User Embeddings are Superpositions of Interpretable Behavioral Modes
The method lets researchers query user embeddings in natural language without training those embeddings on text.
Sparse autoencoders extract behavioral modes that can be automatically named and queried. On synthetic data with known ground truth, they recovered modes better than classical sparse baselines, including from transformer-encoded embeddings. The authors also applied the method to two real-world datasets and demonstrated zero-shot audience retrieval and characterization using natural-language labels. The framework works after training and requires no changes to the upstream embedding system.
ArXiv · AI/CL/LG's note
Sparse autoencoders extract behavioral modes that can be automatically named and queried. On synthetic data with known ground truth, they recovered modes better than classical sparse baselines, including from transformer-encoded embeddings. The authors also applied the method to two real-world datasets and demonstrated zero-shot audience retrieval and characterization using natural-language labels. The framework works after training and requires no changes to the upstream embedding system.
ArXiv · AI/CL/LG's note
score 5