Accuracy–Energy Trade-offs in Heterogeneous Graph Learning with Runtime Analysis

Supervisor: Haleh Dizaji

Author: Balázs Bozsó

Abstract

Heterogeneous graph learning methods, such as relational graph neural networks, have shown strong performance on complex, multi-typed data by explicitly modeling node and relation diversity. However, these gains come at a non-trivial computational cost, raising important questions about their efficiency and environmental impact. This thesis presents a systematic evaluation of accuracy–energy trade-offs in heterogeneous graph learning, comparing representative models, including relational graph convolutional networks, relational graph attention networks, and graph embedding methods, on node classification tasks across benchmark datasets such as DBLP, ACM, IMDB, and OGB-MAG. In addition to predictive performance, the study measures runtime, energy consumption, and CO₂ emissions using hardware-level monitoring tools, enabling a comprehensive assessment of computational cost.

Beyond aggregate comparisons, the analysis investigates how graph size and structural characteristics, such as relation diversity and degree patterns, influence the effectiveness and efficiency of heterogeneous models. By jointly analyzing accuracy, runtime, and energy consumption, this work identifies conditions under which the additional complexity of heterogeneous graph learning is justified, contributing to a more sustainable and informed use of graph-based machine learning methods.