Personalizing Affective Stimuli Using a Recommender Algorithm: An Example with Threatening Words for Trauma Exposed Populations

Experimental paradigms used in affective and clinical science often use stimuli such as images, scenarios, videos, or words to elicit emotional responses in study participants. Choosing appropriate stimuli that are highly evocative is essential to the study of emotional processes in both healthy and clinical populations. Selecting one set of stimuli that will be relevant for all subjects can be challenging because not every person responds the same way to a given stimulus. Machine learning can facilitate the personalization of such stimuli. The current study applied a novel statistical approach called a recommender algorithm to the selection of highly threatening words for a trauma-exposed population (N = 837). Participants rated 513 threatening words, and we trained a user–user collaborative filtering recommender algorithm. The algorithm uses similarities between individuals to predict ratings for unrated words. We compared threat ratings for algorithm-based word selection to a random word set, a word set previously used in research, and trauma-specific word sets. Algorithm-selected personalized words were more threatening compared to non-personalized words with large effects (ds = 2.10–2.92). Recommender algorithms can automate the personalization of stimuli from a large pool of possible stimuli to maximize emotional reactivity in research paradigms. These methods also hold potential for the personalization of behavioral treatments administered remotely where a provider is not available to tailor an intervention to the individual. The word personalization algorithm is available for use online (https://​threat-word-predictor.​herokuapp.​com/​).