A Fast Gradient Boosting Based Approach for Predicting Frags in Tactic Games

Haitao Xiao, Jinzhong Yang, Yuling Liu, Junrong Liu, Dan Du, Zhigang Lu

Abstract: Predicting the probability of scoring a frag in a tactical video game is a challenging task. It is hard for humans to evaluate the real-time game situation and predict whether a player can score in his/her turn. In this paper, we present a fast gradient boosting based approach to this problem consisting of data analysis, feature engineering, and model construction. Firstly, we analyze the game data and identify the key factors that influence the probability of frag scoring. Then, we extract relevant features from game states metadata and map metadata in the feature engineering stage. Finally, we train and predict the probability of scoring a frag using a gradient boosting based method. Our proposed approach achieves an AUC score of 0.8008 on the whole test set, and only takes 156 seconds for 10-fold cross-validation, demonstrating its effectiveness and efficiency.