Beating the Baseline Recommender with Graph & NLP in Pytorch
Eugene Yan
A practitioner implemented product recommendation systems using graph-based and NLP approaches, progressing from matrix factorization through graph random walks and word2vec embeddings to a PyTorch implementation. The PyTorch word2vec model achieved an AUC-ROC of 0.9855 on seen products, outperforming the gensim baseline of 0.9082 and an Alibaba research paper's 0.9327. The implementation addressed memory constraints by using sparse matrices for the transition matrix and negative sampling to reduce computational complexity from millions of weights to thousands.
Why it matters
Beating the baseline using Graph & NLP techniques on PyTorch, AUC improvement of ~21% (Part 2 of 2).
