Unsupervised methods for image clustering do not yet replace social scientists’ gaze

A number of recent papers have used unsupervised image clustering methods for social science questions. However, there has not yet been research examining whether the clusters resemble the kinds of groupings that social scientists identify when conducting traditional image analysis. In response, we compare the clusters emerging through previously recommended unsupervised image clustering methods in social sciences with the manually categorised groupings to determine which, if any, of these methods are able to imitate the ‘social scientist’s gaze’. We use four different conceptual lenses from previous work, theme, content, purpose and typeand use various metrics to assess the similarities. We show that none of the methods can consistently match manual analysis. Therefore, while unsupervised methods for image clustering can identify visually similar images, this may not be sufficient to capture the full breadth of methodological and theoretical approaches common for social sciences. Following this, we discuss the limitations and opportunities further, suggesting that we might consider unsupervised approaches as a distinct gaze or focus on developing methods which align with previous social scientists’ gazes.