The Rating System

Introduction

YaleStation guides employ a scientific technique known as collaborative filtering. This process, which may contain many underlying mathematical algorithms, is used to make personal recommendations, intelligently discover patterns and similarities, and aggregate individual, atomic opinions.

Community members are able to rate establishments listed in the guides on a numeric scale from one (worst) to five (best). Some guides may require that a rating be spread across multiple categories to force a certain thought process and to enable separation of complex individual ratings. For instance, the dining guide requests individual ratings for food, décor, and service. Collected data indicate that the ratings for many restaurants, for instance, differ dramatically between food and décor.

Aggregation of Atomic Ratings

Individual ratings by a single person for a single establishment, known as atomic ratings, must somehow be summarized to create an aggregate rating for a single establishment based upon the ratings of many individuals.

Several weights comprise the algorithm developed by YaleStation to aggregate atomic ratings. The algorithm extends a preference to users who have rated many establishments and have not rated with extremes (e.g., all 1's or all 5's). Preference is also extended to atomic ratings cast more recently than other ratings in the set.

Atomic ratings are normalized based upon other ratings in a given set. For instance, the aggregate rating of an individual establishment may differ from the aggregate rating of that establishment for use in comparison with other establishments.

Fluctuation of Ratings

The rating system uses machine-learning technologies to grow the intelligence of its data mining and aggregation operations. As new ratings are cast, the algorithm becomes further refined and representative of the rating community. Changes to the algorithm may produce changes in the aggregate ratings independent of any changes to the atomic ratings.

Although atomic ratings are never directly altered, the percentage with which they are counted effectively alters the numerical rating. Consequently, as the algorithms improve, original ratings are preserved for reevaluation.

By using collaborative filtering processes on the atomic and aggregated ratings, recommendations are able to be made to the user based upon the user's own ratings set against the ratings of the larger community.

Resultant Products

Lists of similar establishments are generated by examining the ratings of users for a particular establishment and examining what entities they rated. A weak rating in either set weakens the connection. Conversely, a strong rating in both sets improves the connection. The product results in a list of "users who liked this restaurant also liked these restaurants."

Personal recommendations add one degree of complexity to the previous method. In contrast to the two datasets required by the similar establishments system, this method requires at least three datasets. This technique produces a list of "restaurants that I liked had users who liked that restaurant and also liked this restaurant." Like before, the strength of the overall connection depends upon the individual strengths of the underlying connections.