Cooperative Techniques Supporting Sensor-Based People-Centric Inferencing

Abstract

People-centric sensor-based applications targeting mobile device users offer enormous potential. However, learning inference models in this setting is hampered by the lack of labeled training data and appropriate feature inputs. Data features that lead to better classification models are not available at all devices due to device heterogeneity. Even for devices that provide superior data features, models require sufficient training data, perhaps manually labeled by users, before they work well. We propose opportunistic feature vector merging, and the social-network-driven sharing of training data and models between users. Model and training data sharing within social circles combine to reduce the user effort and time involved in collecting training data to attain the maximum classification accuracy possible for a given model, while feature vector merging can enable a higher maximum classification accuracy by enabling better performing models even for more resource-constrained devices. We evaluate our proposed techniques with a significant places classifier that infers and tags locations of importance to a user based on data gathered from cell phones.