Date of Award

1-21-2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Information Science

First Advisor

Nitin Agarwal

Abstract

YouTube has become a dominant arena for global information sharing, where creators and audiences interact through content, comments, and engagement dynamics. Understanding how these interactions shape a channel’s identity requires a comprehensive characterization of both audience behavior and content structure. This dissertation develops a unified framework for characterizing YouTube channels through multimodal analysis that integrates behavioral modeling, dimensionality reduction, clustering, and content-based characterization to provide a holistic view of audience and editorial patterns. The framework begins by analyzing the structural relationships within co-commenter networks, where users who repeatedly comment on the same videos are connected to capture patterns of interaction and community formation. From these networks, a diverse set of structural and clique-based features is extracted to represent audience behavior. To simplify and interpret these high-dimensional features, dimensionality reduction techniques are employed. Principal Component Analysis (PCA) is first applied to identify linear patterns and dominant axes of variation. Next, Graph2Vec embeddings are used to generate vectorized representations of network topology, preserving relational similarities among commenters. These embeddings are further refined using Uniform Manifold Approximation and Projection (UMAP), which retains non-linear proximity relationships while improving visualization and separability. The resulting representations are clustered using algorithms such as K-Means and hierarchical clustering, enabling the discovery of coherent groups of channels that exhibit similar interaction structures and participation tendencies. Building on this analytical foundation, the framework advances to a comprehensive characterization of channels across behavioral and content dimensions. Commenter feature combinations including sentiment, vocabulary diversity, posting cadence, toxicity, and other linguistic or temporal indicators are analyzed in pairwise spaces to identify recurring behavioral profiles. Content characterization is conducted by measuring the semantic alignment among video metadata fields such as titles, descriptions, transcripts, and categories, capturing the degree of thematic and editorial coherence. Inner-content characterization extends this approach within each channel, assessing how consistently content elements are maintained across all videos. To incorporate visual information, video barcodes are generated as compact fingerprints summarizing stylistic and aesthetic regularity or variation across uploads. Clustering is then applied across these behavioral and content features, and results are consolidated through majority voting to ensure stability and reproducibility. Euclidean distance aggregation across feature pairs identifies recurring channel groups that remain consistently close across modalities, revealing archetypes such as behaviorally cohesive yet content-diverse or engagement-stable yet visually inconsistent. After establishing these characterizations, the framework extends to quantifying behavioral deviations through the computation of specialized scores. Commenter behavior scores are estimated using kernel density estimation (KDE) and Gaussian mixture models (GMM), with discriminative features weighted by effect size to emphasize the most informative behavioral indicators. In parallel, engagement scores are derived by analyzing deviations in relationships among core metrics such as views, comments, subscribers, and videos. These two perspectives, commenter and engagement, are integrated at multiple levels: first at the feature level through cosine similarity and PCA loadings, and then at the output level using ensemble aggregation methods including harmonic mean, weighted average with interaction, and agreement-weighted maximum scoring. This dual-layer integration enhances the robustness and interpretability of the analysis and enables validation against channels that were later suspended by YouTube. Overall, this study introduces a comprehensive and scalable framework for characterizing YouTube channels through a multimodal synthesis of structural, behavioral, and content-based analyses. By integrating dimensionality reduction, clustering, and multi-feature characterization, the framework uncovers distinct channel archetypes and interaction patterns. The results demonstrate how audience coordination, editorial coherence, and engagement structure jointly shape visibility, influence, and discourse across digital media ecosystems.

Download

Share

COinS