Opaque User Embeddings Split Into Named Modes

The authors treat dense user embeddings as superpositions of sparse behavioral modes and recover them with dictionary-learning methods. On synthetic data with known ground truth, sparse autoencoders outperformed classical sparse baselines, including for transformer-encoded embeddings. Applied to classical and transformer-based embeddings from two real datasets, the recovered modes could be named and used for zero-shot audience retrieval over representations that were never trained on text. The method is post-hoc and model-agnostic, although interpretability of recovered labels still depends on the evaluation setup.