices for parallel computation); Third, data preprocessing can introduce adversarial denoising (such as removing abnormal game records), feature scaling (such as performing logarithmic transformation on the number of kills), or cross-game correlation encoding (such as calculating the feature difference rate between adjacent games), and this application does not make specific limitations on these.

[0200] Optionally, in this embodiment, the first network model described above may include, but is not limited to, a deep learning architecture for temporal feature fusion, which generates sequence features with high-order semantics by bidirectionally extracting information from the rows (time dimension) and columns (feature dimension) of the feature matrix. For example, the first network model may use a Transformer structure, calculate the correlation weights between different games through a self-attention mechanism (such as identifying the turning point role of key games), or use a Temporal Convolutional Network (TCN) to capture the feature patterns of local time windows (such as the fluctuation pattern of resource acquisition efficiency in three consecutive games). The output of this model is a vector of length n, where each element corresponds to the weighted fused temporal feature representation, which can simultaneously retain the temporal evolution trend and key event features.

[0201] It should be noted that the technical implementation path of the first network model may include the following dimensions: First, the model architecture may use a temporal autoencoder (extracting low-dimensional temporal representations), a graph convolutional network (mining transition patterns between games), or a neural process (modeling the uncertainty of feature distribution); Second, the training strategy may adopt curriculum learning (gradually transitioning from simple game modes to complex scenarios), contrastive learning (enhancing the feature similarity of similar players' behaviors), or multi-task learning (simultaneously predicting retention and payment behavior); Third, deployment optimization may combine quantization compression (reducing model inference latency), knowledge distillation (generating lightweight student models), or dynamic pruning (dynamically adjusting the network path according to the importance of input features), and this application does not make specific limitations in this regard.

[0202] Through the embodiments of this application, the flexible setting of the preset time period can balance the needs of historical behavior analysis and real-time change response, avoiding feature information lag or noise interference caused by fixed periods; the structured storage and processing mechanism of the n·k feature matrix provides a standardized data carrier for multi-dimensional game behavior analysis, improving the scalability and maintainability of feature engineering; the temporal fusion capability of the first network model significantly enhances the semantic expressiveness of sequence features by capturing the implicit correlations and evolutionary patterns between games, providing highly discriminative input features for subsequent second-stay prediction. The overall solution constructs an automated processing link from raw data to high-order features, effectively supporting the intelligent optimization needs of team strategy in complex game environments.

[0203] As an optional approach, the retention probability of each account is determined based on account characteristics, including:

[0204] The sequence features and non-sequence features are concatenated to obtain the account features corresponding to each account;

[0205] The second network model is used to perform feature cross-referencing on account features in the same account group to determine the account retention probability for each account. The target retention prediction model includes the second network model.

[0206] Optionally, in the embodiments of this application, the above-mentioned concatenation may include, but is not limited to, a data fusion method that combines feature vectors of different dimensions into a unified feature representation according to specific rules. The purpose is to preserve the information integrity of the original features and to explore cross-modal correlations. For example, for sequential features (such as the temporal encoding vector of the movement trajectory of the last 10 games) and non-sequential features (such as the scalar value of rank scores), the concatenation operation can connect the two end-to-end in the feature dimension using tensor joins to form a high-dimensional composite vector. In specific implementations, concatenation may be accompanied by dimension alignment processing (such as embedding non-sequential features to the same dimension as sequential features) or the introduction of delimiters (such as inserting special markers to distinguish the source of features) to improve the robustness of subsequent model processing.

[0207] It should be noted that there are several possible paths for the technical implementation of the stitching operation: First, the stitching dimension can be selected as feature-dimensional stitching (horizontally expanding feature length), batch-dimensional stitching (vertically expanding sample size), or channel-dimensional stitching (constructing multi-view features); Second, the stitching timing can be set as static stitching in the preprocessing stage (such as offline feature engineering), dynamic stitching in the model input layer (real-time computation), or progressive stitching in the intermediate layers of the network (staged fusion); Third, the stitching optimization strategy can introduce feature selection mechanisms (filtering low-information features), adversarial regularization (reducing feature redundancy), or attention-weighted stitching (dynamically adjusting feature contribution), and this application does not make specific limitations on these.

[0208] Optionally, in the embodiments of this application, the aforementioned second network model may include, but is not limited to, a deep learning architecture for capturing inter-account synergy, which quantifies the change in retention probability of a single account affected by team configuration by analyzing the interaction relationships of account features within the same team. For example, the second network model may adopt a deep cross-network (DCN) structure, calculating the combined weights between different account features through explicit feature cross-layers (such as the synergistic gain between player A's sniper hit rate and player B's battlefield marker frequency), or dynamically allocating the importance weights of cross-account features using a multi-head attention mechanism. The output layer of this model typically uses a sigmoid function to map cross-features to the 0-1 interval, generating the account's second-day retention probability.

[0209] It should be noted that the structural design of the second network model can be extended to the following directions: First, the basic architecture can be a wide-depth model (balancing memory and generalization), a dual-tower structure (separating the processing of individual and group features), or a capsule network (modeling hierarchical relationships of features); Second, the feature interaction method can be designed as a factorization machine (explicit second-order crossover), a Transformer (global dependency modeling), or a neural decision tree (rule-based crossover); Third, the training optimization strategy can adopt curriculum learning (gradually transitioning from simple crossover patterns to complex relationships), contrastive learning (enhancing the crossover feature discrimination of similar teams), or adversarial training (improving the robustness of the model to noisy features), and this application does not make specific limitations on this.

[0210] Optionally, in the embodiments of this application, the aforementioned feature cross-referencing may include, but is not limited to, techniques for mining nonlinear relationships between features of different accounts through mathematical operations or neural network structures. Its essence is modeling the interaction between individual player attributes and the overall team configuration. For example, for two accounts in the same team, feature cross-referencing can calculate their complementary combat styles through vector outer products, or aggregate the social interaction features of adjacent nodes through graph neural networks. In complex scenarios, feature cross-referencing may introduce gating mechanisms (such as controlling the information transmission ratio) or residual connections (preserving original feature information) to balance the learning depth and stability of the model.

[0211] It should be noted that the modeling method for feature crossover can be enhanced in the following dimensions: First, the crossover granularity can be refined to the field level (e.g., the product of kills and healing), the behavioral pattern level (e.g., the synergy coefficient between aggressive and defensive players), or the semantic level (analyzing the implicit associations of player tags through natural language processing technology); Second, the crossover scope can be limited to intra-team local crossover (only calculating the relationship between accounts within the same team), global crossover (analyzing potential cross-team associations), or hierarchical crossover (progressive processing of intra-team crossover first and then cross-team crossover); Third, the evaluation of crossover effects can introduce interpretability analysis (e.g., quantifying feature contribution with SHAP values), dynamic importance sampling (backward filtering of key crossover items based on prediction results), or causal inference (identifying causal relationship chains between features), which are not specifically limited in this application.

[0212] Through the embodiments of this application, the splicing operation, through a multimodal feature fusion mechanism, fully preserves the original information of individual player behavior and team configuration, providing a high-fidelity data foundation for subsequent analysis; the second network model, through deep feature cross-modeling, accurately captures the nonlinear impact of player synergy and competitive relationships on retention rate, significantly improving the scenario adaptability of second-day retention probability prediction; the diverse implementation paths of feature cross-technology enable the solution to flexibly adapt to the team strategy requirements of different game types (such as team competition and social simulation), breaking through the limitations of traditional single-dimensional feature analysis, and providing scalable technical support for the dynamic optimization of intelligent matching systems.

[0213] As an alternative solution, Figure 7 This is a schematic diagram illustrating another optional account grouping method according to an embodiment of this application, such as... Figure 7 As shown, the above method also includes:

[0214] S702, determine the sample account team set and sample labels based on historical data;

[0215] S704, predict the second-time retention probability of each sample account in each sample account team in the sample account team set by using the initial second-time retention prediction model;

[0216] S706, determine the second-time retention probability of each sample account team based on the second-time retention probability of the sample account team;

[0217] S708, determine the sample strategy retention probability corresponding to the sample account team set based on the sample team retention probability;

[0218] S710 calculates the loss function based on the sample strategy retention probability and sample label, and adjusts the model parameters of the initial retention prediction model based on the loss function until the loss function meets the preset loss condition, and then determines the initial retention prediction model as the target retention prediction model.

