Graph Convolutional Networks for Determining Online Community Echochambers Isaiah Colwell * Project Mentor(s): Razvan Andonie, PhD Echo chambers in online communities arise when users participate predominantly within a single community, or a small selection of communities, often sharing the same users, generating dense intra- community connections while forming few cross-community ties. Detecting and quantifying this behavior is difficult because it manifests in both the structural connectivity of a community and the content of its discourse; methods relying on either signal alone capture only part of the picture. This work proposes a Graph Convolutional Network (GCN) pipeline for detecting and ranking echo chambers across online communities. A two-layer GCN is trained on a large-scale Reddit dataset to classify posts by community. Through message passing, the model produces post-level embeddings that jointly encode content features and graph topology. These embeddings are aggregated to community- level representations and scored along three dimensions: embedding isolation, internal cohesion, and graph-structural insularity, yielding a ranked assessment of echo chamber severity. Echo chamber rankings are compared against graph-only baselines (cross-edge ratio, Louvain community detection) and a content-only baseline (TF-IDF cosine similarity). The moderate correlation between GCN-based scores and graph-only insularity scores indicates that the learned embeddings capture information beyond raw topology. Rankings are stable across scoring weight profiles, and a hierarchical structure emerges that separates extreme, semi-insular, and mainstream communities. These results suggest that GCN-based joint representations offer a more complete characterization of echo chamber behavior than either structural or content methods alone. This project focuses on solving a routing problem used in the real world. It uses computational optimization techniques in order to determine the most efficient and effective travel routes for both electric vehicles (ECV) and internal combustion vehicles (ICCV) using a PRP dataset that incorporates terrain, fuel consumption, and distance constraints. The difference between traditional traveling salesman problems and this is that this project accounts for environmental and resource-based limitations that directly affect the route efficiency. To approach this, two algorithmic strategies were implemented: ant colony swarm optimization and blind search methods. The ant colony algorithm simulates cooperative behavior found in nature. The ant colony swarm algorithm uses the pheromone levels from ants to find the shortest path based on accumulated knowledge of previous paths. On the other hand, blind search provides a baseline by exploring possible routes without the heuristic guidance. By comparing these methods, the project evaluates differences in efficiency, accuracy, and computational cost. The system was fully implemented in code, allowing for simulation and analysis of route performance under varying conditions. The results show that ant colony swarm and other swarm- based approaches significantly outperform blind search in identifying the optimal route especially when terrain and fuel constraints are included. Overall, this project highlights how advanced algorithms can improve transportation efficiency and decision-making for scenarios in the real world. Presentation Type: Oral Presentation (May 20, 9:30am–5:00pm) Keywords: Optimization, Programming, Algorithms, Ant Colony Swarm, Blind Search SOURCE Form ID: 117 Presentation Type: Oral Presentation (May 20, 9:30am–5:00pm) Keywords: Machine Learning, Graph Networks, Social Networks SOURCE Form ID: 208 Optimizing Electric and Internal Combustion Vehicle Routes Jessica Henry Project Mentor(s): Donald Davendra, PhD
112
Made with FlippingBook interactive PDF creator