1 3 Evolving Systems DOI 10.1007/s12530-014-9123-z ORIGINAL PAPER Affect detection from non-stationary physiological data using ensemble classifiers Omar AlZoubi · Davide Fossati · Sidney D’Mello · Rafael A. Calvo Received: 30 April 2014 / Accepted: 4 November 2014 © Springer-Verlag Berlin Heidelberg 2014 base classifier and the alpha factor used to update the mem- ber classifiers of the ensemble. Finally, the corrugator and zygomatic facial EMGs were found to be more reliable measures for detecting the valence component of affect compared to other channels. Keywords Affect · Emotion · Classifier ensembles · Physiological · Non-stationary 1 Introduction There is increased motivation in using physiological sig- nals in affect detection systems that detect either discrete emotional categories or affective dimensions of valence and arousal (Kim and André 2008; Picard et al. 2001; AlZoubi et al. 2011). Physiological responses such as facial muscle activity, skin conductivity, heart activity, and respiration have all been considered as potential markers for recognizing affective states (Whang and Lim 2008). Despite high clas- sification rates achieved under laboratory conditions (Kim et al. 2004; Lichtenstein et al. 2008), the changing nature of physiological signals introduces significant challenges when one moves from the lab and into the real world (Plarre et al. 2011; AlZoubi et al. 2012). In particular, physiological data is expected to exhibit daily variations or non-stationarities (Picard et al. 2001; AlZoubi et al. 2011), which introduce difficulties for building effective classification models on future data (i.e., signals generated by the same individual across time). The ability to integrate automatic affect detec- tion capabilities in computer systems depends largely on the underlying models of affect, and how these models can adapt to the changing nature of physiological data. Previous research has shown that affective physiological data exhibited daily variations (Picard et al. 2001; AlZoubi Abstract Affect detection from physiological signals has received considerable attention. One challenge is that physiological measures exhibit considerable variations over time, making classification of future data difficult. The present study addresses this issue by providing insights on how diagnostic physiological features of affect change over time. Affective physiological data (electrocardiogram, elec- tromyogram, skin conductivity, and respiration) was col- lected from four participants over five sessions each. Clas- sification performance of a number of training strategies, under different conditions of features selection and engi- neering, were compared using an adaptive classifier ensem- ble algorithm. Analysis of the performance of individual physiological channels for affect detection is also provided. The key result is that using pooled features set for affect detection is more accurate than using day-specific features. A decision fusion strategy which combines decisions from classifiers trained on individual channels data outperformed a features fusion strategy. Results also show that the per- formance of the ensemble is affected by the choice of the O. AlZoubi (*) · D. Fossati Computer Science, Carnegie Mellon University in Qatar, Doha, Qatar e-mail: oalzoubi@cmu.edu D. Fossati e-mail: dfossati@cmu.edu S. D’Mello Department of Computer Science, The University of Notre Dame, Notre Dame, IN, USA e-mail: sdmello@nd.edu R. A. Calvo School of Electrical and Information Engineering, The University of Sydney, Sydney, NSW, Australia e-mail: Rafael.Calvo@sydney.edu.au