[0219] Optionally, in this embodiment, the aforementioned sample account team set may include, but is not limited to, a historical team data set used to train the initial second-time retention prediction model. This set consists of player team combinations from past matchmaking games, with each sample team set representing a team strategy that has been actually applied. For example, the sample account team set may contain player team records from different seasons and map modes, with each record containing characteristic data of team members (such as character configuration and combat style tags) and corresponding retention results. In specific implementations, the construction of the sample set must adhere to the principles of temporal continuity (e.g., covering 30 consecutive days of matchmaking data) and balanced distribution (e.g., stratified sampling by rank) to avoid model bias caused by data skew.

[0220] It should be noted that the dimensions of the sample account team set can be expanded in the following ways: First, the time span can be selected from short-term intensive sampling (such as high-frequency battle data from the last 3 days) or long-term sparse sampling (such as cross-quarter data to capture seasonal patterns); Second, the data sources can integrate official server battle records, test server experimental data, and enhanced data generated by simulators; Third, the sampling strategy can be selected from stratified sampling (dividing levels according to rank), cluster sampling (grouping based on player behavior patterns), or active learning (prioritizing samples with high model prediction uncertainty), and this application does not make specific limitations in this regard.

[0221] Optionally, in this embodiment, the sample labels may include, but are not limited to, the actual retention status markers corresponding to each team set in the sample account team set. These markers are used to supervise the model's learning of the correlation between team-based strategies and retention results. The generation of sample labels must be based on actual player behavior logs. For example, if a player in a team logs into the game the day after forming a team, it is marked as retained (label 1); if at least one player does not log in, it is marked as not fully retained (label 0). In extended implementations, sample labels can be refined into multi-category labels (such as 7-day retention, 30-day retention) or continuous value labels (such as retention days) to adapt to the needs of different business scenarios.

[0222] It should be noted that the annotation system for sample labels can be enriched through the following dimensions: First, the label type can be subdivided into behavioral depth labels (such as whether at least one game has been completed after logging in), paid behavior labels (such as whether a recharge occurred during the retention period), or social behavior labels (such as whether new friends have been added); Second, the annotation timing can be set to immediate annotation (judged 24 hours after the end of the game) or delayed annotation (observing the behavior decay curve within 72 hours); Third, noise processing can include majority voting (cross-validation of multiple data sources), fuzzy logic (allowing partial retention status), or adversarial training (enhancing robustness to labeling errors), and this application does not make specific limitations on this.

[0223] Optionally, in this embodiment, the aforementioned sample account retention probability may include, but is not limited to, the predicted value of the retention tendency of a single player in historical samples by the initial retention prediction model, which is calculated through model forward propagation. For example, for player A in a sample team, the model outputs a retention probability of 0.85 based on their historical kill efficiency, social interaction frequency, and current team role configuration, indicating that the model believes that the player has an 85% probability of logging in the next day under a similar team strategy. The calculation of the sample account retention probability needs to be aligned with the true label distribution, and prediction bias can be reduced through probability calibration techniques (such as PlattScaling). The aforementioned sample team retention probability may include, but is not limited to, an aggregated evaluation index of the retention potential of a single sample team, which is generated by regularizing the sample account retention probabilities of all players in the team. For example, a weighted average method is used to calculate the sample team retention probability, where the team leader's account has a weight of 40%, and other members each have a weight of 20%. In complex scenarios, a synergy coefficient can be introduced. When a team contains player combinations with a historical team win rate exceeding a threshold (e.g., 70%), the aggregation result is non-linearly amplified (e.g., multiplied by a coefficient of 1.2). The above-mentioned sample strategy retention probability can include, but is not limited to, the overall effect evaluation of a specific team strategy within a sample account team set. It is derived by statistically analyzing the central tendency of the retention probabilities of all sample teams under that strategy. For example, the median retention probability of 20 sample teams under the same strategy can be taken, or its 90th percentile can be calculated to assess the upper limit potential of the strategy. The generation of sample strategy retention probabilities needs to consider the spatiotemporal limitations of strategy application, such as evaluating strategies during peak weekend periods separately to avoid confusion with weekday strategies.

[0224] Furthermore, the aforementioned loss functions can include, but are not limited to, quantitative metrics that measure the difference between the model's predicted results and the true labels, which guide the adjustment of model parameters. Commonly used loss functions include binary classification cross-entropy loss (suitable for 0 / 1 labels), mean squared error loss (suitable for continuous value labels), and contrastive loss (suitable for difference learning of similar strategies). For example, in a binary classification scenario, the loss function is calculated as the retention probability p of the strategy for the i-th sample. i and real label y i The loss value is determined using the following formula:

[0225] Loss=-Σy i *log(p i )+(1-y i )*log(1-p i );

[0226] The construction of the loss function may incorporate regularization terms (such as L2 regularization to prevent overfitting) or sample weights (such as assigning higher weights to high-value player groups). The aforementioned pre-defined loss conditions may include, but are not limited to, stopping criteria for the model training process, which define when to terminate model optimization. Pre-defined loss conditions can be set as absolute thresholds (such as the loss value falling below 0.1), relative change thresholds (such as a loss decrease of less than 1% for 5 consecutive rounds), or early stopping mechanisms (the validation set loss increases for 3 consecutive rounds). In distributed training scenarios, pre-defined loss conditions may include multi-node synchronous detection (such as half of the nodes reaching a threshold) or dynamic adjustment mechanisms (elastically extending the training rounds based on remaining computing resources).

[0227] Through the embodiments of this application, the diversity construction strategy of the sample account team set enhances the model's ability to understand the distribution of historical data and reduces the risk of overfitting due to data bias; the multi-dimensional definition of sample labels enables the model to capture the complex causes of retention behavior and improve the business guidance value of the prediction results; the incremental optimization mechanism of the initial second-time retention prediction model achieves a balance between model accuracy and training efficiency through dynamic loss function adjustment; the flexible configuration of preset loss conditions ensures the model's optimal convergence state under resource constraints. The overall solution constructs a complete closed loop from data preparation, model training to effect verification, providing a reusable technical framework for the intelligent upgrade of game matching systems.

[0228] As an optional approach, an initial retention prediction model is used to predict the retention probability of each sample account within each sample account group in the sample account group set, determining the retention probability of each sample account, including:

[0229] The initial retention prediction model is used to predict the retention probability of each sample account in each sample account team in the sample account team set, and the initial retention probability of each sample account is determined.

[0230] The probability calibration function is used to calibrate the retention probability of the initial sample accounts, and the retention probability of the sample accounts corresponding to each sample account is determined.

[0231] Optionally, in the embodiments of this application, the aforementioned probability calibration function may include, but is not limited to, a mathematical mapping function used to correct the deviation between the model's predicted probability distribution and the true label distribution. This function improves the interpretability and business applicability of the probability values ​​by fitting the relationship between the predicted probability and the empirical probability. For example, isotonic regression can be used to construct a monotonically non-decreasing function curve, correcting the initial model output probability of 0.8 to 0.75 to match the actual retention distribution; or Platt Scaling (logistic regression calibration) can be used to adjust the probability distribution through the sigmoid function. The parameters of the calibration function are typically learned based on validation set data, which can effectively alleviate the probability distortion problem caused by training data distribution shift or model overfitting.

[0232] It should be noted that there are multiple possibilities for the design and application scenarios of probabilistic calibration functions: First, the calibration function type can include parametric methods (such as Beta calibration introducing two learnable parameters), non-parametric methods (such as histogram binning calibration), or hybrid methods (using different calibration strategies in segments); Second, the calibration timing can be set to offline batch calibration (full correction of historical prediction results), online real-time calibration (dynamic adjustment based on streaming data), or phased calibration (coarse-grained first, then fine-grained); Third, the calibration target can be specific to a particular player group (such as calibrating high-ranking players separately), a specific game mode (such as calibrating ranked matches and entertainment modes separately), or a specific time period (such as calibrating independently during holiday events). This application does not make specific limitations in this regard.

[0233] Through the embodiments of this application, the basic predictive capability of the initial second-time retention prediction model provides reliable raw data input for calibration optimization, avoiding the uncontrollability of prediction results caused by directly using an untrained model; the probability calibration function corrects model bias in a data-driven manner, significantly improving the statistical consistency between the predicted probability and the actual retention distribution, and enhancing the credibility of business decisions; the diversified calibration strategy design enables the solution to flexibly cope with the data distribution characteristics of different game scenarios (such as sparse high-ranking player data and dense novice player data), reducing the risk of calibration failure due to data skew. Finally, through the synergistic effect of initial prediction and calibration optimization, a high-precision and highly robust second-time retention probability prediction system is formed, providing a quantitative basis for the generation of intelligent team allocation strategies.

