Abstract
Graph Domain Adaptation (GDA) transfers knowledge from labeled source graphsto unlabeled target graphs, addressing the challenge of label scarcity. In thispaper, we highlight the significance of graph homophily, a pivotal factor forgraph domain alignment, which, however, has long been overlooked in existingapproaches. Specifically, our analysis first reveals that homophilydiscrepancies exist in benchmarks. Moreover, we also show that homophilydiscrepancies degrade GDA performance from both empirical and theoreticalaspects, which further underscores the importance of homophily alignment inGDA. Inspired by this finding, we propose a novel homophily alignment algorithmthat employs mixed filters to smooth graph signals, thereby effectivelycapturing and mitigating homophily discrepancies between graphs. Experimentalresults on a variety of benchmarks verify the effectiveness of our method.