On cross-domain social semantic learning

Approximately 2.4 billion people are now connected to the Internet, generating massive amounts of data through laptops, mobile phones, sensors and other electronic devices or gadgets. This massive explosion of data provides tremendous opportunity to study, model and improve conceptual and physical systems from which the data is produced. Making sense of this data algorithmically can be a complex process, specifically due to two reasons. Firstly, the data is generated by different devices, capturing different aspects of information and resides in different web resources/ platforms on the Internet. Therefore, even if two pieces of data bear singular conceptual similarity, their generation, format and domain of existence on the web can make them seem considerably dissimilar. Secondly, since humans are social creatures, the data often possesses inherent but murky correlations, primarily caused by the causal nature of direct or indirect social interactions. The main objective of this dissertation is to develop learning algorithms that can identify specific patterns in one domain of data which can consequently augment predictive performance in another domain. Our work presents a series of solutions to address the key challenges in cross-domain learning, particularly in the field of social and semantic data. We propose the concept of bridging media from disparate sources by building a common latent topic space. This allows information transfer between social and non-social domains, fostering real-time socially relevant applications. Using these various disparate data from different domains, my dissertation aims to assert that intelligent learning is a mixture of two parts: combinatorial knowledge representation from diverse data, and transferring the gained knowledge appropriately to tackle a new task which could not be solved elegantly without the synergy.