Do different recommender types shape opinion convergence differently?

Online stores frequently use multiple recommender algorithms simultaneously. Amazon, for instance, has both "Frequently bought together" and "Customers who viewed this item also viewed" recommendation lists. Each is trained differently and recommends different groups of products. Each creates a different product network — the structure of which products link to which other products via recommendation.

The Maleki Shoja and Tabrizi finding is that the network type matters for opinion convergence. Whether a pair of connected products has converging ratings (similar reviews) or diverging ratings (different reviews) depends on which type of recommender created the link. Frequently-bought-together networks tend to produce one pattern of convergence; co-viewed networks produce another.

This decouples the question of "do recommendations affect ratings" from "which kind of recommendation does what." The mechanism: different recommendation types nudge different population subsets to encounter different items, and those subsets bring different prior expectations. People who buy two items together for a specific use-case develop a different review pattern than people who view both but might buy only one. The recommender shapes both the audience and the comparative frame, which shapes the ratings.

The practical implication for platforms: choosing which recommender to deploy is not just a click-rate decision. It actively shapes the rating ecosystem — what reviews look like, how they correlate, what kind of word-of-mouth propagates. The platform's recommender choice is upstream of the data the platform later analyzes for product insights, creating a feedback loop the platform might not be aware it's creating.

Source: Recommenders General