Talk by Meher Chaitanya on Adjacency Search Embeddings

On February 14, 2025, 13:30-14:30, we are happy to host a talk by Meher Chaitanya adjacency search embeddings. The talk will take place in Erzherzog-Johann-Platz 1, 1040 Wien, in Seminarraum FB 02 10 (second floor). You are most welcome to join us.

Title: Adjacency Search Embeddings

Speaker: Meher Chaitanya, ETH Zurich

Abstract: In this talk, I will introduce Adjacency Search Embeddings inspired by threshold models in opinion dynamics. Specifically, I will discuss Maximum Adjacency Search (MAS) and Threshold-based Adjacency Search (TAS), which utilize both a node and a subset of its neighborhood to identify nodes that are well-integrated into cohesive network structures. This provides valuable context for generating higher-order representations. When combined with the skip-gram model, our approaches outperform other shallow embedding techniques in tasks like link prediction and node classification. We substantiate the applicability of our approaches, shedding light on their aptness for specific graph scenarios. By integrating our mechanisms as a preprocessing step, we achieve notable improvements in node classification performance across various GNNs - including GCN, GraphSage, and Gatv2 - on both attributed and non-attributed networks. Additionally, we demonstrate how these methods can serve as positional encoders in graph transformers.

About: Meher recently finished his Ph.D. at ETH Zurich under the supervision of Ulrik Brandes. His research is about mathematical models of social influence and machine learning on graphs.