Avoiding Bias in Text Clustering Using Constrained K-means and May-Not-Links

Abstract

Abstract. In this paper we present a new clustering algorithm which extends the traditional batch k-means enabling the introduction of domain knowledge in the form of Must, Cannot, May and May-Not rules between the data points. Besides, we have applied the presented method to the task of avoiding bias in clustering. Evaluation carried out in standard collections showed considerable improvements in effectiveness against previous constrained and non-constrained algorithms for the given task.

Improving Alternative Text Clustering Quality in the Avoiding Bias Task with Spectral and Flat Partition Algorithms

“…Firstly, we survey normalised cut, a very effective spectral clustering algorithm introduced by Shi and Malik in [5], and its constrained counterpart, constrained normalised cut, introduced by Ji et al in [4]. Afterwards, we outline soft constrained k-means, a constrained clustering algorithm based on k-means introduced by Ares et al in [3].…”

“…In order to address these limitations, Ares et al introduced in [3] two kinds of non absolute constraints: May-Links and May-Not-Links, which indicate that two documents are, respectively, likely or not likely to be in the same cluster. The implementation of these constraints alters again the assignment process of the documents.…”

“…If we are trying to avoid the tendency (bias) of the clustering algorithm to fall in a certain grouping of the data that is being clustered the task is called Avoiding Bias. This problem has been tackled by several authors in the last years, which have proposed a wide range of approaches, ranging from distance learning [2] to using constraints [3]. However, it should be underlined that avoiding bias is still a clustering process, where the main focus is providing the user with a meaningful grouping of the data.…”

“…Concretely, we test two different approaches which use a strategy similar to the one in [3] (using negative constraints to steer the clustering process away from the known clustering), making use of spectral clustering techniques to try to attain that high quality. The first one is introducing negative constraints in the constrained normalised clustering approach proposed by Ji et al in [4].…”

“…The first one is introducing negative constraints in the constrained normalised clustering approach proposed by Ji et al in [4]. The second one is introducing the soft constrained k-means algorithm proposed by Ares et al in [3], which has been shown to have good results, in the second phase of a normalised cut clustering algorithm [5]. The experiments carried out with these approaches showed that, while the first approach does not yield good results, the combined one (normalised cut plus soft constrained k Means) outperforms soft constrained k-means in terms of quality of the results while keeping a good avoidance of the known clustering.…”