Using a data-driven approach and machine learning to coach at the collegiate level

Photo by Dave Adamson on Unsplash

The University of Illinois Urbana-Champaign (UIUC) believes that technology is a powerful tool for driving results and innovation on campus. Their chief information officer, Mark Henderson, developed a task force—called the Data and Technology Innovation Lab—to identify department challenges and task individuals to build innovative solutions using technology. One area where UIUC identified an opportunity was sports analytics using machine learning (ML).

Building game planning sheets and uncovering data insights

Inspired by what they were seeing in professional sports, Greg Gulick, interim CIO, and the Division of Intercollegiate Athletics started to explore how they could use data and shift their approach to coaching football. To help with this, associate athletic director Nick Rogers and head football coach Lovie Smith hired Kinglsey Osei-Asibey as director of analytics and football technology. After joining the Illinois football program, Kingsley began to instruct the coaching staff each week on how they can use data to inform their preparation for the weekly game.

According to Kingsley, “All football games can be viewed as a sequence of plays, where each play is an action taken by players based on personnel, field positon, down and distance, and other factors. To optimize the chance of winning a game, coaches must optimize the winning chance of each play. I allow the data to tell me everything that the opponent is doing.”

However, uncovering data insights is complex given all the factors that go into calling a play and the amount of data available for coaches to analyze. Typically, coaches generate game planning sheets manually through heuristics (decision-making strategies)—zone, down, and distance—only using a small number of features from available datasets to analyze the best play types. Coaching staffs have to create game planning sheets each week with new data from each opponent, which is time consuming and difficult to scale.

To address the challenge, Kingsley, UIUC’s Data and Technology Innovation Lab, and the Amazon Machine Learning (ML) Solutions Lab came together to build a ML solution. Over the past several months, they worked together to build a ML model that generates weekly game planning sheets for the coaches based on an ever-growing dataset.

With this ML model in place, the coaches generate updated weekly game planning sheets to prepare for their opponent—providing improved insights on their performance strategies and saving them time. In parallel, Amazon Web Services (AWS) is conducting ML knowledge transfer sessions with the UIUC Data and Technology Innovation Lab so that they can learn how to re-train the model and drive improved accuracy over time.

An ML approach to game preparation

The team arranged the data into a supervised ML problem where the labels were the success or failure of a play. To prepare the dataset, they encoded the categorical columns using one-hot encoding, and the dataset was balanced using upsampling techniques. The team built an XGBoost model with Amazon SageMaker to predict the result of a play with 65.2 percent weighted accuracy. During training, Amazon SageMaker HPO (Hyper-parameter Optimization) was used to improve the model performance.

Based on the feature importance scores from the XGBoost model and correlation analysis, they chose the features for the new game planning sheets. These included additional opponent-related features as suggested by the model and presented to the UIUC football coaches for feedback. During the football season, coaches will use these new templates to generate the most frequently used play types in various scenarios each week by running a simple AWS Lambda call on the most recent dataset. The XGBoost model will run every three to four weeks to examine the feature importance score and see if there are new important features that should be included in the game planning sheet templates.

Kingsley and the Illinois coaching staff recognize how hard it can be to change coaching behaviors, but they are confident this is critical step forward for the football program. In the words of Kingsley, they are demonstrating to the coaching staff that, “there is more to preparing for a game than watching film. Your data can tell a story for you.”

Hear more about this story in the webinar, “Cloud Across Campus: How one university is reaping big rewards.” And learn more about how AWS machine learning is ushering in the next wave of technical sports innovation, and the cloud for higher education.