[0234] As an alternative solution, Figure 8 This is a schematic diagram illustrating another optional account grouping method according to an embodiment of this application, such as... Figure 8 As shown, a genetic algorithm is used to iterate through multiple account teams N times according to the team's second-time retention probability, including:

[0235] S802, during the i-th iteration, obtain the team retention probability corresponding to each account team in each account team set;

[0236] S804, swap the first account team and the second account team in each account team set whose team retention probability satisfies the first sorting condition, and determine the account team set used in the (i+1)th round of iteration.

[0237] Optionally, in this embodiment, the first ranking condition may include, but is not limited to, a rule-based standard for screening teams with high retention potential. This standard determines candidate teams to participate in the genetic algorithm crossover operation by ranking or dividing the second-time retention probability of all teams in the account team set into quantiles. For example, the first ranking condition may be set to select the top 20% of teams with the second-time retention probability, or to select all teams whose second-time retention probability exceeds the historical average. In specific implementations, the first ranking condition may be combined with a dynamic threshold adjustment mechanism (such as gradually increasing the screening standard according to the current iteration round) or grouped hierarchical screening (such as selecting the top teams after grouping and modeling by team size).

[0238] It should be noted that the rules for setting the first sorting condition can be extended in the following dimensions: First, the sorting method can be absolute sorting (descending order by retention probability value), quantile sorting (selecting the top 25% quantiles), or sliding window sorting (comparing only adjacent strategies); Second, the selection mechanism can include tournament selection (randomly selecting several strategies to choose the best), roulette wheel selection (random selection based on probability weighting), or adaptive selection (dynamically adjusting the selection pressure according to the iteration rounds); Third, the constraints can include a limit on the number of teams (e.g., no more than 3 teams per exchange), timestamp filtering (selecting only strategies generated in the most recent 3 rounds), or load balancing conditions (excluding strategies with excessive server load). This application does not make specific limitations in this regard.

[0239] Optionally, in the embodiments of this application, the aforementioned element exchange may include, but is not limited to, the crossover operation used in genetic algorithms to generate new strategies, which explores the potential optimization space of team strategies by exchanging player combinations between different teams. For example, in the process of exchanging teams A (players 1, 2, 3, 4) and B (players 5, 6, 7, 8), players 2 and 6 may be swapped to form new teams A' (1, 6, 3, 4) and B' (5, 2, 7, 8). The specific forms of element exchange may include single-point exchange (exchanging only one player), multi-point exchange (exchanging multiple players), or block exchange (exchanging the entire role / function group).

[0240] It should be noted that there are multiple technical implementation paths for element swapping: First, the swapping targets can be limited to teams of the same size (only 4-player teams swapping with 4-player teams), teams of different sizes (3-player teams swapping with 5-player teams), or teams with matching roles (only swapping players with the same role); Second, the swapping trigger conditions can be based on player characteristic similarity (such as rank difference less than a threshold), historical synergy effect (such as team win rate higher than a benchmark value), or random probability; Third, the evaluation of swapping effects can introduce a simulation prediction mechanism (predicting changes in the second-day retention probability after the swap), a real-time rollback mechanism (revoking the swap when the effect is not as expected), or a multi-objective trade-off (simultaneously optimizing the second-day retention probability and matching time), which this application does not specifically limit.

[0241] Optionally, in this embodiment, the set of accounts used in the (i+1)th iteration can include, but is not limited to, the next-generation strategy set generated by a genetic algorithm, which inherits the high-quality strategies from the previous iteration and introduces new exploratory strategies. For example, the (i+1)th round set may contain the following three types of strategies: retaining the 10% strategies with the highest second-to-last retention probability in the previous i rounds (elite retention), 70% new strategies generated through element swapping (crossover recombination), and 20% strategies generated by random perturbation (mutation exploration). The size of this set may be dynamically adjusted (e.g., automatically expanded or contracted according to the convergence speed) or kept fixed.

[0242] It should be noted that the generation strategy of the account team set used in the (i+1)th iteration can be optimized in the following dimensions: First, the set update mechanism can include complete replacement (using the new generation strategy for the entire set), partial replacement (retaining some historical strategies), or hybrid injection (combining with an external strategy library); Second, the maintenance of set diversity can introduce niche techniques (restricting the repeated generation of similar strategies), entropy constraints (forcing the maintenance of strategy differences), or adversarial sample insertion (increasing strategy robustness); Third, the set size control can adopt a fixed capacity (e.g., always maintaining 100 strategies), elastic scaling (dynamically adjusting according to strategy quality), or phased expansion (small-scale rapid iteration in the early stage, and large-scale fine search in the later stage). This application does not make specific limitations on this.

[0243] Through the embodiments of this application, the dynamic screening mechanism of the first ranking condition effectively balances the retention of high-quality strategies and the expansion of the exploration space, avoiding the premature convergence problem caused by traditional fixed thresholds; the multi-mode operation design of element exchange enhances the algorithm's global search capability in high-dimensional strategy space, and significantly improves the generation efficiency of potential high-quality strategies by combining targeted exchange with random perturbation; the intelligent update strategy of the iterative set, by introducing elite retention and diversity maintenance mechanisms, ensures that the algorithm continuously accumulates optimization experience during the iteration process, while avoiding getting trapped in local optima. The overall solution constructs an efficient and stable strategy evolution link, providing reliable technical support for the real-time squad assignment requirements in complex game environments.

[0244] As an alternative solution, such as Figure 8 As shown, the first and second account teams in each account team set whose team retention probabilities satisfy the first sorting condition are swapped to determine the account team set used in the (i+1)th iteration, including:

[0245] Get the first account in the first account team of each account team set whose account retention probability satisfies the second sorting condition, and get the second account in the second account team of each account team set whose account retention probability satisfies the second sorting condition;

[0246] Swap the account teams to which the first and second accounts belong in each account team set to determine the third and fourth account teams. Replace the first and second account teams with the third and fourth account teams respectively to determine the account team set used in the (i+1)th iteration.

[0247] Optionally, in this embodiment, the second sorting condition may include, but is not limited to, a secondary rule for screening players with high retention potential within an account team. This rule prioritizes players within the team based on the account's second-day retention probability to determine candidate individuals for the exchange. For example, the second sorting condition may be set to select players in the top 50% of the account's second-day retention probability, or to select players whose probability value is higher than the team average. In practical applications, the second sorting condition may be combined with dynamic thresholds (such as automatically adjusting the screening ratio in each iteration) or composite conditions (such as simultaneously satisfying a probability threshold and a combat style matching degree) to accurately locate players who significantly influence the team's second-day retention probability.

[0248] It should be noted that the rules for setting the second sorting condition can be expanded in the following dimensions: First, the screening method can include quantile screening (e.g., selecting players in the top 30% quantile), sliding window screening (comparing only the probability differences between adjacent players), or clustering screening (sorting within groups based on the similarity of player characteristics); Second, the dynamic adjustment mechanism can introduce a time decay factor (e.g., lowering the screening threshold for recently active players), a load balancing factor (e.g., relaxing the player screening conditions in high-load servers), or a player grouping strategy (e.g., using different thresholds for novice players and experienced players); Third, the constraints can include role and function restrictions (e.g., only allowing players with the same role to participate in exchanges), social relationship restrictions (e.g., players with a friend relationship cannot be separated), or device performance restrictions (e.g., high-latency players cannot team up with low-latency players), which are not specifically limited in this application.

[0249] Optionally, in this embodiment, the aforementioned third and fourth account teams may include, but are not limited to, new team combinations generated by swapping selected players from the first and second account teams. These combinations inherit some characteristics of the original teams and introduce new possibilities for player collaboration. For example, the first account team originally included players A (0.8% retention rate) and B (0.7%), and the second account team included players C (0.75%) and D (0.6%). If players B and C are swapped, the third account team becomes A and C, and the fourth account team becomes B and D. The generation of the new teams must meet game rule constraints (such as player limits and role balance) and may trigger a subsequent re-evaluation of the retention rate.

