KABADDI Analytics

Kabaddi is a contact team sport of Indian-origin. It is a highly strategic game and generates a significant amount of data due to its rules. However, data generated from kabaddi tournaments has so far been unused, and coaches and players rely heavily on intuitions to make decisions and craft strategies. This paper provides a quantitative approach to the game of kabaddi. The research derives outlook from an analysis performed on data from the 3rd Standard-style Kabaddi World Cup 2016, organised by the International Kabaddi Federation. The dataset, which consists of 66 entries over 31 variables from 33 matches, was manually curated. This paper discusses and provides a quantitative perspective on traditional strategies and conceptions related to the game of kabaddi such as attack and defence strategies. Multiple hypotheses are built and validated using student’s t-test. This paper further provides a quantitative approach to profile an entire tournament to gain a general understanding of the strengths of various teams. Additionally, team-specific profiling, through hypotheses testing and visualisation, is presented to gain a deeper understanding of team’s behaviour and performance. This paper also provides multiple models to forecast the winner. The model-building includes automatic feature selection techniques and variable importance analysis techniques. Generalised linear model with and without an elastic net, recursive partitioning and regression tree, conditional inference tree, random forest, support vector machine (linear and radial) and neural network-based models are built and presented. Ensemble models use generalised linear model and random forest model techniques as ensemble method to combine outcome of a generalised linear model with the elastic net, random forest, and neural network-based models. The research discusses the comparison between models and their performance parameters. Research also suggests that ensemble technique is not able to boost up accuracy. Models achieve 91.67%-100% accuracy on cross-validation dataset and 78.57%-100% on test set. Results presented can be used to design in-game real-time winning predictions to improve decision-making. Results presented can be used to design agent and environments to train artificial intelligence via reinforced learning model.