Crowdsourced correlation clustering with relative distance comparisons

TL;DR

This paper introduces a new correlation clustering method based on relative distance comparisons, enabling direct clustering from human-provided comparisons without intermediate steps, and demonstrates its effectiveness through experiments.

Contribution

It defines a novel correlation clustering variant for relative distance data and provides approximation and practical algorithms with empirical validation.

Findings

Our method outperforms existing complex methods on synthetic data.

It efficiently produces intuitive clusterings from real human comparison data.

The approach is suitable for crowdsourced clustering tasks.

Abstract

Crowdsourced, or human computation based clustering algorithms usually rely on relative distance comparisons, as these are easier to elicit from human workers than absolute distance information. A relative distance comparison is a statement of the form "item A is closer to item B than to item C". However, many existing clustering algorithms that use relative distances are rather complex. They are often based on a two-step approach, where the relative distances are first used to learn either a distance matrix, or an embedding of the items, and then some standard clustering method is applied in a second step. In this paper we argue that it should be possible to compute a clustering directly from relative distance comparisons. Our ideas are built upon existing work on correlation clustering, a well-known non-parametric approach to clustering. The technical contribution of this work is twofold. We first define a novel variant of correlation clustering that is based on relative distance comparisons, and hence suitable for human computation. We go on to show that our new problem is closely related to basic correlation clustering, and use this property to design an approximation algorithm for our problem. As a second contribution, we propose a more practical algorithm, which we empirically compare against existing methods from literature. Experiments with synthetic data suggest that our approach can outperform more complex methods. Also, our method efficiently finds good and intuitive clusterings from real relative distance comparison data.