[0250] It should be noted that there are multiple optimization paths for the generation logic of the third and fourth account teams: First, the team reorganization strategy can include completely random swapping (randomly selecting swap positions), heuristic swapping (prioritizing swapping players with complementary characteristics), or template-driven swapping (adjusting based on the structure of historically high-retention teams); Second, team size control can be set to a fixed size (e.g., strictly maintaining a 4-person team), a flexible size (allowing a mix of 3-5 person teams), or dynamic splitting (splitting a 5-person team into a 3-person team + a 2-person team); Third, synergy evaluation can introduce a predictive mechanism (predicting changes in the second-day retention probability of the team after the swap), a rollback mechanism (canceling the replacement if the second-day retention probability of the new team is lower than that of the original team), or multi-objective optimization (considering both the social interaction index and tactical diversity within the team). This application does not make specific limitations on these aspects.

[0251] Optionally, in the embodiments of this application, the above replacement may include, but is not limited to, the operation of removing the original team from the account team set and inserting the newly generated team. The purpose is to improve the overall retention potential of the set through policy updates. For example, in an iteration process containing a set of 100 account teams, if 20 new teams are generated through swapping, the replacement operation may eliminate the 20 old teams with the lowest retention probability, or randomly replace some old teams to maintain the stability of the set size. The design of the replacement strategy needs to balance the needs of exploration (introducing new strategies) and development (retaining high-quality strategies) to avoid premature convergence to a local optimum.

[0252] It should be noted that the execution method of the replacement operation can be enhanced in the following dimensions: First, the replacement triggering condition can be based on strategy quality (such as only replacing teams with a second-day retention probability lower than the current average), resource consumption (such as replacing inefficient strategies whose computation time exceeds a threshold), or external events (such as forcibly replacing old strategies after a game version update); Second, the replacement scope restriction can be set to local replacement (only replacing child strategies derived from the same parent strategy), global replacement (allowing arbitrary replacement across strategy sets), or hierarchical replacement (replacing secondary strategies first and then the core strategy); Third, the replacement effect evaluation can be combined with offline simulation (simulating the replacement effect on historical data), online A / B testing (gray-scale replacement of some strategies and detection of indicators), or player feedback analysis (collecting player satisfaction ratings after the replacement), and this application does not make specific limitations in this regard.

[0253] Through the embodiments of this application, the refined screening mechanism of the second sorting condition can accurately locate high-potential individual players, improving the effectiveness of exchange operations and the efficiency of strategy optimization; the generation process of the third and fourth account teams, through a dynamic reorganization strategy, significantly enhances the diversity of player synergy within the teams, avoiding the strategy homogenization problem caused by traditional fixed rules; the intelligent execution mechanism of the replacement operation, through multi-dimensional constraints and evaluations, ensures that the quality of the strategy set continues to improve during the iteration process, while maintaining a reasonable exploration space. The overall solution constructs a complete evolutionary link from individual screening to group optimization, providing technical support that balances efficiency and stability for real-time team formation needs in complex game environments.

[0254] As an optional approach, accounts are grouped based on a grouping strategy to obtain multiple sets of account teams, including: generating a first set of account teams using a random allocation rule; generating a second set of account teams using a label matching rule; merging the first and second sets of account teams to generate multiple sets of account teams; and using a genetic algorithm to iterate through the multiple sets of account teams N times according to the team retention probability, including: obtaining the team retention probability for each account team during the i-th iteration; sorting the account teams according to the team retention probability, selecting the first P account teams and the last Q account teams, where P and Q are positive integers; swapping the elements of the target account teams whose team retention probabilities satisfy the third sorting condition in the last Q account teams to determine the updated last Q account teams; and combining the updated last Q account teams with the first P account teams to determine the set of account teams used in the (i+1)-th iteration, wherein the first P account teams and the last Q account teams belong to the same strategy.

[0255] Optionally, in this embodiment, the aforementioned first type of account team set may include, but is not limited to, an initial team-assignment strategy set generated by a random allocation rule. This set uses an unbiased random algorithm to assign player accounts to different teams, exploring the diversity of team-assignment strategies. For example, in a matchmaking pool containing 100 players, the random allocation rule can divide players into 25 teams of 4 players or 20 teams of 5 players, ensuring that the team size conforms to the game rules. The generation of such a set does not rely on any prior knowledge, making it suitable for scenarios lacking historical data or requiring rapid generation of candidate strategies. It can effectively avoid the local optimum trap caused by manual rules, providing a broad exploration space for genetic algorithms.

[0256] It should be noted that the generation strategy for the first type of account team set can be expanded in the following dimensions: First, the random allocation rule can be subdivided into completely random (unconstrained uniform distribution), stratified random (random within the group after grouping by rank), or weighted random (allocating selection probability according to player activity); Second, team size control can adopt a fixed size (such as a strict 4-person team), a flexible size (allowing 3-5 person teams to coexist), or a dynamic size (adaptively adjusting according to the real-time number of players); Third, the time attribute can be set to cold start random (used when there is no historical data), periodic random (periodically reset to explore new strategies), or event-triggered random (activated when new players flood in), and this application does not make specific limitations on this.

[0257] Optionally, in this embodiment, the aforementioned tag matching rules may include, but are not limited to, grouping strategies based on player characteristic tags. These strategies assign similar or complementary players to the same team using a predefined tag system (such as combat style, social attributes, and role specialization). For example, tag matching rules can combine "assault" players and "support" players in a preset ratio, or evenly distribute highly interactive players to each team based on the "social activity" tag. Such rules can be formulated through human experience or historical data analysis, and can quickly generate team strategies that meet specific tactical objectives. However, it is necessary to avoid over-reliance on fixed rules, which could lead to insufficient strategic innovation.

[0258] It should be noted that there are several possibilities for the implementation of tag matching rules: First, the tag types can cover combat attribute tags (such as output / defense / healing), social behavior tags (such as follower / lone wolf), or economic strategy tags (such as resource hoarding / fast consumption); Second, the matching logic can be designed as homogeneous matching (teaming up similar tags), complementary matching (combining different tags), or hybrid matching (homogeneous core tags and complementary auxiliary tags); Third, the dynamic adjustment mechanism can introduce tag weight learning (automatically adjusting tag importance based on historical effects), tag timeliness decay (old tags gradually becoming invalid), or context-aware tags (dynamically switching the tag system based on map mode). This application does not make specific limitations on these aspects.

[0259] Optionally, in this embodiment, the third sorting condition may include, but is not limited to, the criteria for determining the teams to be optimized among the remaining Q account teams, which are set based on the ranking or probability value of the team's second-time retention probability. For example, the third sorting condition may be set to a second-time retention probability lower than the average of the current iteration round, or in the bottom 30% quantile range. Teams that meet the condition will be marked as inefficient strategies and need to introduce new player combinations through element swapping to improve their potential. The dynamic adjustment capability of this condition directly affects the optimization efficiency of the algorithm; for example, the screening threshold can be gradually increased in the later stages of the iteration to accelerate convergence.

[0260] It should be noted that the rules for setting the third sorting condition can be enriched through the following dimensions: First, the sorting method can be global sorting (sorting all teams uniformly), group sorting (sorting by team size or pattern), or local sorting (comparing only within the same type of strategy); Second, the selection mechanism can include hard truncation (strictly selecting the last Q teams), probability sampling (selecting teams to be optimized according to probability distribution), or clustering screening (selecting teams with low clustering of strategy features); Third, the constraints can include timestamp restrictions (only processing strategies generated in the most recent N rounds), resource consumption restrictions (excluding strategies with excessively high computational costs), or player experience protection (prohibiting frequent adjustments to high-scoring strategies), which are not specifically limited in this application.

[0261] Through the embodiments of this application, the random generation mechanism of the first type of account team set provides baseline diversity for strategy exploration, avoiding the algorithm from getting trapped in local optima too early; the label matching rule accelerates the discovery of high-quality strategies through structured strategy generation, complementing the random strategy; the dynamic filtering capability of the third sorting condition enables accurate positioning and targeted optimization of inefficient strategies, significantly improving iteration efficiency; the conflict repair mechanism ensures the legality and stability of the strategy iteration process, reducing system anomalies caused by logical errors.

[0262] As an alternative solution, Figure 9 This is a schematic diagram illustrating another optional account grouping method according to an embodiment of this application, such as... Figure 9 As shown, the target account team set is determined based on the policy retention probability corresponding to each account team set generated in N rounds of iteration, including:

[0263] When the number of iterations reaches N, the set of accounts and teams corresponding to the maximum retention probability of the strategy in the Nth iteration is selected as the target account and team set.

[0264] If the number of iterations has not reached N and the increase of the maximum value of the strategy retention probability in M ​​consecutive iterations is less than or equal to the first threshold, the iteration is terminated early and the set of account teams corresponding to the maximum value of the strategy retention probability in the current iteration is selected as the target account team set, where M is a positive integer less than N.

