On the quality of ART1 text clustering

Abstract

There is a large and continually growing quantity of electronic text available, which contain essential human and organization knowledge. An important research endeavor is to study and develop better ways to access this knowledge. Text clustering is a popular approach to automatically organize textual document collections by topics to help users find the information they need. Adaptive Resonance Theory (ART) neural networks possess several interesting properties that make them appealing in the area of text clustering. Although ART has been used in several research works as a text clustering tool, the level of quality of the resulting document clusters has not been clearly established yet. In this paper, we present experimental results with binary ART that address this issue by determining how close clustering quality is to an upper bound on clustering quality.

Introduction

We consider the application of clustering to the selforganization of a textual document collection. Clustering is the operation by which similar objects are grouped together in an unsupervised manner (Jain et al., 1999, Kaufman and Rousseeuw, 1990). Hence, when clustering textual documents, one is hoping to form sets of documents with similar content. Instead of exploring the whole collection of documents, a user can then browse the resulting clusters to identify and retrieve relevant documents. As such, clustering provides a summarized view of the information space by grouping documents by topics. Clustering is often the only viable solution to organize large text collections into topics. The advantage of clustering is realized when a training set and classes definitions are unavailable, or when creating them is either cost prohibitive due to the collection shear size or unrealistic due to the rapidly changing nature of the collection.

We specifically study text clustering with Adaptive Resonance Theory (ART) (Carpenter and Grossberg, 1995, Grossberg, 1976) neural networks. ART neural networks are known for their ability to perform on-line and incremental clustering of dynamic datasets. Contrary to most other types of artificial neural networks such as the popular Back-propagation Multi-Layer Perceptron (MLP) (Rumelhart, Hinton, & Williams, 1986), ART is unsupervised and allows for plastic yet stable learning. ART detects similarities among data objects, typically data points in an N-dimensional metric space. When novelty is detected, ART adaptively and autonomously creates a new category. Another advantageous and distinguishing feature of ART is its ability to discover patterns at various levels of generality. This is achieved by setting the value of a parameter known as vigilance and denoted by ρ, ρ∈(0,1]. ART stability and plasticity properties as well as its ability to process dynamic data efficiently make it an attractive candidate for clustering large, rapidly changing text collections in real-life environments. Although ART has been investigated previously as a means of clustering text data, due to numerous variations in ART implementations, experimental data sets and quality evaluation methodologies, it is not clear whether ART performs well in this type of application. Since ART seems to be a logical and appealing solution to the rapidly growing amount of textual electronic information processed by organizations, it would be important to eliminate any confusion surrounding the quality of the text clusters it produces. In this paper, we present experimental results with a binary ART neural network (ART1) that address this issue by determining how close clustering quality achieved with ART is to an expected upper bound on clustering quality. We will consider other versions of ART in future work.