Predicting Trait Impressions of Faces using Classifier Ensembles
In the experiments presented in this chapter, single classifiers and classifier ensembles are trained to detect the social meanings people perceive in facial morphology. Our first concern in designing this study was developing a sound ground truth for this problem domain. Our goal was to collect a set of faces that exhibit strong human consensus in a comprehensive set of trait categories. To accomplish this objective, four artists were asked to construct 480 stimulus faces, using the composite program FACES, with an eye towards making faces they thought were clearly intelligent, unintelligent, mature, immature, warm, cold, social, unsocial, dominant, submissive, trustworthy, and untrustworthy. Subjects then rated the 480 faces using the same twelve descriptors. Since traits are correlated, this process succeeded in creating trait classes that averaged 111 faces. Single classifiers and ensembles were then trained to match the bipolar extremes of the faces in each of the six trait dimensions of intelligence, maturity, warmth, sociality, dominance, and trustworthiness. With performance measured by the Area under the Receiver Operating Characteristic curve (AROC) and averaged across all six dimensions, results show that single classifiers, espeicially linear support vector machines (0.74) and Levenberg-Marquardt neural networks (0.73), performed as well as human raters (0.77). These single classifiers, however, performed poorly in the trait dimension of maturity. Ensembles of 100 Marquardt neural networks, constructed using bagging (0.76), random subspace (0.77), and class switching (0.75), compared equally well to rater performance, but were better than the single classifiers at handling all six trait dimensions. The Random subspace AROC, averaged across the six dimensions, exactly matched rater performance. We conclude that machine learning methods, especially ensembles, are as capable of perceiveing the social impressions faces make on the general observer as are most human beings. Also included in this chapter are two sections of background material: a brief overview of the relevant person perception literature and a complete tutorial, suitable for the novice, on the single classifier and ensemble systems used in the experiments reported in this study.