An esport match prediction method based on a graph neural network
By constructing a heterogeneous graph of esports competitions and introducing a multi-layer attention mechanism, the problems of insufficient information fusion and poor interpretability in esports competition prediction are solved, achieving efficient and accurate competition prediction and tactical analysis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SUZHOU UNIV
- Filing Date
- 2026-03-20
- Publication Date
- 2026-06-05
AI Technical Summary
Existing esports prediction methods struggle to effectively integrate multi-source heterogeneous information and ignore structural dependencies and semantic relationships between entities, resulting in poor prediction accuracy and a lack of interpretability.
A graph neural network-based approach is used to construct a heterogeneous graph of e-sports competitions. Through node-level, semantic-level, and group-level attention mechanisms, the attention weights between nodes and their neighbors are calculated, the contribution weights of different meta-paths to the competition results are learned, and information is fused. Finally, the prediction results are output through a classifier.
It significantly improves the accuracy and interpretability of match predictions, captures key factors in match outcomes, provides intuitive tactical insights, and supports match analysis and decision-making.
Smart Images
Figure CN122153458A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of knowledge graph technology, specifically to a method for predicting esports matches based on graph neural networks. Background Technology
[0002] With the rapid development of the esports industry, especially the popularity of multiplayer online battle arena (MOBA) games (such as League of Legends and Honor of Kings), esports match prediction has become an important requirement for data analysis, event operation, and betting decisions. Traditional match prediction methods mainly rely on statistical regression models (such as logistic regression and support vector machines) or time series models (such as recurrent neural networks). These methods are usually based on single-dimensional historical data (such as team win rate, player KDA, hero selection rate, etc.), making it difficult to effectively integrate multi-source heterogeneous match information.
[0003] Esports competitions involve various entity types (such as teams, players, and heroes) and their complex interactions (such as players using heroes, teams including players, and team matchups), forming a naturally heterogeneous graph structure. Traditional prediction methods often process this information as independent feature vectors, ignoring the structural dependencies and semantic relationships between entities. This makes it difficult for models to capture the complex tactical coordination and adversarial relationships behind the outcome of a match. For example, information such as a player's performance when using a specific hero or the team's collaborative efficiency under a specific lineup is difficult to effectively represent and integrate in traditional feature engineering.
[0004] In recent years, graph neural networks have demonstrated powerful capabilities in processing graph-structured data, providing new insights into modeling complex relationships in esports. However, most existing graph neural network-based prediction methods are designed for homogeneous graphs, making them difficult to directly apply to heterogeneous graph scenarios in esports where multiple entity types coexist. Furthermore, existing methods often treat all neighboring nodes and relationship paths equally when processing heterogeneous nodes, failing to distinguish the varying importance of different players, heroes, or tactical paths to the match outcome. This leads to noise during information aggregation and insufficient semantic understanding, thus affecting prediction accuracy.
[0005] Furthermore, existing esports prediction models generally suffer from poor interpretability, making it difficult to explain the basis of predictions to coaches, analysts, or viewers, such as "why a certain team is more likely to win" or "which factors have the greatest impact on the outcome." This lack of interpretability limits the model's application value in practical scenarios such as tactical analysis and decision support. Summary of the Invention
[0006] Purpose of the invention: The purpose of this invention is to provide an e-sports competition prediction method based on graph neural networks, which solves the problems of insufficient fusion of multimodal heterogeneous information such as players, heroes, and teams, imprecise structural modeling, and poor predictive interpretability in existing e-sports competition prediction methods.
[0007] Technical Solution: The present invention provides an esports match prediction method based on graph neural networks, comprising the following steps:
[0008] (1) Construct a heterogeneous graph of e-sports competitions, which includes multiple types of nodes and multiple relational edges connecting the nodes. The multiple types of nodes include at least team nodes, player nodes and hero nodes.
[0009] (2) Extract and assign initial feature vectors to various nodes in the heterogeneous graph;
[0010] (3) Through the node-level attention mechanism, for the preset meta-path, the attention weight between the central node and its neighboring nodes is calculated, and the features of the neighboring nodes are aggregated according to the attention weight to obtain the semantic specific representation of each node under the meta-path.
[0011] (4) Through the semantic level attention mechanism, the contribution weight of different meta-paths to the competition results is learned, and the semantic specific representation of the same node under different meta-paths is fused according to the contribution weight to obtain the fused representation of the node.
