A cluster-based approach to selecting representative stimuli from the International Affective Picture System (IAPS) database

Abstract

The International Affective Picture System (IAPS; Lang, Bradley, & Cuthbert, 2008) is a stimulus database that is frequently used to investigate various aspects of emotional processing. Despite its extensive use, selecting IAPS stimuli for a research project is not usually done according to an established strategy, but rather is tailored to individual studies. Here we propose a standard, replicable method for stimulus selection based on cluster analysis, which re-creates the group structure that is most likely to have produced the valence arousal, and dominance norms associated with the IAPS images. Our method includes screening the database for outliers, identifying a suitable clustering solution, and then extracting the desired number of stimuli on the basis of their level of certainty of belonging to the cluster they were assigned to. Our method preserves statistical power in studies by maximizing the likelihood that the stimuli belong to the cluster structure fitted to them, and by filtering stimuli according to their certainty of cluster membership. In addition, although our cluster-based method is illustrated using the IAPS, it can be extended to other stimulus databases.

Keywords: Cluster analysis; Emotion; IAPS; International affective picture system; Stimulus selection.

Figures

Fig. 1

Correlations between the pleasure/valence arousal, and dominance dimensions, with deviations from linearity that give rise to the specific shapes of the relationships

Fig. 2

Various clustering indices indicate different “optimal” values for k. These graphs may change slightly with every run of the clustering algorithm, due to the random seeds that k-means uses. As such, 100,000 repetitions were run on the k-means clustering algorithm each time, with a range for k from 2 to 8, and with the values of the Caliński, Ball, Hartigan, and SSI criteria computed each time (with the Ball criterion having to be minimized, unlike the other three criteria, which must be maximized). The average values for these criteria were then computed across all of the repetitions and indicated (left to right, and top to bottom) that three, eight, eight, and three clusters should be extracted, respectively

Fig. 3

Data structure of the IAPS images. It is worth noting that large portions of the 3-D space remain unpopulated, signaling either that the IAPS does not cover those combinations between valence arousal, and dominance, or that photographic material in general would have difficulty with this

Fig. 4

The amount of dissimilarity (as computed using the R package clue: Hornik, 2005) between cases is accounted for by ever-increasing values for k

Fig. 5

Cluster boxplots, for each dimension. The boxplots indicate, for each cluster (coded by colors), the spread of cases assigned to it, in terms of valence arousal, and dominance. The boxplot widths are proportional to the cluster sample sizes

Fig. 6

Bivariate scatterplots showing the default classifications of cases and the uncertainties provided by Mclust() in R. The uncertainties are coded using one of three symbols: ringed black dots for candidates with a high certainty of cluster membership; orange (light gray) asterisks for less clear cluster memberships; and red (dark gray) squares for cases to avoid using as stimuli, with very unclear memberships. Point size is an additional indicator for the level of classification uncertainty, with larger points indicating higher uncertainty

Fig. 7

Selection of the 20 most likely IAPS cases per cluster, for the k = 5 clustering solution. The color coding was chosen to be consistent with Fig. 8 below

Fig. 8

Selection of the five most likely IAPS cases per cluster, for the k = 5 clustering solution, along with IAPS image codes. The color coding was chosen to be consistent with Fig. 7 above