[Paper Review] Efficient Feature Representations for Cricket Data Analysis using Deep Learning based Multi-Modal Fusion Model

Souridas Alaka, R Sreekumar+1 more|arXiv (Cornell University)|Aug 16, 2021

Sports Analytics and Performance9 references4 citations

TL;DR

This paper proposes a deep learning-based multi-modal fusion framework that learns adaptive, contrastive loss-optimized embeddings for cricket players and teams using IPL data. By jointly training player embeddings and integrating pre-match pitch reports via BERT, the model achieves 95% accuracy in predicting overall run-rate classes—dramatically outperforming cross-entropy baseline models (22.5% accuracy), demonstrating the efficacy of contrastive representation learning for sports analytics.

ABSTRACT

Data analysis has become a necessity in the modern era of cricket. Everything from effective team management to match win predictions use some form of analytics. Meaningful data representations are necessary for efficient analysis of data. In this study we investigate the use of adaptive (learnable) embeddings to represent inter-related features (such as players, teams, etc). The data used for this study is collected from a classical T20 tournament IPL (Indian Premier League). To naturally facilitate the learning of meaningful representations of features for accurate data analysis, we formulate a deep representation learning framework which jointly learns a custom set of embeddings (which represents our features of interest) through the minimization of a contrastive loss. We base our objective on a set of classes obtained as a result of hierarchical clustering on the overall run rate of an innings. It's been assessed that the framework ensures greater generality in the obtained embeddings, on top of which a task based analysis of overall run rate prediction was done to show the reliability of the framework.

Motivation & Objective

• Address the lack of effective, inter-related feature representations for players and teams in cricket analytics.
• Overcome data scarcity in sports representation learning by using a siamese network with contrastive loss.
• Improve run-rate prediction accuracy by jointly learning player embeddings and incorporating pre-match pitch reports.
• Demonstrate that meaningful, generalizable embeddings enhance downstream prediction tasks in sports data analysis.
• Evaluate the impact of multimodal inputs (player features + pitch reports) on predictive performance in T20 cricket.

Proposed method

• Apply hierarchical clustering to overall run rates to define discrete classes for contrastive learning.
• Train a siamese network with contrastive loss to learn meaningful, low-dimensional embeddings for players and teams.
• Use a separate branch to encode pre-match pitch reports via a fine-tuned BERT model into dense embeddings.
• Concatenate player embeddings and pitch embeddings as input to a final prediction head for run-rate class prediction.
• Train the prediction model using cross-entropy or contrastive loss, with frozen embedding weights during fine-tuning.
• Perform 5-fold cross-validation with fixed test sets per run-rate class to ensure robust evaluation.

Experimental results

Research questions

• RQ1 Can contrastive loss-based representation learning produce more generalizable and informative embeddings for cricket players and teams compared to standard cross-entropy training?
• RQ2 How does incorporating pre-match pitch reports affect the accuracy of run-rate prediction in T20 cricket?
• RQ3 To what extent do learned embeddings improve the performance of downstream prediction tasks in data-scarce sports analytics?
• RQ4 What is the relative contribution of player embeddings versus pitch report embeddings in predicting overall run-rate classes?
• RQ5 Can multimodal fusion of player features and textual pitch reports lead to more robust and accurate run-rate predictions?

Key findings

• The contrastive representation learning framework achieved 95% accuracy in predicting overall run-rate classes when pitch report data was included, compared to only 22.5% with cross-entropy baseline.
• The model with contrastive embeddings and pitch report input showed a 72.5 percentage point improvement in accuracy over the cross-entropy baseline.
• Including pitch report embeddings increased prediction accuracy by 5% (from 90% to 95%) when contrastive embeddings were used, indicating the value of multimodal input.
• The cross-entropy baseline model performed poorly (22.5% accuracy) due to ineffective player embedding learning, highlighting the limitations of standard classification objectives in data-scarce settings.
• Confusion matrices revealed that intermediate run-rate classes were more prone to misclassification, likely due to insufficiently representative input features.
• The framework demonstrated robustness and generalization, with 90% accuracy even without pitch data, underscoring the quality of the contrastive embeddings.