[0012] (5) Construct node groups, taking the two opposing teams and their associated player nodes and hero nodes as a node group respectively. Learn the interaction weights between the two node group representations through the group-level attention mechanism, and fuse them according to the interaction weights to obtain the final confrontation representations of the two opposing teams.
[0013] (6) Input the adversarial representation into the classifier and output the prediction result of the match outcome.
[0014] Furthermore, in step (3), the node-level attention mechanism is as follows: for each meta-path, the features of different types of nodes are projected onto the same semantic space through a type-specific transformation matrix, and the attention coefficients of the central node and its neighboring nodes are calculated in the semantic space to distinguish the importance of different neighboring nodes to the central node.
[0015] Furthermore, in step (4), the semantic attention mechanism specifically involves: summarizing the semantic-specific representations of all nodes under a single meta-path, learning the global importance weight of that meta-path, and achieving adaptive fusion of different semantic information carried by multiple meta-paths.
[0016] Furthermore, in step (5), the group-level attention mechanism is specifically as follows: each of the opposing teams is regarded as a node group containing its internal players and heroes. After the internal representations of the two node groups are initially fused, the mutual influence between the groups is learned through the attention mechanism to explicitly model the adversarial relationship in the game.
[0017] Furthermore, during model training and prediction, the attention weights generated by one or more layers of node-level attention mechanisms, semantic-level attention mechanisms, or group-level attention mechanisms are output to explain the key factors affecting the prediction results of the competition.
[0018] The present invention discloses an e-sports match prediction system based on graph neural networks, comprising:
[0019] Graph construction module: used to construct a heterogeneous graph of e-sports competitions, which includes various types of nodes and various relational edges connecting the nodes. The various types of nodes include at least team nodes, player nodes and hero nodes.
[0020] Feature initialization module: used to extract and assign initial feature vectors to various nodes in the heterogeneous graph;
[0021] Node-level attention module: It is used to calculate the attention weight between the central node and its neighboring nodes for a preset meta-path through a node-level attention mechanism, and aggregate the features of the neighboring nodes based on the attention weight to obtain the semantic-specific representation of each node under the meta-path.
[0022] Semantic Attention Module: Used to learn the contribution weights of different meta-paths to the competition results through a semantic attention mechanism, and to fuse the semantic-specific representations of the same node under different meta-paths using the contribution weights to obtain the fused representation of the node.
[0023] Group-level attention module: used to construct node groups, taking the two opposing teams and their associated player nodes and hero nodes as a node group respectively. The interaction weights between the representations of the two node groups are learned through the group-level attention mechanism, and then fused according to the interaction weights to obtain the final confrontation representation of the two opposing teams.
[0024] Prediction output module: This module takes the adversarial representation as input to the classifier and outputs the prediction results of the match outcome.
[0025] An electronic device according to the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when loaded onto the processor, implements any of the methods described herein.
[0026] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any of the methods described herein.
[0027] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: Improved prediction accuracy and scientific decision-making: By introducing a multi-layer attention mechanism, this invention can effectively integrate the complex heterogeneous relationships between teams, players, and heroes, capturing key factors in determining the outcome of a match. Experiments show that on the League of Legends professional match dataset, the prediction accuracy of this method is approximately 12% higher than traditional statistical models and approximately 7% higher than time series models based on recurrent neural networks, significantly improving the scientific rigor and reliability of match predictions and providing strong support for match analysis, tactical formulation, and betting decisions.
[0028] Enhanced Feature Fusion and Semantic Understanding Capabilities: Through node-level attention mechanisms, the model can distinguish the differences in importance of different players and heroes to the team's performance; through semantic-level attention mechanisms, the model can automatically identify the degree of influence of different meta-paths (such as "team-player-hero" and "team-hero-team") on the match results; through group-level attention mechanisms, the model can further fuse the overall competitive information of the two teams. This multi-layered, fine-grained attention fusion mechanism significantly improves the model's ability to understand complex match semantics.
[0029] Enhancing model interpretability and tactical analysis value: The attention weights output by this method can be used to interpret the basis of predictions, such as showing which players or heroes have the greatest impact on the outcome, which tactical paths are more decisive, and which team collaboration patterns are more effective. This provides coaching teams, commentators, and viewers with intuitive tactical insights, enhancing the model's practical value and credibility.
[0030] Achieving efficient training and real-time prediction: By employing graph-structured representation and attention-based parallel computation, this method achieves high training and inference efficiency while maintaining prediction accuracy. On a single RTX 3090 GPU, the model can be trained quickly on historical season data (training 10,000 matches in approximately 2 hours) and supports dynamic predictions during real-time matches (single prediction time < 50ms), meeting the needs of scenarios such as live event broadcasts and real-time data dashboards.
