Browse Theses

For many users of information systems, "information overload" has become a problem: the amount of information they must sift through has reached the point where it is overwhelming. Collaborative Filtering (CF) techniques, which attempt to predict what information will meet a user's needs based on data coming from similar users, are becoming increasingly popular as ways to combat this information overload. While accuracy has been a major focus of CF, in practice, efficiency, privacy, and reliability are also important issues in CF in order to enhance the performance, security, and robustness of CF systems.

Efficiency refers to the computational cost of CF algorithms. We propose two efficient algorithms that are able to learn low-dimensional linear models (either unconstrained or non-negativity constrained) from a very large number of ratings.

Privacy-preserving CF provides protection against divulgence of personal information. Ratings (and even their existence) can reveal information about individuals' personal preferences. We introduce two data reconstruction methods to reveal limitations in existing privacy-preserving schemes and propose a new two-way communication scheme to help users preserve more privacy.

Reliability is the ability to detect and prevent malicious attacks that might make specific items appear more or less popular than they truly are. We describe a series of approaches to detect a diverse and general set of recommendation attacks and demonstrate their effectiveness with data.

The research work introduced in this thesis will benefit both users and implementers of information access systems. Research results will have significant implications for a variety of adaptive information systems that rely on users' input for learning purposes.