The project explores using the LightGCN model for movie recommendation. The repository was forked from here.

Specifically, we are using the Movie Lens Small Dataset. This dataset is retrived from https://grouplens.org/datasets/movielens/ and is preprocessed with the file ml-latest-small.py and ml-latest-small-new.py.

We modified the original LightGCN to incorporate the movies' rating by the users, as well as the movies' genre. We call this new model "Modified LightGCN". We also implemented a few baselines (SVD, SVDpp and KNN) taking into account the movie ratings. The baselines selected for experiment include KNN-based models with different similarity measures (Cosine Similarity, Pearson Similarity, Pearson-Baseline, and Mean Squared Difference), SVD and SVDpp. We perform experiments comparing the original LightGCN, Modified LightGCN, filtered LightGCN (where we only consider giving movie recommendations to active users) and baselines on the prediction task of Movie Lens Dataset.

For original LightGCN, use: python main.py --decay=1e-4 --lr=0.001 --layer=3 --seed=2020 --dataset="ml-latest-small" --topks="[20]" --recdim=64 --testbatch 61 --epochs 500

For Modified LightGCN training with ratings, add: --multiplied

For Modified LightGCN training with movie categories, add: --genre

Need to delete the s_pre_sdj_mat.npz after running to reinitialize the adjacency matrix.

The process for executing the experiments have been encapsulated in baselines.py and baseline_experiments.py. To run the experiments, simply run baseline_experiments.py with the following configuration: --dataset="ml-latest-small" --topks="[20]" --recdim=64 --testbatch 12

The baseline experiment results will be written into the directory data/baseline_results_ML.

For details on the baseline experiments (e.g. trainable parameters), please refer to the comments in baselines.py and baseline_experiments.py.