An Augmentation for the Trueskill Video Game rater with attention to offense/defense and player contribution variables

Edward
Tensorflow

To solve the player/team evaluation problem we use a kaggle dataset on European Soccer Games available here. This data set contains seven tables: countries, leagues, matches, player, player attributes, teams, and team attributes. We are primarily concerned with match data from different teams. The player and team attributes are from FIFA (the video game) and are unlikely to be used

We attempt to create a solution that is efficiently scalable to a large number of teams and players. Individual player skill is modeled as a Gaussian, with the mean as their skill rating and variance as therandomness in their performance. The team rating is modeled dependently on the player skills and theirindividual impact to the team(all imapacts summing to 1), ignoring team attributes given by the dataset. We will use this model to augment the existing TrueSkill model

One key change we will be making with our new model is that we will be implementing the model using Edward and conducting inference using Black Box Variational Inference. Guo (2011) uses Variational EM with exact updates

Our evaluation criteria will follow that of Guo (2011) which is an estimate of classification accuracy using the area under the curve measure of the ROC curve. In hyperparameter tuning, we will evaluate theimpact of our player-based scoring model by comparing the AUC of our model and the the original

Shengbo Guo: Bayesian Recommendar Systems
Ralf Herbrich, Tom Minka, Thore Graepel: Trueskill
Data.