A new study emerging from the MM4SPA project has just been published in Applied Intelligence!
In this study, the authors compared several common data representation formats that have recently been used to process football tracking data with deep learning architectures. Based on their benefits and drawbacks, a new graph-based representation and a corresponding hybrid network were proposed.
In a comprehensive state-of-the-art comparison on the same generic classification task, it was found that this new representation achieves top results while being 100 times leaner than networks of similar capabilities. The resulting TGNet is a light-weight, hybrid graph-based network that may be used for all deep learning applications involving tracking data and is now being deployed in several follow-up studies.