Weight Reviews

In content aggregation systems, the goal is typically to generate numerical scores to estimate aggregate preferences from human feedback. However, ratings from different users are likely to have different levels of value for this task. In the context of location based rating:

A person that visits every pizza place in New York searching for the best slice likely has better discernment on the rating of an NYC pizza joint than the tourist who visits the Sabaro in Times Square.

Additionally, The user that leaves all 5-star or all 1-star reviews probably is a less discerning reviewer than the one who leaves a variety of ratings.

Finally, the rating from a user that has only rated a few things is more likely to be an inauthentic review.

Implicit in the structured data of a user's ratings over time is a pattern of data in which the usefulness of their recommendations for others can be reasonably well estimated. Averaging these ratings by a series of weights based on past reviews could yield better aggregate ratings, and make the system less susceptible to manipulation by motivated actors.

Though the intervention sketched out here is focused on the goal of accuracy in rating for an average, and thus would have a significant homogenizing effect, that goal is maleable, and the point of differential value for reviews still applies. Ratings could instead be weighted by their similarity to the characteristics a user has expressed an implicit or explicit interest in. A user who consistently leaves comments about the quality of the mobility accommodation in a space might be better served by rating aggregations that weight ratings that mention mobility features highly.