Grant: Discovery Grants Program - Individual

Networks (a.k.a., graphs) are the universal data structures for representing the relationships between interconnected objects. They are ubiquitous in a variety of disciplines and domains ranging from computer science, social science, economics, medicine, to bioinformatics. Representative examples of real-world networks include social networks, airline networks, protein-protein interaction networks, and molecular structures. Network analysis techniques can be used for applications such as recommending friends to users in a social network, predicting the roles of proteins in a protein-protein interaction network, and predicting the properties of molecule structures for discovering new drugs. The interest for analyzing and mining networks has been growing in both academia and industry.Many network-based applications share the common problem of making predictions. This opens many avenues for exploring the use of techniques based on artificial intelligence and machine learning. These methods are exceptionally good at extracting relevant patterns from a vast amount of historical data and using those patterns to make predictions. The first and one of the most critical step towards using machine learning methods is to effectively represent data, which largely determines the prediction performance. Recently, deep learning methods have shown great success in prediction tasks for various domains including computer vision, speech recognition, and natural language understanding. These methods learn effective representations of data by stacking multiple non-linear neural networks to create “deep” neural networks. Even though deep learning techniques are very effective and networks are ubiquitous, deep learning techniques for analyzing networks is an underexplored territory. This research program aims to develop novel deep learning methods for learning effective representations of networks which would empower applications in a variety of disciplines and domains. It is a very challenging problem since most existing techniques of deep learning, such as convolutional neural networks and recurrent neural networks, are developed for data with 2D grid structures (e.g., images) or sequences (e.g., natural languages and speech). However, real-world networks are much more complex, which are large-scale, dynamic, and heterogeneous. This research program aims to overcome these limitations, understand the principles and develop new methods for effectively representing networks. This program would contribute to Canada's lead in artificial intelligence and machine learning by providing the first-mover advantage in the domain of deep learning based graph representation. It would also contribute to our theoretical understanding of deep learning methods and would extend the scope of deep learning to previously unexplored domains and applications.