Clustering and Analyzing 5v5 NHL Shot Location Data

I presented my work on clustering NHL player 5v5 shot heatmaps at the Carnegie Mellon Sports Analytics Conference poster competition.

Abstract

Heatmap-style visualizations are a popular and intuitive way to display shot location data in hockey. However, these visualizations do not allow for large scale player comparison. In this poster, I provide a solution to this issue by clustering NHL players based on their even strength shot locations over three seasons (2017-18 to 2019-20) and analyzing the results.

The clustering consists of two stages. First, I fit a 565 square shot density polygrid to each player’s shot locations through the use of 2D-Gaussian kernel density estimation. Following that, I perform Ward's linkage hierarchical clustering to group players based on their shot polygrids.

This methodology results in 10 clusters that can be broken down into three subgroups: home plate forwards, perimeter forwards and defencemen. Upon forming these clusters, I analyze the results through comparing overall shot heatmaps, the top scorers, and the distribution of expected unblocked shooting percentage, actual unblocked shooting percentage, and player height for each cluster.