Prior knowledge of human activities from social data

Abstract

We explore the feasibility of utilizing large, crowd-generated online repositories to construct prior knowledge models for high-level activity recognition. Towards this, we mine the popular location-based social network, Foursquare, for geo-tagged activity reports. Although unstructured and noisy, we are able to extract, categorize and geographically map people's activities, thereby answering the question: what activities are possible where? Through Foursquare text only, we obtain a testing accuracy of 59.2% with 10 activity categories; using additional contextual cues such as venue semantics, we obtain an increased accuracy of 67.4%. By mapping prior odds of activities via geographical coordinates, we directly benefit activity recognition systems built on geo-aware mobile phones.