eugeneyan/recsys-nlp-graph

Implements Skip-gram embeddings and graph-based sequence generation over product co-purchase networks using PyTorch, with NLP side information (titles, categories) to handle cold-start items. Compares five distinct approaches—matrix factorization with L2 regularization and bias terms, Node2Vec, Gensim Word2Vec, and PyTorch Skip-gram with negative sampling—evaluated on Amazon electronics and books datasets with precision-recall curves under different evaluation protocols (all vs. seen products only).

145 stars. No commits in the last 6 months.