Original article was published by Tim de Boer on Deep Learning on Medium
Idea For Predicting Soccer Goals Using GPS Data And Deep Learning
Recently, I started learning about Deep Learning via the fast.ai Deep Learning For Coders course taught by Jeremy Howard.
Numerous of applications have been developed with the help of deep learning, with one of the most prominent field the image classification, detection and recognition.
A deep learning model is able to detect every possible pattern that may be contained within an image.
A simple example could be classifying three types of bear: grizzly, black and teddy bears. Nowadays, and with the help of higher level abstract libraries such as fast.ai or Keras, a deep learning model could be developed in a short amount of code and can learn to accurately classifiy the three types of bear in under 1 minute of training.
Not all data in the world comes in images. Deep learning can of course be deployed on more types of data apart for images, such as tabular or text data. Unfortunately, some data can be difficult to visualize and computionally expensive to run a deep learning model on, such as GPS data. However, these datapoints could be tranformed into images over a certain time interval. Training a deep learning model with these images turns out to be really successful.
An example of this method was mentioned in the course: fast.ai student Oleg Izmerly successfully deployed a Deep Learning Model for Online Fraud Detection at the company Splunk by transforming mouse movements and clicks into an image. Direction, speed and acceleration are represented in the different colors of the lines and mouse click action are represented in the larger color circles.
Exactly this method could be used for analysing soccer games!
Mouse clicks can be replaced with passes or shots on goal, with the color of the circle representing the speed of the pass or shot. The lines would represent the direction, speed and acceleration of the players. Using a deep learning model, it may be possible to detect patterns which predict a successful attack, a goal, an injury, and probably even more.
I hope you had fun reading my first blog post. Send me a message in case you have more ideas about this model, or if you are interested in deploying this model at your soccer club 🙂