From Physical to Social Sensing: Analyzing User-generated Content for Better Sensing

While physical sensing has been used for a long time to observe events and natural phenomena, social sensing has emerged as a new paradigm that uses humans as “sensors” (also known as social sensors) to report their observations about the physical world. Social sensing transforms how we sense the world and complements physical sensing by substantially extending the horizon we know about the world in real time. This thesis addresses two main challenges in the sensing area: (1) Proposing a scalable framework for searching physical sensors; and (2) Enabling social sensing through efficient methods for predicting the geolocation of social sensors. While the number of physical sensors deployed around the world is growing at a rapid pace, an important challenge is to determine which physical sensors should be selected to retrieve the desired data for monitoring, detecting or discovering a phenomenon. Considering the Internet of Things (IoT) as an enabler for physical sensing, we proposed a scalable context-aware sensor search framework to deal with the dramatic increase of sensors and consequently the data collected by those sensors in the future IoT. We also perform novel adjustments to increase the efficiency of our system and maintain its accuracy against dynamic nature of IoT environment. Experimental results over large scale datasets show the efficiency and scalability of our proposed framework. Social sensors, which are social media users, can react to phenomena and events of many kinds such as earthquake by making posts about it. Consequently, social sensing enables a wide range of new applications and services such as rapid disaster response and opinion analysis. Social media users should provide their locations to be useful as social sensors, however, majority of them do not declare their locations. To address this issue, we have focused on user geolocation which is the task of predicting the location of a user from available sources of information, such as text posted by that individual, or network relationships with other individuals. To this end, we propose novel predictive models based on the following intuitions: 1) People who are well-known in their local communities (local celebrities) can be used as important clues to geolocate their followers; 2) Users who live nearby are more likely to influence each other and discuss more local topics; hence, semantic similarity between users' posts can predict their geographic proximity. Moreover, we present an efficient joint model based on the fusion of neural networks to incorporate different types of available information including text, users' social relationships, and metadata fields ebmeded in tweets and profiles such as time zone and user language. We have utilized four standard Twitter geolocation benchmark datasets to evaluate our methods. The results indicate that the proposed methods significantly improve the performance of user geolocation in terms of accuracy and mean/median error distances, and yield state-of-the-art results over all datasets