Structured Probabilistic Models for Computational Social Science

Опубликовано 6 марта 2018, 4:36

With today’s abundance of data, probabilistic models have an opportunity to answer fundamental questions about human behavior and interactions. However, unlike standard inference tasks, socio-behavioral outcomes are frequently interdependent, rely on heterogenous observations, and require complex reasoning. To effectively capture this structure, we need richer modeling frameworks. In this talk, I present my research on developing probabilistic methods that address the needs of computational social science by: 1) making interrelated inferences 2) discovering patterns and causal relationships, and 3) fusing noisy domain knowledge with statistical signals. First, I introduce modeling templates that exploit useful structural patterns to make collective, consistent predictions. Next, I present algorithms that bolster domain knowledge by learning causal relationships and structural patterns directly from data. These approaches have shed light on identifying stances in online debates, detecting social media indicators of alcoholism relapse, and inferring causal links between factors that affect mood. Finally, I outline a research agenda that unifies structured methods with promising embedding and latent variable models to address questions in causality and social science.

See more at microsoft.com/en-us/research/v...