There’s Something About You: Epistemic Recommendation for Latent Interest Discovery

Abstract

How well does a recommender system know you? These systems are typically trained on the silhouette of user activity to predict immediate engagement, yet this narrow focus may paradoxically expose how incomplete the system’s knowledge of the user really is. Rather than recommending from established user preferences, this paper considers recommendations as a means to enrich our knowledge of the user. We introduce epistemic gain (EG), a signal that measures how much a recommender would learn about a user from each candidate item’s outcome. For each candidate, EGRec (Epistemic Gain-based Recommendations) constructs two hypothetical scenarios (the user engages, or does not) and measures the resulting catalog-wide relevance distribution divergence. EG is a versatile exploration signal applicable across paradigms, whether as a blended scoring component, an engagement-weighted self-regulating score, a context feature for bandit methods, or a component of an RL reward objective. To account for practical scalability considerations, a lightweight prediction head trained on frozen model embeddings approximates EG in a single forward pass. A dual regret framework, analyzing both reward and coverage regret simultaneously, allows us to prove that under structural conditions on the EG signal, EGRec achieves coverage regret bounded independently of the time horizon, while greedy relevance incurs linear regret. We evaluate EGRec on KuaiRec and MovieLens-1M with two sequential architectures. Combining EG with bandit exploration consistently leads on both interest coverage and accuracy across most evaluated configurations. Bandits supply item-reward-level uncertainty calibration while EG provides latent-interest-level learning direction. A comparison against a novelty-baseline method shows that EG carries information beyond category novelty alone. Moreover, EG signal quality depends on base model expressiveness, a dependency that serves as a practical diagnostic for when the approach is most effective. The most valuable recommendation may end up not being the one a user is most likely to click, but rather, the one from which the system stands to learn the most about the user.