홈 › Papers

The Power of Pixels: Exploring the Potential of CNNs for Expected Goals (xG) in Football

Expected Goals (xG) is a popular metric in football that estimates the likelihood of a shot resulting in a goal. Traditional xG models are useful and are continuously improved incorporating more attributes and data. Leveraging the potential shown by deep learning models in image recognition tasks, we present xG-CNN, a convolutional neural network that computes xG based on the spatial distribution of the opponents in front of the shooter. We describe the input image representation and CNN architecture employed and explain how we trained/validated it using a publicly available dataset. Our goal is to determine whether a CNN-based xG model outperforms previous models. Despite the initial expectations, the model with an AUC ROC of 0.801 did not outperform the publicly announced state-of-the-art models. This could be due to the sample size or not incorporating additional context, such as body type, in the input representations or through separate CNN models. The approach presented here is also relevant to similar metrics and to promote good scientific practices, we have provided the code and data used to generate the results.

Code (1)

Similar Papers 제목 키워드 기반