There has been an increasing trend of performing inference on data collected by smartphones to provide context-aware location-based services. When this inference is performed using supervised analysis, these services need ground truth if high accuracies are desired. While accuracy is less of a concern for services targeted at individuals, it is important when individual data is aggregated for semantic analysis of a population. However, traditional techniques for obtaining ground truth such as paid crowdsourcing are challenging in this domain since the ground truth is uniquely available to the user. Therefore, the user needs to be the source of ground truth for these services.

This motivates the need for Participatory Classification, a framework that is able to satisfy the need for minimally invasive, ongoing, ground truth collection from regular users at scale. We present an architecture that can be used to enable this framework for such services, and evaluate the framework in the context of an end-to-end prototype that we built. The prototype minimizes the burden on the user while classifying trips by travel mode, and uses the classified trips to generate a personalized carbon footprint for the user and aggregate data such as commute mode share, for use by urban planners. With this prototype, we collected 7439 labelled sections from 44 unpaid volunteers over a total period of 3 months.




Download Full History