Research

The first systematically generalizable and scalable reasoning model on KG. It models one-step logic inference as triangle update on graph inspired by logic programming and formalize it under the designed GNN framework. By recursively applying triangle updates, the model could reason with logic rules of any length with strong scalability. It can also identify reasonable patterns and conduct partial reasoning with the help of auxiliary edges on graphs.

A model-agnostic enhancement pipeline for small molecule property prediction. It uses a encoder and a retriever to retrieve similar molecules out of the whole dataset as supportive evidence. It also introduces a refinement stage to utilize only supportive evidence for downstream task. It will soon be sealed as a toolkit.

A graph-matching model with theoretical guarantees. It proposes that checking the condition of a rule is equivalent to performing subgraph isomorphism tests. It trains a graph neural network (GNN) on the line graph and exploited the trained model with matching-based grading. The extracted rules server as a tool for the interpretability analysis.

A well-functioning pretrain-finetune paradigm on knowledge graph with great generalizability and interpretability. It formulates complex logical queries as masked predictions on graph patterns and introduces a two-stage masked pre-training strategy. It also proposes a KG triple transformation method to enable transformer to handle KG elegantly and a mechanism to unify different tasks of knowledge graph problems.