Generalized Learning Models for Structured Data
Structures are present in almost everything around us. In most of the systems that we interact with, or the way we interact with them, some emergent structure can often be observed. A simple sentence is a sequence of words. A small classroom of interacting students can be depicted as a network with each student defining a node of it. The emergent structures, therefore, highlight the inter-relatedness of different entities within systems, where while each entity has a significant individuality, it is also a component of a larger structure. This structural information, combined with the individual knowledge, can assist the task of learning properties in such systems. On a social network, for instance, we can learn link related properties between users by learning from the users as well as the graph of several users on the same network. Similarly, in an interactive sequence of a click-stream on a system, using the ordered information of these click actions, we may be able to learn the intent of a user performing the clicks.
In this dissertation, we present the concept, methodology, and experiments for performing generalized learning from structural data. We discuss the emergence of structures within datasets, and the entire approach to learn from those structures. We provide a methodology for capturing the structures and assembling the information hidden within the structures. We present a concept of neural aggregation which helps combine information from complex structures while ensuring the learning capability of the models. We present several neural network based architectures for learning different properties from sequential and graphical structures. The dissertation provides the general approach, as well as specific learning frameworks for problems and datasets across several application domains.