• Gunn, S. R.
• Harris, C. J.
• Reed, Philippa A. S.
• Lee, K. K.

Abstract

Many classification techniques emphasize obtaining a good classification rate. In our view, a more important issue is to be able to interpret the underlying model. The aim of this work is to build data driven classifiers that provide enhanced understanding of a system through visualisation of I/O relationships, in addition to good predictive performance for a set of imbalanced data. The problem of data imbalance is addressed by incorporating a different misclassification cost for each class and an appropriate performance criteria. The Support vector Parismonious ANalysis Of VAriance (SUPANOVA) technique has been used successfully for regression problems in generating interpretable models. In this paper, we modify<br/>SUPANOVA so that it can be applied to the domain of classification problems enabling a predictive model with a high degree of interpretability to be recovered. Here, the problem of classifying and predicting fatigue crack initiation sites, through microstructure quantification in Austempered Ductile Iron (ADI), is considered. SUPANOVA selects a sparse set of components from the model for easy visualisation. Results from the modified SUPANOVA technique provide good performance with 5 components selected out of the possible 512 as significant components. The components selected are consistent with prior knowledge of metallurgists working on the material. With this modelling knowledge, the key production and microstructure features can be identified to optimise automotive materials performance.

Topics

• impedance spectroscopy
• microstructure
• crack
• fatigue
• iron