Deep Learning for Football Outcomes Prediction based on Football Rating System

Abstract

The prediction models of association football can be categorized into three natural clusters, which are the statistical models, the machine learning and probabilistic graphical models and rating systems. The prediction may focus on matches outcomes prediction (win, draw and lose) or the number of goals scored obtained by home and away team. Match result prediction is very important as a benchmark for the management team to assess the chances of winning, drawing and losing for a particular team before the match starts. In the 2017 soccer challenge, the conventional machine learning and ensemble methods, as well as probabilistic graphical models such as k-Nearest Neighbor (k-NN), Gradient Boosting Tree (GBT) and Bayesian Networks (BN) have been a choice of researchers that participate the challenge and successfully dominate the challenge. Even so, there are some well-known techniques such as Neural Networks (NN) and Deep Neural Networks (DNN) that are left in this challenge although this technique has advantages in producing good results in many fields including solving real world problems which the results are comparable and sometimes superior to expert knowledge in some cases. In this paper, we propose a football match outcome prediction based on pi-rating system using TabNet, a DNN architecture for tabular data. The experiment cover two parts namely: (1) generates pi-rating system from 216, 743 instances of raw football dataset and (2) predicts 206 football match outcomes using TabNet. As a result, the proposed prediction model based on the pi-rating system using TabNet successfully overcomes the existing model for football match outcomes for both scoring rules, in terms of accuracy and Rank Probability Score (RPS). The findings from this study are important because they can be used for future researchers in developing new football match outcome prediction models that incorporate several new features along with other features. This is important to improve the predictive performance of football prediction models using other advanced techniques.