Date of Award


Document Type

Thesis (Ph.D.)

Department or Program

Department of Computer Science

First Advisor

V.S. Subrahmanian


Human behaviors in a group setting involve a complex mixture of multiple modalities: audio, visual, linguistic, and human interactions. With the rapid progress of AI, automatic prediction and understanding of these behaviors is no longer a dream. In a negotiation, discovering human relationships and identifying the dominant person can be useful for decision making. In security settings, detecting nervous behaviors can help law enforcement agents spot suspicious people. In adversarial settings such as national elections and court defense, identifying persuasive speakers is a critical task. It is beneficial to build accurate machine learning (ML) models to predict such human group behaviors. There are two elements for successful prediction of group behaviors. The first is to design domain-specific features for each modality. Social and Psychological studies have uncovered various factors including both individual cues and group interactions, which inspire us to extract relevant features computationally. In particular, the group interaction modality plays an important role, since human behaviors influence each other through interactions in a group. Second, effective multimodal ML models are needed to align and integrate the different modalities for accurate predictions. However, most previous work ignored the group interaction modality. Moreover, they only adopt early fusion or late fusion to combine different modalities, which is not optimal. This thesis presents methods to train models taking multimodal inputs in group interaction videos, and to predict human group behaviors. First, we develop an ML algorithm to automatically predict human interactions from videos, which is the basis to extract interaction features and model group behaviors. Second, we propose a multimodal method to identify dominant people in videos from multiple modalities. Third, we study the nervousness in human behavior by a developing hybrid method: group interaction feature engineering combined with individual facial embedding learning. Last, we introduce a multimodal fusion framework that enables us to predict how persuasive speakers are.

Overall, we develop one algorithm to extract group interactions and build three multimodal models to identify three kinds of human behavior in videos: dominance, nervousness and persuasion. The experiments demonstrate the efficacy of the methods and analyze the modality-wise contributions.