Research On The Theory And Application By Applying Machine Learning To Player Ratings In Football Games

Football,as the world's first sports,creats huge econimic benefits every year.As one of the most important part of football,player ratings can be applied to many other scenarios of football field,such as match result prediction and evaluating player values in the transfer market.Thus the accurate and scientific evaluation system is urgent needed for the player ratings.Machine learning can be a good solution for this problem.Using machine learning to get player ratings is an interdisciplinary research involving comprehensive evaluation and machine learning.This thesis is mainly expanded from the following three aspects: data acquisition,model construction in player ratings as well as application of player ratings.The lack of data is the primary problem for ordinary researchers in football-related research.Especially for researchers who research on machine learning,they often lack the support of professional sports data provider.For most researchers,the internet is nearly the only way to obtain the data,and these incomplete and unreliable data can barely support researcher's research.Therefore,this thesis proposes a method,which uses web crawler technology,to get complete data of a whole season.Besides,in the data processing,this thesis proposes a similarity algorithm based onplayers' names,to match data form different data sets which obtained from web crawlers.Model construction in player ratings is the key work of this thesis.Good model owns accuracy and reliability of player ratings and gives rise to the improvement of the application value of player ratings.In this thesis,we propose a model base on regression algorithm.This model using player statistics of training set to train the model by fitting expert ratings of training set.After this training,we use player statistics of training set to calculate model ratings,and compare model ratings with expert ratings of testing set.The results show that our ratings are fit the expert ratings well and better than the ratings of whosored which are the most popular ratings in top football leagues.We also propose an improved model which combines cluster and regression.Firstly,we cluster players by their characteristics.Secondly,we generate regression model for each cluster.Thirdly,we get the model ratings of testing set based on different model.The results show that combination of cluster and regression can improve the model ratings to be closer to expert ratings.Accurate player ratings model can be used to different scenarios of football field,such as team selection,transfer decision and match result prediction.This thesis mainly presents the application of player ratings on match result prediction.Based on PageRank algorithm,we propose a method which calculate team ratings by player ratings to predict match result.The results show that the method based on player ratings achieves better accuracy than method based on historical results.