ExUP recommendations: Inferring user's product metadata preferences from single-criterion rating systems

Highlights：

• Most online rating schemes for products allow users to give a single rating, typically on a five-point Likert scale, which hides how a product’s attributes may be influencing an individual’s experience with the product. In this work, we propose a method to combine these ratings with multi-valued, multi-dimensional product data to infer granular user preferences for the purpose of generating product recommendations.

• We implemented a prototype system to demonstrate the effectiveness of our method, and tested it on a large volume of user ratings for movies, and the associated movie attributes.

• Our experimental results showed that, compared with the two best-performing existing state of the art methods, our method provided review score predictions with up to: 47.7% greater precision, 6.9% greater recall, and 20.5% greater F-measure than existing methods.

摘要

Recommendation systems make use of complex algorithms and methods to provide recommendations to consumers. Typically, online rating schemes use a single rating metric that captures the overall user experience with a product. Nevertheless, this might hinder the intricacies of how a product's attributes influence an individual's preferences. While it is possible to use sentiment and semantic analysis to interpret free text in user reviews, if available, to gain insight into a user's reasons for a product rating, these methods are expensive to implement and error prone, and rely on significant data input from the user. To overcome these challenges, we propose a method for inferring user preferences and generating recommendations without relying on the availability or quality of text reviews. Specifically, our method is designed to use existing product metadata and user rating patterns to shed light on how the attributes of a product correspond to individual preferences. Our method uses only the user's history of ratings and the corresponding product attributes to generate predicted ratings for products a user has not yet experienced. This work extends existing work in this area by focusing on multi-valued attributes, and considering the distinct impact of each attribute value in a user's preferences. In terms of computational complexity, our method runs in linear time, making it feasible for real-time implementations. Our experimental results showed that, compared with the two best-performing existing state of the art methods, our method provided review score predictions with up to: 47.7% greater precision, 6.9% greater recall, and 20.5% greater F-measure than existing methods.