Inductive Matrix Completion Using Graph Autoencoder for Sparse Rating Matrices

PyTorch implementation of Inductive Matrix Completion Using Graph Autoencoder (Shen et al., CIKM 2021) for matrix completion on a very sparse MovieLens subset (610 users × 4,980 movies, ~1% observed), with movie genres and release year as side information, compared with a biased ALS baseline.

Held-out test RMSE: tuned GNN 0.8682, biased ALS 0.8720, training-mean predictor 1.0369. See report.pdf.

Layout

• cf/: data, rating graphs, features, IMC-GAE model, trainer (full-batch Adam, early stopping, cross-validation), ALS baseline
• experiments/: Optuna search and final fit (notebook_imc_gae.ipynb), baselines, plots

Credits

R. Breaud, S. H. Nguyen, A. Aidoudi - Supervised by Alexandre Vérine.

Ratings from MovieLens (F. M. Harper and J. A. Konstan, The MovieLens Datasets: History and Context, ACM TiiS 2015); not endorsed by GroupLens.