Reproduce “Embarrassingly Shallow Autoencoders for Sparse Data” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Harald Steck
EASE^R (Embarrassingly Shallow AutoEncoder) introduces a linear autoencoder for recommender systems that achieves state-of-the-art ranking accuracy on sparse datasets like MovieLens and Netflix Prize, outperforming complex deep learning models while offering significantly faster training times due to its closed-form solution.
Key insights
- The model reveals that the inverse of the data Gram matrix (X^TX) serves as a conceptually correct similarity measure for recommendation.
- EASE^R learns both positive and negative relationships between items, with approximately 60% of learned weights being negative, indicating the importance of capturing dissimilarities.
- The learned weight matrix can be interpreted in terms of conditional independence in graphical models, providing theoretical grounding for the approach.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.