Effective feature construction by maximum common subgraph sampling

• Published in: Machine learning Vol. 83; no. 2; pp. 137 - 161
• Main Authors: Schietgat, Leander, Costa, Fabrizio, Ramon, Jan, De Raedt, Luc
• Format: Journal Article
• Language: English
• Published: Boston Springer US 01-05-2011; Springer Nature B.V
• Subjects: Algorithms; Artificial Intelligence; Computation; Computer Science; Construction; Control; Data mining; Decision trees; Graphs; Mathematical models; Mechatronics; Mining; Natural Language Processing (NLP); Robotics; Sampling; Searching; Simulation and Modeling; Structure-activity learning; Subgraph mining; Chemoinformatics; Feature generation
• ISSN: 0885-6125; 1573-0565
• DOI: 10.1007/s10994-010-5193-8

The standard approach to feature construction and predictive learning in molecular datasets is to employ computationally expensive graph mining techniques and to bias the feature search exploration using frequency or correlation measures. These features are then typically employed in predictive models that can be constructed using, for example, SVMs or decision trees. We take a different approach: rather than mining for all optimal local patterns, we extract features from the set of pairwise maximum common subgraphs. The maximum common subgraphs are computed under the block-and-bridge-preserving subgraph isomorphism from the outerplanar examples in polynomial time. We empirically observe a significant increase in predictive performance when using maximum common subgraph features instead of correlated local patterns on 60 benchmark datasets from NCI. Moreover, we show that when we randomly sample the pairs of graphs from which to extract the maximum common subgraphs, we obtain a smaller set of features that still allows the same predictive performance as methods that exhaustively enumerate all possible patterns. The sampling strategy turns out to be a very good compromise between a slight decrease in predictive performance (although still remaining comparable with state-of-the-art methods) and a significant runtime reduction (two orders of magnitude on a popular medium size chemoinformatics dataset). This suggests that maximum common subgraphs are interesting and meaningful features.