Reproduce “User Embeddings are Superpositions of Interpretable Behavioral Modes” 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
Jonathan Kozaczuk, Zheng Dong, Noyan Evirgen, Keivan Majidi, Eddie K. H. Ng, Chris Porter, Qi Zhao, Jia Geng and 2 more
The authors use a sparse autoencoder (SAE) to decompose existing user embeddings into behavioral features, then use language-model labels to support natural-language audience retrieval. In a controlled transformer-embedding test, the SAE recovered 73 of 75 known modes at decoder–centroid cosine similarity of at least 0.7, with 0.58 alignment versus 0.53 for FastICA; MovieLens ranking was not uniformly better than K-Means.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.