PerspectiveSpace: Opinion Modeling with Dimensionality Reduction

Abstract

Words mean different things to different people, and capturing these differences is often a subtle art. These differences are often “a matter of perspective”. Perspective can be taken to be the set of beliefs held by a person as a result of their background, culture, tastes, and experience. But how can we represent perspective computationally?

In this paper, we present PerspectiveSpace, a new technique for modeling spaces of users and their beliefs. PerspectiveSpace represents these spaces as a matrix of users, and data on how people agree or disagree on assertions that they themselves have expressed. It uses Principal Component Analysis (PCA) to reduce the dimensionality of that matrix, discovering the most important axes that best characterize the space. It can then express user perspectives and opinions in terms of these axes. For recommender systems, because it discovers patterns in the beliefs about items, rather than similarity of the items or users themselves, it can perform more nuanced categorization and recommendation. It integrates with our more general common sense reasoning technique, AnalogySpace, which can reason over the content of expressed opinions.

An application of PerspectiveSpace to movie recommendation, 2-wit, is presented. A leave-one-out test shows that PerspectiveSpace captures the consistency of users’ opinions very well. The technique also has applications ranging from discovering subcultures in a larger society, to building community-driven web sites.