Identifying salient topics for personalized place similarity

The ability to find similar places is an important component to geographic information retrieval applications as varied as travel recommendation services, marketing analysis tools, and socio-ecological research. Using generative topic modelling on a large collection of place descriptions, we can represent places as distributions over thematic topics, and quantitatively measure similarity for places modelled with these topic signatures. However, existing similarity measures are context independent; in cognitive science research there exists evidence that when people perform similarity judgments they will weigh properties differently depending on personal context. In this paper we present a novel method to re-weight the topics that are broadly associated with a place, based on users’ interests inferred from sample place similarity rankings. We evaluate the method by training topics associated with texts about places, and perform a user study that compares user-provided similar places to those provided by automatically personalised place rankings. The results demonstrate improved correspondence between user rankings and automated rankings when personalised weights are applied.