[0031] Support for cross-game and cross-event migration: The heterogeneous graph modeling framework of this method has good generalization ability. By adjusting the node type and meta-path definition, it can be quickly adapted to different e-sports projects (such as "DOTA 2" and "Honor of Kings") or different event systems (such as professional leagues and international invitational tournaments), reducing the cost of repeated modeling and improving the applicability of the method.
[0032] Reduce reliance on manual feature engineering: Traditional esports prediction methods heavily rely on expert experience for feature design and selection. This method automatically learns node representations and relation weights through graph neural networks, significantly reducing manual intervention and improving the level of modeling automation and system robustness.
[0033] Promoting Data-Driven Decision-Making and Intelligent Esports Development: This method provides the esports industry with a systematic tool for extracting knowledge from massive amounts of competition data and assisting in decision-making. It helps to promote the development of esports training, tactical analysis, and event operation towards data-driven and intelligent directions, and has significant social and economic value. Attached Figure Description
[0034] Figure 1 This is a schematic diagram of the model structure of the present invention;
[0035] Figure 2 This is a flowchart of the present invention. Detailed Implementation
[0036] The technical solution of the present invention will be further described below with reference to the accompanying drawings.
[0037] like Figure 2 As shown, this embodiment of the invention provides an esports match prediction method based on graph neural networks, including the following steps:
[0038] S1. Construct a heterogeneous graph of e-sports competitions, defining three types of nodes: team nodes, player nodes, and hero nodes, as well as various relationship edges between them, such as "team-contains-player", "player-uses-hero", "team-selects-hero", etc.
[0039] S2. Extract feature vectors for each type of node: Team node features may include historical win rate, economic data, team collaboration indicators, etc.; Player node features may include individual KDA, hero pool depth, recent performance, etc.; Hero node features may include skill type, positioning, version strength, etc.
[0040] S3. Employ a node-level attention mechanism to calculate the attention weight between a node and its neighbors for each meta-path (such as "team-player-hero"), and distinguish the importance of different neighbors to the central node.
[0041] S4. Employ a semantic-level attention mechanism to learn weights for multiple meta-paths (such as "team-player-team", "team-hero-team", "team-player-hero-player-team", etc.) to capture the contribution of different semantic paths to the competition results.
[0042] S5. Introduce a node group-level attention mechanism, treating the two opposing teams as two node groups, learning the node representations within the group and further performing inter-group attention fusion to enhance the team's overall representation discrimination ability.
[0043] S6. After multi-layer attention fusion, the vector representations of the two teams are obtained, input into the classifier to calculate the win / loss probability, and output the prediction result.
[0044] S7. An end-to-end training method is adopted, and the cross-entropy loss function is used for optimization. Attention weights are used to provide prediction interpretability.
[0045] The specific implementation process of step S1 above is as follows:
[0046] S11, Define Heterogeneous Maps , where node type mapping function {team, player, hero}, edge type mapping function {Contains, Uses, Selects, Matchups};
[0047] S12. Construct a graph based on historical match data. Each match corresponds to a subgraph, which includes nodes of two teams, nodes of players from both sides, and nodes of the heroes they selected.
[0048] S13. Assign an initial feature vector to each node. The features are derived from structured statistical data or embedded representations.
[0049] The specific implementation process of step S3 above is as follows:
[0050] S31. For each metapath Extract the meta-path neighbor set for each node.
[0051] S32. Project the features of various node types to the same semantic space using a type-specific transformation matrix:
[0052]
[0053] S33. Calculate the attention coefficient between node i and its neighbor j:
[0054]
[0055] in This represents vector concatenation. A learnable attention vector;
[0056] S34. Normalize attention weights using softmax:
[0057]
[0058] S35. Weighted aggregation of neighbor features yields a semantically specific representation of node i under this meta-path:
[0059] The specific implementation process of step S4 above is as follows:
[0060] S41. Suppose there are P meta-paths, resulting in P sets of semantically specific node representations.
[0061] S42. Learn the importance weights of each meta-path through semantic-level attention:
[0062]
[0063] in These are learnable parameters;
[0064] S43. Normalization yields the meta-path weights:
[0065]
[0066] S44. Weighted Fusion Multi-Semantic Representation:
[0067]
[0068] The specific implementation process of step S5 above is as follows:
[0069] S51. Define the two opposing teams as groups GA and GB, with each group containing player nodes and hero nodes for that team.
