Multitasking Neural Network for Cricket Team Selection with Complex Dynamics

Supervised multitasking machine learning is a prominent approach for predicting complex contexts where several target variables need to be addressed simultaneously. Predicting the team composition in cricket is such a problem, where the ratio of batsmen to bowlers to all-rounders is determined concomitantly, yet existing research studies have only treated the problem of composition prediction as a simple machine learning task. This study advocates and implements a novel Multitasking Neural Network (MNN) model to facilitate selection of team composition in cricket. The dataset, consisting of 323 records, was created using web scraping from multiple cricket data repositories and weather data sites, focusing on Team Sri Lanka. After preprocessing, feature engineering, and splitting the data into an 80:20 ratio for training and testing, four MNN models were developed. Each model accepted ten input dimensions and predicted three outputs simultaneously. The models were evaluated based on test precision and test loss, with the model using the mean squared logarithmic error (MSLE) as the loss function outperforming others with a precision of 76% and a test loss of 1. Integrating MSLE reduces the impact of data imbalance within the target classes while according to domain experts’ recommendations, this performance is reasonably acceptable. Future work will aim to generalize the model for other cricket teams, addressing data imbalance by increasing the size of the dataset.