New AI Turns Barely-There Reviews Into Smarter Recommendations

Online stores, streaming services, and apps all use recommender systems to guess what you'll like. Usually, the more you rate and review, the better the suggestions. But most people don't write detailed reviews every time. That leaves big gaps in the data, and recommendations often get worse for exactly the people who need them most: new or casual users.

The new system, called MOSAIC, borrows an idea from academic publishing. In academic peer review, an editor reads reviews from several experts and writes a meta-review—a single summary that captures the main points. MOSAIC does something similar. For each user, it looks at reviews written by similar people, pulls out the key opinions about specific things like price, quality, or convenience, and combines them into a personalized meta-review. That fills in missing details without pretending the user wrote them.

It also solves a second problem: incomplete reviews. A review might say great product without saying why. MOSAIC uses the collective language of neighbors to infer what attributes that person probably cares about. It then predicts ratings better and can even show attribute-level explanations, like this user values battery life and durability, rather than just saying you'll like it.

In tests on four real-world datasets, MOSAIC beat current top models in both rating accuracy and explanation quality. It especially helped users with very little history. The honest catch: the system depends on having enough similar users with useful reviews. If a user is truly one of a kind, or if their neighbors' reviews are also empty, MOSAIC can't perform magic. Still, it's a practical step toward recommendations that work even when we don't leave long reviews.