Empirical Analysis of Predictive Algorithms for Collaborative Filtering

The paper studies algorithms for collaborative filtering, where a database of user preferences is used to predict additional items a new user might like. It describes and implements several prediction methods, including techniques based on correlation coefficients, vector-based similarity calculations, and statistical Bayesian methods such as Bayesian clustering and Bayesian networks with a decision tree at each node. To evaluate them, the authors use two classes of metrics: one measuring average absolute deviation across individual predictions, and another estimating the utility of a ranked list of suggestions based on the probability that a user sees a recommendation.

Experiments spanned three application areas, four experimental protocols, and both evaluation metrics. Bayesian networks with decision trees and correlation methods outperformed Bayesian clustering and vector similarity across a wide range of conditions, while the choice between correlation and Bayesian networks depended on the dataset, the application (ranked versus one-by-one presentation), and the availability of votes. Other practical considerations included database size, prediction speed, and learning time, making the study a systematic guide to selecting collaborative filtering algorithms for different settings.