Can attention mechanisms reveal which user taste explains each recommendation?

Single-vector user representations treat tastes as monolithic. A user who likes both horror movies and comedies gets one latent vector encoding the union, and at recommendation time, the dominant taste tends to overtake the list. The conventional fix is to bolt a diversity-enhancing reranker on top — but that admits the underlying model can't represent the user's tastes correctly, only mask the symptom.

AMP-CF restructures the representation. Each user has multiple latent personas, each capturing a different taste cluster. When scoring a candidate item, an attention mechanism weights the personas by their relevance to that item — a user's "horror persona" lights up for horror candidates and stays quiet for comedies. The user representation becomes candidate-conditional in a way single-vector models can't be: same user, different effective vector depending on what's being scored.

This buys two distinct goods at once. Recommendations become diverse without a separate diversity step because the inactive personas surface their preferences when their kind of item shows up. Recommendations become explainable because each item can be attributed to the persona that gave it the highest weight — "we recommended this because of your horror taste, not your comedy taste." The Taste Distribution Distance metric the paper introduces measures whether the recommendation list proportionally matches the user's full range of interests, which diversity metrics don't capture.

Source: Recommenders Architectures