Subject-specific predictive modelling for urban affect analysis

Abstract

Recent developments in crowd-sourced data collection and machine intelligence have facilitated data-driven analyses of the affective qualities of urban environments. While past studies have focused on the commonalities of affective experience across multiple subjects, this paper demonstrates an integrated framework for subject-specific affective data collection and predictive modelling. For demonstration, 10 field observers recorded their affective appraisals of various urban environments along the scales of Liveliness, Beauty, Comfort, Safety, Interestingness, Affluence, Stress and Familiarity. Data was collected through a mobile application that also recorded geo-location, date, time of day, a high resolution image of the user's field of view, and a short audio clip of ambient sound. Computer vision algorithms were employed for extraction of six key urban features from the images - built score, paved score, auto score, sky score, nature score, and human score. For predictive modelling, K-Nearest Neighbour and Random Forest regression algorithms were trained on the subject-specific datasets of urban features and affective ratings. The algorithms were able to accurately assess the predicted affective qualities of new environments based on the specific individual's affective patterns.