Modelling Uncertainty and Vagueness within Recommender Systems via Nonparametric Predictive Inference

Abstract

The way in which we learn is the subject of considerable research within multiple disciplines. There is also a vast amount of on-line material available to us, causing decision-making to become increasingly difficult. Learning preferences for decision-making processes has been an area of substantial research in recent years given the introduction of Recommender Systems (RSs). RSs help in decision-making processes by recommending items of interest and filtering out undesired items, they need to learn preferences by extracting information about both the user and the item. This thesis presents a novel approach of incorporating vagueness and uncertainty into recommendations via Nonparametric Predictive Inference (NPI). This approach is termed the Uncertainty Interval (UI); it is a modified version of Nonparametric Predictive Utility Intervals. There are four UI approaches presented: UIUntrans, UIAbs, UISq and UIRt. Each algorithm is evaluated and compared with a similar technique, Robust Bayesian Correlation Learning. The UIAbs algorithm has superior performance to the other $UI$ approaches and is applied to real world data. The width of the interval reflects the amount of information available to the RS, with a wider interval indicating little or no information. The interval narrows as more information is incorporated into the UI algorithm.