[0265] If the number of iterations has not reached N and the maximum value of the second-time retention probability of the strategy is greater than or equal to the second threshold, the iteration is terminated early and the set of accounts and teams corresponding to the maximum value of the second-time retention probability of the strategy in the current iteration is selected as the target set of accounts and teams.

[0266] Optionally, in this embodiment, the number of iterations may include, but is not limited to, a threshold for the number of optimization rounds performed by the genetic algorithm, which is used to control the balance between algorithm runtime and computational resource consumption. The setting of the number of iterations needs to comprehensively consider the policy convergence speed and system real-time requirements. For example, it may be set to 10-20 rounds in real-time matching scenarios, while in offline policy optimization it can be extended to 50-100 rounds. The dynamic adjustment mechanism for the number of iterations may combine policy quality assessment (such as the rate of increase in second-retention probability) or resource detection (such as CPU / memory utilization). For example, when it is detected that the time consumed in a single iteration exceeds a threshold, the total number of iterations is automatically reduced.

[0267] It should be noted that the strategy for controlling the number of iterations can be extended through the following dimensions: First, the dynamic adjustment mechanism can be based on convergence speed indicators (such as the rate of reduction of the variance of the second-retention probability), resource utilization (such as forced termination when memory usage exceeds 80%), or external events (such as iteration reset triggered by hot updates of the game version); Second, the execution mode can be designed as full iteration (strictly completing N rounds), elastic iteration (dynamically adjusting within the range of N rounds), or segmented iteration (dividing N rounds into multiple sub-stages for separate optimization); Third, the interruption recovery mechanism can include breakpoint continuation (saving intermediate states for subsequent continuation), strategy snapshot (periodically storing copies of the optimal strategy), or asynchronous execution (continuing background calculations without affecting real-time services), and this application does not make specific limitations on this.

[0268] Optionally, in this embodiment, the maximum value of the second-time retention probability of the aforementioned strategy may include, but is not limited to, the highest evaluation score of all account team sets in the current iteration round, which reflects the estimated effect of the optimal team allocation strategy explored by the algorithm. For example, in a set of 50 account teams, if the second-time retention probability of a certain set reaches 0.78, then this value is the maximum value in the current round. The maximum value may fluctuate and increase with the iteration process, or it may stagnate due to getting stuck in a local optimum, requiring a termination condition to determine the effectiveness of the optimization process.

[0269] It should be noted that there are several possible application scenarios for the maximum second-retention probability of a strategy: First, the benchmark can be set as the global historical maximum (across all iteration rounds), the sliding window maximum (comparing only the most recent M rounds), or the cluster maximum (comparing by strategy type); Second, the optimization direction can include single-objective optimization (pursuing only the maximum value increase), multi-objective optimization (considering both strategy diversity and computational cost), or constraint optimization (increasing the maximum value while meeting the upper limit of matching time); Third, the verification mechanism can introduce A / B testing (online verification of the actual effect of the maximum value strategy), adversarial attack testing (detecting the robustness of the maximum value strategy), or player surveys (collecting subjective experience ratings), and this application does not make specific limitations in this regard.

[0270] Optionally, in the embodiments of this application, the aforementioned first threshold may include, but is not limited to, an incremental threshold used to determine when the strategy optimization process has stalled. This threshold is defined as the upper limit of the relative increase in the maximum value of the strategy's second-time retention probability over M consecutive iterations. For example, if the first threshold is set to 0.5%, and the maximum increase in the value over 5 consecutive iterations does not exceed this threshold, the algorithm is considered unable to effectively improve the strategy quality. The setting of the first threshold needs to consider the model's prediction accuracy (such as the minimum effective unit of change in the second-time retention probability) and business sensitivity (such as the minimum acceptable optimization magnitude), and is usually determined through historical data statistics or simulation experiments.

[0271] Optionally, in this embodiment, the second threshold may include, but is not limited to, the absolute target value of the strategy's second-day retention probability, which reflects the business's minimum expected standard for the effectiveness of the team-based strategy. For example, if the second threshold is set to 0.75, then when a strategy with a second-day retention probability ≥ 0.75 is generated in a certain iteration, it is considered that the business objective has been achieved and the optimization is terminated early. The second threshold may be dynamically adjusted according to the game's operational stage. For example, a lower threshold may be set during the new user growth period to quickly deploy strategies, while the threshold may be increased during the mature period to pursue the ultimate effect.

[0272] It should be noted that the threshold setting rules can be enriched in the following dimensions: First, the calculation method of the first threshold can be an absolute difference (e.g., a fixed 0.5%), a relative proportion (e.g., an increase of 1.2% compared to the previous round), or a dynamic adjustment value (automatically calculated based on historical volatility); Second, the triggering logic of the second threshold can be designed as a static threshold (fixed value), a dynamic threshold (fluctuating with the number of active players), or a composite threshold (requiring the strategy's second-day retention probability ≥ X and the player complaint rate ≤ Y); Third, the threshold linkage mechanism can include tiered triggering (e.g., lowering the second threshold after reaching the first threshold to accelerate convergence), conditional mutual exclusion (allowing only one termination condition to take effect), or weight superposition (combining multiple conditions to calculate the termination priority), and this application does not specifically limit this.

[0273] Optionally, in the embodiments of this application, the aforementioned early termination of iteration may include, but is not limited to, an operational mechanism that interrupts the genetic algorithm optimization process when specific conditions are met. This mechanism is used to balance the marginal benefits of computational resource consumption and strategy quality improvement. For example, when the algorithm reaches the second threshold within 10 rounds, or when the improvement in optimization effect is less than the first threshold for three consecutive rounds, the iteration is terminated and the current optimal strategy is output. The early termination mechanism needs to be accompanied by exception handling logic (such as recording the reason for interruption) and result verification procedures (such as performing offline simulation verification on the final strategy).

[0274] Through the embodiments of this application, the dynamic control mechanism for the number of iterations significantly improves resource utilization efficiency and reduces redundant computation while ensuring strategy quality; the multi-dimensional comparison and verification mechanism of the maximum second-retention probability of the strategy enhances the reliability and business adaptability of the optimal strategy; the flexible configuration of the threshold system supports differentiated needs under different operational scenarios, such as quickly obtaining feasible strategies during the launch of new features and finely optimizing long-term effects during stable operation; the intelligent decision-making capability to terminate iterations early avoids falling into ineffective computation loops by real-time detection of the marginal benefits of the optimization process. The overall solution constructs a flexible and scalable optimization termination system, providing key technical support for balancing the real-time performance and strategy quality of the game matching system.

[0275] in, Figure 10 This is a schematic diagram illustrating another optional account grouping method according to an embodiment of this application, such as... Figure 10 The diagram illustrates the interrelationships between the various processes in this application.

[0276] The following specific examples will further explain this application:

[0277] This application proposes a player teaming scheme based on a genetic algorithm, applicable to team-based multiplayer competitive games, overcoming the limitation of existing schemes that cannot be directly optimized based on player retention. This application first initializes multiple player teaming strategies. For each strategy, a retention prediction model predicts the average retention of all teams, using this average retention as the evaluation metric. Then, a genetic algorithm is used to adjust the teaming scheme by swapping players between teams for each strategy, forming a new strategy. The average retention of teams is used as the metric to eliminate poor strategies and retain good ones. Through multiple iterations of the genetic algorithm, the optimal teaming strategy is finally obtained. Because the average retention of teams is used as the optimization metric, the final teaming strategy can improve player retention.

[0278] This technical solution primarily serves team-based competitive multiplayer games, where player teams are a crucial component that directly impacts the overall gaming experience.

[0279] 1) Multiplayer games usually emphasize teamwork. A reasonable team division plan can help players find teammates with similar styles, making teamwork smoother.

[0280] 2) A good team formation plan can promote interaction between players, thereby enhancing the social experience;

[0281] 3) A well-planned team composition can create diverse tactical combinations, enhancing the game's fun. A good team composition can improve the player's in-game experience, thereby optimizing player retention.

[0282] Therefore, the main purpose of this application is to find the optimal team composition scheme that maximizes the overall retention of all players in a game through multiple iterations of a genetic algorithm, thereby optimizing the player experience and retention.