[0070] S52. Perform intra-group representation fusion on each group of nodes to obtain a preliminary group representation vector;
[0071] S53. Through a group-level attention mechanism, the interaction weights between the two sets of representations are learned and further fused to obtain the final team confrontation representation.
[0072] S54. Input the two sets of final representations into a multilayer perceptron for classification and output the win / loss probabilities.
[0073] The specific implementation process of step S7 above is as follows:
[0074] S71, Let the training labels be... Indicates the win or loss, and the predicted probability is: ;
[0075] S72. Using the binary cross-entropy loss function:
[0076]
[0077] S73. Update all attention mechanism and transformation matrix parameters through backpropagation;
[0078] S74. During training, node-level, semantic-level, and group-level attention weights can be output to provide interpretability for prediction decisions.
[0079] Taking professional matches of "Honor of Kings" as an example, the specific implementation includes the following steps: By constructing a heterogeneous graph composed of three types of nodes—"team," "player," and "hero"—and combining a three-layer attention mechanism at the node, semantic, and group levels, multimodal information is integrated to achieve accurate prediction of the outcome of the match. The specific implementation process is as follows:
[0080] S1. Construct and initialize the heterogeneous graph of the "Honor of Kings" competition.
[0081] S11, such as Figure 1 As shown, a heterogeneous graph is constructed for each professional "Honor of Kings" match. The node types include: Team, Player, and Hero. Relationship types include: contain (team → player), use (player → hero), pick (team → hero), and vs (team ↔ team). Metapaths are predefined to better integrate node information on the heterogeneous graph, including: "team-player-hero", "player-hero-player", "hero-player-hero", etc.
[0082] S12. Assign an initial feature vector to each node:
[0083] Team node characteristics: historical win rate, average economy per game, number of towers destroyed, team fight win rate, etc.
[0084] Player characteristics include KDA, average damage per game, economic conversion rate, and hero pool breadth.
[0085] Hero node characteristics: positioning (tank / assassin / mage, etc.), skill type, version win rate, ban rate, etc.
[0086] S13. All features are standardized, and at the same time... Figure 2 As shown, initialize the projection matrix. Mapping to a unified feature space using a type-specific projection matrix.
[0087] S2, Node-level attention computation, such as Figure 2 As shown in the node-level attention diagram, the process includes projecting node features, calculating node attention weights, and finally weighted aggregation to obtain the final fusion result of neighbor information for a given node.
[0088] S21. Project the original features of the three types of nodes onto a unified semantic space using the projection matrix Mϕ:
[0089]
[0090] in Nodes of the same type share the same projection matrix.
[0091] S22, such as Figure 1 As shown in (a), the attention weights between nodes and their neighbors are calculated for each meta-path.
[0092] S23. Taking the meta-path "team-player-hero" as an example:
[0093] For a given team node TA, its neighbors include players. Weights are calculated using an attention mechanism, such as: The semantic representation of the team under this path is obtained after weighted aggregation.
[0094] S24. Similarly, calculate other meta-paths such as "team-hero-team" and "player-hero-player" to obtain semantically specific embedding matrices.
[0095] S25, such as Figure 1 As shown in the node-level attention process, each meta-path performs the above calculations for team A and team B respectively, resulting in two sets of semantically specific embeddings:
[0096] Team A:
[0097] Team B:
[0098] S3, Semantic Attention Fusion
[0099] S31, such as Figure 1 As shown in (b), attention fusion is achieved by weighting the importance of multiple meta-paths.
[0100] S32. For team A, its multiple meta-path representation The weights of each path are calculated using semantic-level attention:
[0101]
[0102] S33, such as Figure 2 As shown, after performing semantic-level attention calculations on each team, a weighted fusion is obtained to obtain the semantic-level embedding vector ZA for team A. Similarly, the semantic-level embedding vector ZB for team B is obtained.
[0103] S4, Node-level Attention Fusion
[0104] S41, such as Figure 1 As shown in (c), the two teams are regarded as two node groups, and attention fusion between the groups is performed.
[0105] S42, semantic embeddings ZA and ZB of team A and team B are input into the group-level attention module to calculate inter-group interaction weights:
[0106]
[0107] S43. After weighted fusion, the final adversarial representation vector Z is obtained.
[0108] S5, Prediction and Training, such as Figure 2 As shown, the final embedding is obtained after node group attention calculation. The loss is then calculated based on the final embedding and the target task.
[0109] S51, such as Figure 1 As shown in (d), the final representation is that Z is input to a multilayer perceptron for classification:
[0110]
[0111] in This indicates the probability that team A will win.
