Representation Learning with Relational Data

点击翻译按钮获取中文摘要

Graphs are a ubiquitous data structure employed extensively within computer science and related fields. Social networks, molecular graph structures, biological protein-protein networks, recommender systems-all of these domains and many more can be readily modeled as graphs, which capture relations (i.e., edges) between individual entities (i.e., nodes). However, graphs are not only useful as structured knowledge repositories: many machine learning applications seek to make predictions or discover new patterns using graph-structured data as input. For example, one might wish to classify the role of a protein in a biological interaction graph, recommend new friends to a user in a social network, or predict new therapeutic applications of existing drug molecules whose structure can be represented as a graph. The central problem in machine learning with graphs is finding a way to encode information about graph-structure into a machine learning model. For example, in the case of link prediction in a social network, one might want to encode pairwise properties between nodes, such as relationship strength or the number of common friends. Or in the case of classifying a protein's role in a biological interaction network, one might want to include information about the structure of the protein's local graph neighborhood. In my research program, I will explore a nascent, but quickly developing class of approaches to machine learning with graphs: approaches based on graph representation learning (GRL). The key idea behind these approaches is to embed nodes, or entire (sub)graphs, as points in a learned low-dimensional vector space and to use neural networks to reason about relational interactions in these learned vector spaces. Whereas traditional approaches would extract graph statistics as a pre-processing step before applying standard machine learning algorithms, GRL approaches directly learn using graph-structured data in an end-to-end fashion. While still in their nascency, these methods have shown considerable promise across numerous graph analysis tasks. Due to the ubiquity of graph-structured data, GRL has wide range of potential applications, ranging from chemical synthesis to social network analysis. In previous research, I have developed GRL methods to predict drug-disease interactions, to model complex social interactions in web forums, and to power a production-scale recommender system at Pinterest Inc. Over the coming years, I will continue to pursue these core application themes, especially applications related to computational social science and computational biology. Many domain scientists in biology and the social sciences now possess massive troves of structured data but lack the computational tools to effectively understand and use it. A key focus of my research will be developing GRL-based models that can aid such domain experts in leveraging this data for large-scale knowledge discovery.