[0283] Figure 11 This is a schematic diagram illustrating another optional account grouping method according to an embodiment of this application, such as... Figure 11 The diagram shows the overall framework of this application, which includes, but is not limited to, the following parts:

[0284] 1) Squad strategy initialization: Initialize some player squad strategies as the basis for subsequent optimization. Each squad strategy will correspond to an in-game squad result.

[0285] 2) Day 2 Retention Prediction Model: This model is used to predict the day 2 retention rate of all players in a team under the current team division strategy. This allows us to calculate the average day 2 retention rate of all player teams in the game as an evaluation metric for the team division strategy, and use this metric as the target for optimization of the genetic algorithm.

[0286] 3) Genetic Algorithm Iteration: The team division strategy is continuously adjusted through the genetic algorithm to eventually generate the optimal solution.

[0287] The following is a detailed introduction to each:

[0288] 1) Squad strategy initialization:

[0289] The team composition strategy initialization phase requires generating team composition strategies for subsequent optimization. There are two initialization methods: random initialization and custom rule initialization. Random initialization randomly divides players into different teams, suitable for scenarios where no data analysis of the team composition scheme has been performed in the early stages, and there are no prior team composition rules. If preliminary analysis of the team composition rules has been conducted in the early stages, and it is desired that the team composition scheme can be adjusted based on the existing one, then the custom rule initialization scheme can be used.

[0290] 2) Secondary retention prediction model:

[0291] The retention prediction model is used to predict the retention rate of players within a team. Figure 12This is a schematic diagram of another optional account grouping method according to an embodiment of this application, the structure of which is as follows: Figure 12 As shown.

[0292] The model's input consists of features from all players within the same team. These player features are divided into two parts: non-sequential features and sequential features. Sequential features are the characteristics of a player's previous n games that day, including match rankings, kill counts, assist counts, and other match-related features. Assuming each game has k features, these k features from the n games can be represented as an n×k sequence based on the chronological order of the matches. For sequential features, a transformer model is used. Simply put, this involves using a self-attention mechanism to weight and fuse the n-length sequential features, ultimately obtaining a fused sequential feature representation. Non-sequential features include a player's current level, rank, match tags, etc. Match tags are labels related to player matchmaking, such as the player's playstyle and tactical coordination.

[0293] The features input to the model are processed using a deep neural network (DNN) to obtain the player's second-day retention prediction. The second-day retention prediction model needs to be trained before application. Here, a supervised training method is used, where the training samples and labels are the features of the historical player teams and their corresponding player retention labels, respectively. A standard binary classification loss function is constructed based on the model's prediction results and the retention labels to guide the model's learning. Assume that for the i-th player, the DNN model predicts a second-day retention probability of p. i (The value range is [0,1]), and its actual retention label is y. i (The value can only be 0 or 1), then the loss function is:

[0294] Loss=-Σy i *log(p i )+(1-y i )*log(1-p i );

[0295] Using gradient descent based on this loss function, the learnable parameters in the model are iterated and eventually converged to obtain the player second-time retention prediction model. This model is then used to predict player second-time retention, and the average second-time retention of all players within the team is calculated to obtain the team's average second-time retention.

[0296] Since this application needs to determine the merits of the team allocation strategy based on the average second-retention rate of each team, it requires obtaining the most accurate second-retention probability prediction possible. However, DNN models do not directly optimize the accuracy of the output probability during training; instead, they optimize the classification loss function. Furthermore, the training samples are often sampled, resulting in distribution differences compared to the original samples. This can lead to inconsistencies between the model's output probability values ​​and the actual class distribution. To obtain the most accurate second-retention probability possible, this application employs Isotonic Regression for probability calibration. Isotonic Regression is a non-parametric regression method that uses a monotonically non-decreasing function to fit the data, thereby achieving probability calibration.

[0297] 3) Genetic Algorithm:

[0298] The optimization process of the squad strategy in this application adopts a scheme based on genetic algorithms. Genetic algorithms are optimization and search algorithms based on the principles of natural selection and genetics. Their basic idea is to improve the quality of the solution generation by generation by simulating the process of biological evolution and through operations such as selection, crossover, and mutation. This application mainly adopts the mutation operation.

[0299] Specifically, each initial team strategy can be viewed as an individual in the genetic algorithm. The average second-time retention of players within a team can be calculated using a retention prediction model, thus calculating the average second-time retention of all teams corresponding to each team strategy. This average retention is equivalent to the fitness score of the genetic algorithm; strategies with higher fitness scores have a higher probability of being retained in the next generation. In each iteration of the genetic algorithm, a mutation operation is used. For each strategy, a new strategy is formed by swapping players in two teams within the current strategy's team. When selecting players for swapping, this application uses a heuristic method. First, based on the average second-time retention of each team under the current strategy, team 1 (where players need to be swapped) is sampled. The lower the average second-time retention of a team, the higher the probability of it being sampled. Then, based on the second-time retention of players in team 1, player A (where players have a lower second-time retention) is sampled. Next, player B is sampled from the list of swappable players; again, the lower the second-time retention, the higher the probability of it being sampled. Finally, swapping A and B completes one mutation operation.

[0300] Assume there are n strategies in each generation. Through mutation, n new strategies are generated. Based on the average retention rate of the teams under each strategy, only n strategies are retained in the next generation for the next operation. After k iterations, the optimal team allocation strategy in the k-th iteration is selected as the final team allocation strategy.

[0301] After the team division scheme proposed in this application was launched in the game, it achieved significantly better results than the original online optimal rule scheme that had been tested many times:

[0302] Comparison of different solutions Day 1 retention rate Related solutions 65.39％ This application 65.59％

[0303] Table 1

[0304] As shown in Table 1, online experiments demonstrate that the user retention rate using this application is higher than the optimal related solutions tested in previous online experiments. To ensure online response time, the number of initial strategies and the number of genetic generations were configured very conservatively. Theoretically, increasing the number of initial strategies and the number of genetic iterations could achieve better results.

[0305] It is understood that in the specific embodiments of this application, data such as user information are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0306] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0307] According to another aspect of the embodiments of this application, an account grouping apparatus for implementing the above-described account grouping method is also provided. For example... Figure 13 As shown, the device includes:

[0308] Grouping module 1302 is used to group accounts based on a grouping strategy to obtain a set of multiple account teams, wherein one account team in the set of multiple account teams includes multiple accounts determined according to the grouping strategy;

[0309] The prediction module 1304 is used to predict multiple account teams using a target retention prediction model, and to determine the strategy retention probability corresponding to each account team set. The target retention prediction model is used to predict the team retention probability of each account team in each account team set, and to determine the strategy retention probability based on the team retention probability. The target retention prediction model includes a model trained on an initial retention prediction model using sample account team sets and sample labels. The sample labels are used to indicate the retention status of the sample account team sets.

[0310] Processing module 1306 is used to perform N rounds of iteration on multiple account team sets according to the team second retention probability using a genetic algorithm. In the i-th round of iteration, the strategy second retention probability corresponding to each account team set generated in the i-th round of iteration is determined by the target second retention prediction model. Here, N and i are positive integers, and i≤N.

[0311] The generation module 1308 is used to determine the target account team set based on the policy retention probability corresponding to each account team set generated by N rounds of iteration, wherein the target account team set represents the account team set whose policy retention probability satisfies the preset condition.

[0312] As an optional solution, the above-mentioned device is used to predict multiple account teams using a target retention prediction model in the following manner to determine the strategy retention probability corresponding to each account team: predicting each account in each account team in the multiple account teams using the target retention prediction model to determine the account retention probability corresponding to each account; determining the team retention probability corresponding to each account team based on the account retention probability; and determining the strategy retention probability corresponding to each account team based on the team retention probability.

[0313] As an optional approach, the aforementioned device is used to predict the retention probability of each account in each account team within a set of multiple account teams by using a target retention prediction model in the following manner: obtaining the account features corresponding to each account in each account team; and determining the retention probability of each account based on the account features.

[0314] As an optional solution, the above-mentioned device is used to obtain the account characteristics corresponding to each account in each account team in the following manner: obtaining the sequence characteristics corresponding to each account, wherein the sequence characteristics are used to represent the game data of the account within a preset time period; obtaining the non-sequence characteristics corresponding to each account, wherein the non-sequence characteristics are used to represent the account parameters of the account; and determining the account characteristics based on the sequence characteristics and non-sequence characteristics.

[0315] As an optional solution, the above-mentioned device is used to obtain the sequence features corresponding to each account in the following manner: obtaining a feature matrix of size n·k generated by the first n games of each account within a preset time period, wherein each account in each game corresponds to k game features, and n and k are both positive integers; using a first network model to perform weighted fusion of the feature matrix to determine the sequence features, wherein the target retention prediction model includes the first network model.

