SQ-Offensive-Rebounding

Link to Kaggle Leaderboard (I came in first place!)

Predicting Offensive Rebounds in Basketball: A Location-Based Approach

Basketball is a fast-paced sport, rich in strategies and opportunities for data analysis. One critical moment that can shift the momentum of a game is the rebound after a missed shot. The ability to anticipate and secure an offensive rebound can create additional opportunities to score, making it a valuable skill on the court. The challenge presented here is to leverage the on-court location data of players to predict the probability of a team securing an offensive rebound after a shot miss.

Dataset Overview

The datasets provided for this challenge come in two flavors:

1. Play-by-Play Data (PBP): This dataset captures play-level details. Crucially, it records the team names, and for the training set, whether a rebound is offensive or defensive.

2. Location Data (Locs): This is the centerpiece of the challenge. For each play, it records the on-court coordinates of each player at the moment a shot is taken. This dataset differentiates between the shooter, the offensive players, and the defensive players.

A critical note is that all the shots considered in both the training and test sets are misses, which are then rebounded either offensively or defensively.

Objective

Given this data, the task is to predict the likelihood of an offensive rebound post a missed shot. The prediction will be based on the positions of the players at the time of the shot.

My Approach

In my exploration of this problem, I have undertaken a comprehensive approach to feature engineering and model testing:

1. Distance Features: Calculated distances between key players and positions to gauge the immediate opportunity or challenge in securing a rebound.

2. Angle Features: Considered angles to understand players' relative positions to the basket.

3. Box-Out Responsibility: Paired players based on potential matchups, considering 'boxing out' in rebound scenarios.

4. Complex Feature Engineering Developed features using the results of EDA and basketball domain knowledge

5. Hyperparameter Tuning: Tuning hyperparameters using gridsearch crossvalidation.

6. Building an Ensemble: Combining the prediction power of several diverse ML algorithms to leverage the strengths of each model.

7. Model Exploration: Beyond traditional algorithms, I experimented with Graph Neural Networks (GNNs) and Convolutional Neural Networks (CNNs) to capture the spatial relationships between players.