[0112] S52. Training using the binary cross-entropy loss function: Calculate the loss based on the predicted probability and the true label.
[0113]
[0114] S53. Use the loss to backpropagate and update all learnable parameters in the network.
[0115] S54, such as Figure 2 As shown, after backpropagation, it is determined whether the loss value of the final model is in a convergent state. If it is converged, the training ends and the model's parameter weights are saved. If it is not converged, the above training steps S2-S5 are continued until the model converges.
[0116] In summary, this invention achieves efficient modeling and accurate prediction of complex heterogeneous relationships in "Honor of Kings" matches through a multi-level attention mechanism, demonstrating strong practical value and interpretability.
Claims
1. A method for predicting esports matches based on graph neural networks, characterized in that, Includes the following steps: (1) Construct a heterogeneous graph of e-sports competitions, which includes multiple types of nodes and multiple relational edges connecting the nodes. The multiple types of nodes include at least team nodes, player nodes and hero nodes. (2) Extract and assign initial feature vectors to various nodes in the heterogeneous graph; (3) Through the node-level attention mechanism, for the preset meta-path, the attention weight between the central node and its neighboring nodes is calculated, and the features of the neighboring nodes are aggregated according to the attention weight to obtain the semantic specific representation of each node under the meta-path. (4) Through the semantic level attention mechanism, the contribution weight of different meta-paths to the competition results is learned, and the semantic specific representation of the same node under different meta-paths is fused according to the contribution weight to obtain the fused representation of the node. (5) Construct node groups, taking the two opposing teams and their associated player nodes and hero nodes as a node group respectively. Learn the interaction weights between the two node group representations through the group-level attention mechanism, and fuse them according to the interaction weights to obtain the final confrontation representations of the two opposing teams. (6) Input the adversarial representation into the classifier and output the prediction result of the match outcome.
2. The e-sports match prediction method based on graph neural networks according to claim 1, characterized in that, In step (3), the node-level attention mechanism is as follows: for each meta-path, the features of different types of nodes are projected onto the same semantic space through a type-specific transformation matrix, and the attention coefficients of the central node and its neighboring nodes are calculated in the semantic space to distinguish the importance of different neighboring nodes to the central node.
3. The e-sports match prediction method based on graph neural networks according to claim 1, characterized in that, In step (4), the semantic attention mechanism is specifically as follows: sum up the semantic-specific representations of all nodes under a single meta-path, learn the global importance weight of the meta-path, and realize the adaptive fusion of different semantic information carried by multiple meta-paths.
4. The e-sports match prediction method based on graph neural networks according to claim 1, characterized in that, In step (5), the group-level attention mechanism is as follows: each of the opposing teams is regarded as a node group containing its internal players and heroes. After the internal representations of the two node groups are initially fused, the mutual influence between the groups is learned through the attention mechanism to explicitly model the adversarial relationship in the game.
5. The e-sports match prediction method based on graph neural networks according to claim 1, characterized in that, During model training and prediction, the attention weights generated by one or more layers of node-level attention mechanisms, semantic-level attention mechanisms, or group-level attention mechanisms are output to explain the key factors affecting the prediction results of the competition.
6. An eSports match prediction system based on graph neural networks, characterized in that, include: Graph construction module: used to construct a heterogeneous graph of e-sports competitions, which includes various types of nodes and various relational edges connecting the nodes. The various types of nodes include at least team nodes, player nodes and hero nodes. Feature initialization module: used to extract and assign initial feature vectors to various nodes in the heterogeneous graph; Node-level attention module: It is used to calculate the attention weight between the central node and its neighboring nodes for a preset meta-path through a node-level attention mechanism, and aggregate the features of the neighboring nodes based on the attention weight to obtain the semantic-specific representation of each node under the meta-path. Semantic Attention Module: Used to learn the contribution weights of different meta-paths to the competition results through a semantic attention mechanism, and to fuse the semantic-specific representations of the same node under different meta-paths using the contribution weights to obtain the fused representation of the node. Group-level attention module: used to construct node groups, taking the two opposing teams and their associated player nodes and hero nodes as a node group respectively. The interaction weights between the representations of the two node groups are learned through the group-level attention mechanism, and then fused according to the interaction weights to obtain the final confrontation representation of the two opposing teams. Prediction output module: This module takes the adversarial representation as input to the classifier and outputs the prediction results of the match outcome.
7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is loaded into the processor, it implements the method described in any one of claims 1-5.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-5.