[0316] As an optional approach, the aforementioned device is used to determine the account retention probability corresponding to each account based on account features in the following manner: concatenating sequential features and non-sequence features to obtain the account features corresponding to each account; using a second network model to perform feature cross-referencing on the account features in the same account group to determine the account retention probability corresponding to each account, wherein the target retention prediction model includes the second network model.

[0317] As an optional solution, the above-mentioned device is also used to: determine the sample account team set and sample labels based on historical data; predict each sample account in each sample account team in the sample account team set using an initial retention prediction model, and determine the sample account retention probability corresponding to each sample account; determine the sample team retention probability corresponding to each sample account team based on the sample account retention probability; determine the sample strategy retention probability corresponding to the sample account team set based on the sample team retention probability; calculate the loss function based on the sample strategy retention probability and sample labels, and adjust the model parameters of the initial retention prediction model based on the loss function until the loss function meets the preset loss condition, and determine the initial retention prediction model as the target retention prediction model.

[0318] As an optional approach, the aforementioned device is used to predict the retention probability of each sample account in each sample account team within the sample account team set using an initial retention prediction model in the following manner: predicting the retention probability of each sample account in each sample account team within the sample account team set using an initial retention prediction model to determine the initial retention probability of each sample account; calibrating the initial retention probability of the sample account using a probability calibration function to determine the retention probability of each sample account.

[0319] As an optional approach, the above-mentioned device is used to perform N rounds of iteration on multiple account team sets according to the team second retention probability using a genetic algorithm in the following manner: during the i-th round of iteration, the team second retention probability corresponding to each account team in each account team set is obtained; the first account team and the second account team in each account team set whose team second retention probability satisfies the first sorting condition are swapped to determine the account team set used in the (i+1)-th round of iteration.

[0320] As an optional approach, the aforementioned apparatus is used to determine the set of account teams used in the (i+1)th iteration by swapping the first account team and the second account team in each account team set whose retention probability satisfies the first sorting condition in the following manner: obtaining the first account in the first account team of each account team set whose retention probability satisfies the second sorting condition and the second account in the second account team of each account team set whose retention probability satisfies the second sorting condition; swapping the account teams to which the first account and the second account belong in each account team set to determine the third account team and the fourth account team; and using the third account team and the fourth account team to replace the first account team and the second account team respectively to determine the set of account teams used in the (i+1)th iteration.

[0321] As an optional solution, the above-mentioned device is used to group accounts based on a grouping strategy to obtain multiple sets of account teams in the following manner: generating a first set of account teams using a random allocation rule; generating a second set of account teams using a tag matching rule; merging the first and second sets of account teams to generate multiple sets of account teams; the above-mentioned device is used to perform N rounds of iteration on the multiple sets of account teams according to the team retention probability using a genetic algorithm in the following manner: during the i-th iteration, obtaining the team retention probability corresponding to each account team; sorting the account teams according to the team retention probability, and selecting the first P account teams and the last Q account teams, where P and Q are both positive integers; swapping the elements of the target account teams whose team retention probability satisfies the third sorting condition in the last Q account teams to determine the updated last Q account teams; combining the updated last Q account teams with the first P account teams to determine the set of account teams used in the (i+1)-th iteration, wherein the first P account teams and the last Q account teams belong to the same strategy.

[0322] As an optional approach, the aforementioned device is used to determine the target account team set based on the policy retention probability corresponding to each account team set generated by N iterations in the following manner: when the number of iterations reaches N, the account team set corresponding to the maximum policy retention probability in the Nth iteration is selected as the target account team set; when the number of iterations has not reached N, and the increase in the maximum policy retention probability in M ​​consecutive iterations is less than or equal to a first threshold, the iteration is terminated early and the account team set corresponding to the maximum policy retention probability in the current iteration is selected as the target account team set, where M is a positive integer less than N; when the number of iterations has not reached N, and the maximum policy retention probability is greater than or equal to a second threshold, the iteration is terminated early and the account team set corresponding to the maximum policy retention probability in the current iteration is selected as the target account team set.

[0323] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0324] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0325] According to one aspect of this application, a computer program product is provided, the computer program product comprising a computer program.

[0326] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0327] Figure 14 A schematic block diagram of a computer system architecture for implementing an electronic device according to embodiments of the present application is shown.

[0328] It should be noted that, Figure 14 The computer system 1400 of the electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0329] like Figure 14 As shown, the computer system 1400 includes a central processing unit (CPU) 1401, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 1402 or programs loaded from storage section 1408 into random access memory (RAM). The RAM 1403 also stores various programs and data required for system operation. The CPU 1401, ROM 1402, and RAM 1403 are interconnected via a bus 1404. An input / output interface 1405 (I / O interface) is also connected to the bus 1404.

[0330] The following components are connected to the input / output interface 1405: an input section 1406 including a keyboard, mouse, etc.; an output section 1407 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 1408 including a hard disk, etc.; and a communication section 1409 including a network interface card such as a local area network card, modem, etc. The communication section 1409 performs communication processing via a network such as the Internet. A drive 1410 is also connected to the input / output interface 1405 as needed. A removable medium 1411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 1410 as needed so that computer programs read from it can be installed into the storage section 1408 as needed.

[0331] Specifically, according to embodiments of this application, the processes described in the various method flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 1409, and / or installed from removable medium 1411. When the computer program is executed by central processing unit 1401, it performs various functions defined in the system of this application.

[0332] In such an embodiment, the computer program can be downloaded and installed from a network via communication section 1409, and / or installed from removable media 1411. When the computer program is executed by central processing unit 1401, it performs various functions provided in the embodiments of this application.

[0333] According to another aspect of the embodiments of this application, an electronic device for implementing the above-described account grouping method is also provided, the electronic device being... Figure 1 The terminal device or server shown. This embodiment uses this electronic device as an example for illustration. Figure 15 As shown, the electronic device includes a memory 1502 and a processor 1504. The memory 1502 stores a computer program, and the processor 1504 is configured to execute the steps of any of the above method embodiments via the computer program.

[0334] Optionally, in this embodiment, the aforementioned electronic device may be located in at least one of a plurality of network devices in a computer network.

[0335] Optionally, in this embodiment, the processor may be configured to execute the methods in the embodiments of this application via a computer program.

[0336] Alternatively, as those skilled in the art will understand, Figure 15 The structure shown is for illustrative purposes only. Figure 15 This does not limit the structure of the aforementioned electronic devices. For example, the electronic device may also include components that are more... Figure 15 The more or fewer components shown (such as network interfaces, etc.), or having the same Figure 15 The different configurations shown.

[0337] The memory 1502 can be used to store software programs and modules, such as the program instructions / modules corresponding to the account grouping method and device in this embodiment. The processor 1504 executes various functional applications and data processing by running the software programs and modules stored in the memory 1502, thereby implementing the aforementioned account grouping method. The memory 1502 may include high-speed random access memory and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 1502 may further include memory remotely located relative to the processor 1504, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. Specifically, the memory 1502 may be used, but is not limited to, to store information such as account data. As an example, such as... Figure 15 As shown, the memory 1502 may include, but is not limited to, the grouping module 1302, prediction module 1304, processing module 1306, and generation module 1308 in the account grouping device. Furthermore, it may include, but is not limited to, other module units in the account grouping device, which will not be described further in this example.

[0338] Optionally, the transmission device 1506 described above is used to receive or send data via a network. Specific examples of the network described above may include wired networks and wireless networks. In one example, the transmission device 1506 includes a Network Interface Controller (NIC), which can be connected to other network devices and a router via a network cable to communicate with the Internet or a local area network. In another example, the transmission device 1506 is a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0339] In addition, the aforementioned electronic device also includes: a display 1508 for displaying the aforementioned account group information; and a connection bus 1510 for connecting the various module components in the aforementioned electronic device.

[0340] In other embodiments, the aforementioned terminal device or server can be a node in a distributed system, wherein the distributed system can be a blockchain system, which is a distributed system formed by connecting multiple nodes through network communication. The nodes can form a peer-to-peer network, and any form of computing device, such as a server, terminal, or other electronic device, can become a node in the blockchain system by joining this peer-to-peer network.

[0341] According to one aspect of this application, a computer-readable storage medium is provided, wherein a processor of an electronic device reads computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the electronic device to perform the account grouping method provided in various alternative implementations of the account grouping aspect described above.

