Steam is a video game distribution platform. We employ neighborhood, matrix factorization, and mixed collaborative filtering (CF) methods to predict the number of hours Steam users will play games. We also adapt a regression boosting framework for matrix factorization CF algorithms and apply it to the prediction task. We find that neighborhood methods outperform matrix factorization methods, and a mixed approach outperforms both. Additionally, we find the boosting framework did not meaningful improve performance. To improve predictions, future research should incorporate user friendship networks.
- Brandon Lin
- Chris Painter
- Barry Plunkett
- Stephanie Shi
neighborhood- memory-based methodsfactorization- latent factor modelsboost- boostingensemble- mixed methods
pip install -r requirements.txt
Full original dataset can be found here and is over 200GB. Processed data is too large to include and can be obtained by contacting the owner.