To Learn is to Wander: Learning Across Graphs and Tasks with Random Walks
Wander frames graph learning as filling in a partially observed graph, using random walks as the shared interface.
The paper says that lets one pretrained checkpoint work across homogeneous graphs, multi-relational graphs, different feature spaces, labels, schemas, and tasks. At inference time, Wander can take in more structural context without changing its learned parameters. The authors report state-of-the-art or competitive results on node classification, homogeneous link prediction, and knowledge-graph link prediction. Joint pretraining is described as preserving specialized performance while enabling positive transfer across graph settings. ArXiv · AI/CL/LG's note
The paper says that lets one pretrained checkpoint work across homogeneous graphs, multi-relational graphs, different feature spaces, labels, schemas, and tasks. At inference time, Wander can take in more structural context without changing its learned parameters. The authors report state-of-the-art or competitive results on node classification, homogeneous link prediction, and knowledge-graph link prediction. Joint pretraining is described as preserving specialized performance while enabling positive transfer across graph settings. ArXiv · AI/CL/LG's note
score 5