Axial Transformer Predictions for Real-Time Possession-Based Sports
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing solutions fail to accurately predict player and team performance in possession-based sporting events due to inadequate capture of correlations between teammates, opponents, and contextual features, particularly in real-time scenarios.
Innovation Solution
A transformer neural network, specifically an axial transformer neural network, is employed to generate predictions by incorporating a super feature embedding layer that accounts for unique characteristics of possession-based sports, such as lineup changes, and applies self-attention with an autoregressive mask to handle temporal and spatial interactions between players and teams.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing prediction solutions are used, then the system is simple to implement, but the prediction accuracy is insufficient due to inadequate capture of correlations between teammates, opponents, and contextual features
Solution Approach 1:
The model segments the prediction task into multiple independent prediction heads, each targeting specific metrics (e.g., shots, goals, passes, tackles) for different entities (player, team, match). This segmentation allows each head to specialize in capturing specific correlation patterns without overwhelming the entire model, thereby improving prediction accuracy for each metric while managing overall complexity through modular architecture.
Solution Approach 2:
The patent introduces a temporal dimension by making predictions at multiple time-steps throughout the match rather than a single endpoint prediction. This dimensional expansion allows the model to capture dynamic correlations that evolve over time, such as changing team formations, player fatigue effects, and momentum shifts, significantly improving prediction accuracy while the autoregressive approach manages the complexity of temporal dependencies.
2Measurement precision
If real-time predictions are generated at multiple time-steps, then the prediction granularity and usefulness are improved, but the computational latency increases
Solution Approach 1:
The model performs preliminary processing by embedding all input features (player statistics, team strength, contextual factors) at the beginning of the match. These embeddings are stored and reused at each subsequent time-step prediction, avoiding redundant computation. This preliminary action significantly reduces latency for real-time predictions while maintaining high granularity, as the model only needs to update predictions based on current game state rather than reprocessing all inputs.
Solution Approach 2:
The patent dynamically adjusts prediction parameters based on the current time-step and game state. Instead of using fixed prediction intervals or uniform attention weights, the model adapts parameters such as the number of future time-steps to predict, the scope of contextual features to consider, and the confidence thresholds for predictions. This flexibility optimizes the balance between prediction granularity and computational latency by allocating more computational resources to critical moments in the match.
3Loss of information
If the model incorporates multiple input tensors representing various game features, then the comprehensiveness of contextual information is improved, but the input processing complexity increases
Solution Approach 1:
The patent merges multiple input tensors (player features, team features, contextual features, historical statistics) into a unified feature representation through concatenation and shared embedding layers. This merging process consolidates information from diverse sources into a cohesive input structure that the transformer model can process efficiently. By combining rather than separately processing each input type, the model achieves comprehensive information coverage while reducing the overall processing complexity through unified architecture components.
Solution Approach 2:
The input processing architecture uses universal components that handle multiple types of features simultaneously. For example, the embedding layers and transformer encoder serve as multi-functional blocks that process different feature types (categorical, numerical, temporal) through the same computational pathway. This universality allows the model to incorporate comprehensive contextual information from various sources while avoiding the need for separate specialized processing pipelines for each feature type, thereby managing input processing complexity.
Data Source
AI summary
A method of generating a set of predictions associated with a possession-based sporting event using an axial transformer neural network, the method including: receiving an input tuple, including a set of tensors representing game context, team strength, player strength, live team features, live player features, game events, and a super feature; inputting the input tuple into an axial transformer neural network by inputting each tensor from the set of tensors within a corresponding initial embedding layer; concatenating the initial embedding layers to form a single tensor; applying self-attention to the single tensor; mapping output embeddings from the axial transformer layers to target layers, each of the output embeddings being of a dimension of a target metric; and generating a set of target metric predictions for each of a set of players, one or more teams, and a match, based on the output embeddings from the target layers.