Date of Award

Summer 9-1-2021

Document Type

Thesis (Ph.D.)

Department or Program

Department of Computer Science

First Advisor

V.S. Subrahmanian


In this thesis, we consider the task of inferring the social interactions between humans by analyzing multi-modal data. Specifically, we attempt to solve some of the problems in interaction analysis, such as long-term deception detection, political deception detection, and impression prediction. In this work, we emphasize the importance of using knowledge about the group structure of the analyzed interactions. Previous works on the matter mostly neglected this aspect and analyzed a single subject at a time. Using the new Resistance dataset, collected by our collaborators, we approach the problem of long-term deception detection by designing a class of histogram-based features and a novel class of meta-features we callLiarRank. We develop a LiarOrNot model to identify spies in Resistance videos. We achieve AUCs of over 0.70 outperforming our baselines by 3% and human judges by 12%. For the problem of political deception, we first collect a dataset of videos and transcripts of 76 politicians from 18 countries making truthful and deceptive statements. We call it the Global Political Deception Dataset. We then show how to analyze the statements in a broader context by building a Video-Article-Topic graph. From this graph, we create a novel class of features called Deception Score that captures how controversial each topic is and how it affects the truthfulness of each statement. We show that our approach achieves 0.775 AUC outperforming competing baselines. Finally, we use the Resistance data to solve the problem of dyadic impression prediction. Our proposed Dyadic Impression Prediction System (DIPS) contains four major innovations: a novel class of features called emotion ranks, sign imbalance features derived from signed graphs theory, a novel method to align the facial expressions of subjects, and finally, we propose the concept of a multilayered stochastic network we call Temporal Delayed Network. Our DIPS architecture beats eight baselines from the literature, yielding statistically significant improvements of 19.9-30.8% in AUC.