Representation, Inference and Learning in Structured Statistical Models

Addressing inherent uncertainty and exploiting structure are fundamental to understanding, designing and making predictions in large-scale information, biological and socio-technical systems. Statistical relational learning (SRL) builds on principles from probability theory and statistics to address uncertainty while incorporating tools from logic to represent structure. SRL methods are especially well-suited to domains where the input is best described as a large multi-relational network, such as online social media and communication networks, and we need to make structured predictions.

The first part of the tutorial will provide an introduction to key SRL concepts, including relational feature construction and representation, inference and learning methods for "lifted graphical models." The second part of the tutorial will describe three important challenges in network analysis: graph identification (inferring a graph from noisy observations), graph alignment (mapping components in one graph to another) and graph summarization (clustering the nodes and edges in a graph). I will overview approaches to these problems based on SRL methods, describe available datasets, and highlight opportunities for future research.

Throughout, I will pay particular attention to scaling and make connections to related areas of machine learning such as structured prediction and latent factor models.