ExUP: Inferring multiple-dimension-based ratings from a single-dimension rating system

ExUP: Inferring multiple-dimension-based ratings from a single-dimension rating system Conference

Castellanos, A, Castillo, A, VanderMeer, D. (2013). ExUP: Inferring multiple-dimension-based ratings from a single-dimension rating system .

Collaborative filtering (CF) is a popular method for recommender systems. It clusters users based on similarities in their rating histories. Most rating schemes ask for a single-dimension rating score that reflects the user's overall experience with a product or service. The dimensionality of a user's rating, i.e., the pros and cons that led to the overall score, is not captured in such schemes, which can lead to misattributed preferences and lower-quality recommendations. One solution might be to simply add rating dimensions to capture additional preferences. However, this can lead to survey fatigue, with reduced rating quality and user participation. We introduce a method, ExUP, which takes a single-dimension rating and infers the multi-dimensionality that comprises a user's overall score. We implemented and tested ExUP with real data from a restaurant rating site. Our experimental results show that, compared to CF, ExUP provided (on average): 8.2% greater precision (max: 11.6%), 5.9% greater recall (max: 6.9%), and 6.8% greater F-measure (max: 8%). Further, ExUP's training complexity is many orders of magnitude less than that of CF's training, and only 1.6 times the real-time recommendation complexity of CF, which suggests potential for real-time implementation for both training and prediction/recommendation. © Thirty Fourth International Conference on Information Systems, Milan 2013.