A connectionist approach to structural similarity determination as a basis of clustering, classification and feature detection

Abstract

Many algorithms in machine learning, knowledge discovery, pattern recognition and classification are based on the estimation of the similarity or the distance between the analysed objects. Objects with higher structural complexity often cannot be described by feature vectors without losing important structural information. These objects can adequately be represented in the language of logic or by labeled graphs. The similarity of such descriptions is difficult to define and to compute. In this paper, a cormectionist approach for the determination of the similarity of arbitrary labeled graphs is introduced. Using an example from organic chemistry, the application of the approach within one distance based and one generMisation based classfication algorithm is demonstrated. The generalisation based algorithm forms clusters or subclasses of similar examples of the same class and extracts the parts of the objects which determine the class of the object. The algorithms perform very satisfactorily in comparison with recent logical and feature vector approaches. Moreover, being able to handle structural data directly, the algorithms need only a subset of the given features of the objects for classification.

A Graph Matching Algorithm and Its Application to Conceptual System Translation

“…Second, one can approach the approximate isomorphism problem as the problem of minimizing a certain error function on the space of correspondence matrices subject to structural constraints, and solve the problem using the full arsenal of optimization methods (Gold & Rangarajan 1996;Rangarajan & Mjolsness 1994;Pelillo 1998). Third, a number of neural-net and related iterative approaches have been presented (Schadler & Wysotzki 1997;Schädler & Wysotzki 1999;Melnik, Garcia-Molina, & Rahm 2002). This paper continues work on the simple approach of (Goldstone & Rogosky 2002) that was originally inspired by constraint propagation neural networks for consistent visual interpretation or analogical reasoning.…”

Untitled

“…The objective function here was another approach, Schadler et. al [64] report using a neural-network optimization procedure for 2D chemical graphs.…”