Context-aware group services for proximity-based mobile social networking

Abstract

As social creatures, people spend a lot of time with other people in public places. People can have social interactions in different forms, either long-term or temporary. For instance, in the campus scenario, students who are members of a special interest club may have regular social interactions such as weekly meetings, thus they form a persistent group; in contrast, students who are participating in a poster session might only interact at the session, thus they form a transient group. Unobtrusively recognizing these groups can help automatically perform certain group-specific tasks such as information recommendation, thereby enabling more intelligent social applications especially in pervasive computing environments. Sensor-equipped mobile devices allow users to participate in various social networking services. In this thesis, we consider proximity-based mobile social networking environments where users can share information using peer-to-peer communications via their mobile devices when they are in proximity. Existing research of recognizing groups in proximity-based mobile social networks has not been fully exploited. For persistent groups, existing work clusters mobile users into communities merely based on contact frequency and duration when users are in proximity, missing the location and time contexts of these contacts. For transient groups, existing work is focused on crowd behavior recognition based on people's movement patterns, hence failing to identify stationary groups; there is work on identify transient groups performing similar activities based on sensor data distributions, but it only uses a single sensor modality, hence it cannot distinguish groups with more fine-grained activity similarity. In this thesis, we address these limitations in group recognition using mobile devices. Our basic approach is to first utilize the sensors on mobile devices to extract the relevant contexts of mobile users, then apply clustering algorithms on the social graph constructed from user contexts to form groups, and finally build applications based on the automatically recognized groups. More specifically, we identify persistent information influence-based groups based on space-time contexts of information exchanges, proximity-based transient intentional groups using a probability-based approach based on Bluetooth signal strength, and sub-groups in activity similarity-based groups using a fusion of multimodal sensor data. We use different applications to demonstrate how their performances are improved by using these recognized groups and also build mobile apps to test the grouping accuracy.