[0342] Optionally, in this embodiment, the computer-readable storage medium described above may be configured to store methods for performing the embodiments of this application.

[0343] Optionally, in this embodiment, those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0344] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0345] If the integrated units in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause one or more electronic devices to execute all or part of the steps of the methods described in the various embodiments of this application.

[0346] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0347] In the several embodiments provided in this application, it should be understood that the disclosed application can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of 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 couplings, direct couplings, or communication connections may be through some interfaces; indirect couplings or communication connections between units or modules may be electrical or other forms.

[0348] 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.

[0349] Furthermore, the functional units in the various embodiments of this application 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.

[0350] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for grouping accounts, characterized in that, include: Accounts are grouped according to a grouping strategy to obtain multiple account teams. Each account team in one of the multiple account teams includes multiple accounts determined according to the grouping strategy. The multiple account team sets are predicted using a target retention prediction model to determine the strategy retention probability corresponding to each account team set. The target retention prediction model is used to predict the team retention probability of each account team in each account team set, and to determine the strategy retention probability based on the team retention probability. The target retention prediction model includes a model trained on an initial retention prediction model using sample account team sets and sample labels. The sample labels are used to indicate the retention status of the sample account team sets. The genetic algorithm is used to iterate the multiple account team sets for N rounds according to the team second retention probability. In the i-th round of iteration, the strategy second retention probability corresponding to each account team set generated in the i-th round of iteration is determined by the target second retention prediction model. N and i are positive integers, and i ≤ N. The target account team set is determined based on the strategy retention probability corresponding to each account team set generated by N rounds of iteration, wherein the target account team set represents the account team set whose strategy retention probability satisfies a preset condition.

2. The method according to claim 1, characterized in that, The step of predicting the retention probability of the multiple account teams using a target retention prediction model to determine the policy retention probability for each account team team includes: The target retention prediction model is used to predict the retention probability of each account in each account team in the multiple account team sets. The team retention probability corresponding to each account team is determined based on the account retention probability. The strategy retention probability corresponding to each account's team set is determined based on the team retention probability.

3. The method according to claim 2, characterized in that, The step of predicting the retention probability of each account in each account group within the multiple account group sets using the target retention prediction model includes: Obtain the account characteristics corresponding to each account in each account team; The retention probability of each account is determined based on the account characteristics.

4. The method according to claim 3, characterized in that, The step of obtaining the account characteristics corresponding to each account in each account team includes: Obtain the sequence features corresponding to each account, wherein the sequence features are used to represent the game data of the account within a preset time period; Obtain the non-sequence feature corresponding to each account, wherein the non-sequence feature is used to represent the account parameters of the account; The account features are determined based on the sequence features and the non-sequence features.

5. The method according to claim 4, characterized in that, The step of obtaining the sequence features corresponding to each account includes: Obtain the feature matrix of size n·k generated by each account in the first n games within the preset time period, where each account corresponds to k game features in each game. n and k are both positive integers; The first network model is used to perform weighted fusion on the feature matrix to determine the sequence features, wherein the target second-stay prediction model includes the first network model.

6. The method according to claim 4, characterized in that, Determining the retention probability of each account based on the account characteristics includes: The sequence features and the non-sequence features are concatenated to obtain the account features corresponding to each account; The second network model is used to perform feature cross-referencing on account features in the same account group to determine the account retention probability corresponding to each account, wherein the target retention prediction model includes the second network model.

7. The method according to claim 1, characterized in that, The method further includes: The sample account team set and the sample tags are determined based on historical data; The initial retention prediction model is used to predict the retention probability of each sample account in each sample account team in the sample account team set. The retention probability of each sample account team is determined based on the retention probability of the sample account team. The second-time retention probability of the sample strategy corresponding to the sample account team set is determined based on the second-time retention probability of the sample team. The loss function is calculated based on the second-time retention probability of the sample strategy and the sample label. The model parameters of the initial second-time retention prediction model are adjusted based on the loss function until the loss function meets the preset loss condition. The initial second-time retention prediction model is then determined as the target second-time retention prediction model.

8. The method according to claim 7, characterized in that, The step of predicting the second-time retention probability of each sample account in each sample account team in the sample account team set using the initial second-time retention prediction model includes: The initial retention prediction model is used to predict the retention probability of each sample account in each sample account team in the sample account team set, thereby determining the initial retention probability of each sample account. The retention probability of the initial sample account is calibrated using a probability calibration function to determine the retention probability of the sample account corresponding to each sample account.

9. The method according to claim 1, characterized in that, The step of using a genetic algorithm to iterate through the set of multiple account teams N times according to the second-time retention probability of the teams includes: During the i-th iteration, the retention probability of each account team in each account team set is obtained; Swap the elements of the first account team and the second account team in each account team set whose team retention probability satisfies the first sorting condition, and determine the account team set used in the (i+1)th round of iteration.

10. The method according to claim 9, characterized in that, The step of swapping elements between the first account team and the second account team in each account team set whose team retention probabilities satisfy the first sorting condition, to determine the account team set used in the (i+1)th iteration, includes: Get the first account in the first account team of each account team set whose account retention probability satisfies the second sorting condition, and get the second account in the second account team of each account team set whose account retention probability satisfies the second sorting condition; Swap the account teams to which the first and second accounts belong in each account team set to determine the third and fourth account teams. Replace the first and second account teams with the third and fourth account teams respectively to determine the account team set used in the (i+1)th iteration.

11. The method according to claim 1, characterized in that, The process of grouping accounts based on a grouping strategy to obtain multiple account team sets includes: generating a first type of account team set through a random allocation rule; generating a second type of account team set through a tag matching rule; and merging the first type of account team set and the second type of account team set to generate the multiple account team sets. The step of using a genetic algorithm to iterate through the set of multiple account teams N times according to the team retention probability includes: in the i-th iteration, obtaining the team retention probability corresponding to each account team; sorting the account teams according to the team retention probability, and selecting the first P account teams and the last Q account teams, where P and Q are both positive integers; swapping the elements of the target account teams in the last Q account teams whose team retention probability meets the third sorting condition to determine the updated last Q account teams; combining the updated last Q account teams with the first P account teams to determine the set of account teams used in the (i+1)-th iteration, wherein the first P account teams and the last Q account teams belong to the same strategy.

12. The method according to claim 1, characterized in that, The determination of the target account team set based on the retention probability of the strategy corresponding to each account team set generated by N rounds of iteration includes: When the number of iterations reaches N, the set of accounts and teams corresponding to the maximum retention probability of the strategy in the Nth iteration is selected as the target account and team set. If the number of iterations has not reached N, and the increase in the maximum value of the second retention probability of the strategy in M ​​consecutive iterations is less than or equal to the first threshold, the iteration is terminated early and the set of account teams corresponding to the maximum value of the second retention probability of the strategy in the current iteration is selected as the target account team set, where M is a positive integer less than N. If the number of iterations has not reached N and the maximum value of the second retention probability of the strategy is greater than or equal to the second threshold, the iteration is terminated early and the set of accounts corresponding to the maximum value of the second retention probability of the strategy in the current iteration is selected as the target account set.

13. An account grouping device, characterized in that, include: The grouping module is used to group accounts based on a grouping strategy to obtain multiple account teams, wherein one account team in one of the multiple account teams includes multiple accounts determined according to the grouping strategy; The prediction module is used to predict the multiple account team sets through a target retention prediction model, and determine the strategy retention probability corresponding to each account team set. The target retention prediction model is used to predict the team retention probability of each account team in each account team set, and determine the strategy retention probability based on the team retention probability. The target retention prediction model includes a model obtained by training an initial retention prediction model using sample account team sets and sample labels. The sample labels are used to indicate the retention status of the sample account team sets. The processing module is used to perform N rounds of iteration on the multiple account team sets according to the team second retention probability using a genetic algorithm. In the i-th iteration, the strategy second retention probability corresponding to each account team set generated in the i-th iteration is determined by the target second retention prediction model. N and i are both positive integers, and i ≤ N. The generation module is used to determine the target account team set based on the strategy retention probability corresponding to each account team set generated in N rounds of iteration, wherein the target account team set represents the account team set whose strategy retention probability satisfies a preset condition.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein the program, when executed by a processor, performs the method described in any one of claims 1 to 12.

15. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method described in any one of claims 1 to 12.

16. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to execute the method described in any one of claims 1 to 12 through the computer program.

Images