Generative Causal Explanations for Graph Neural Networks

                      Wanyu Lin 1 2 Hao Lan 2 Baochun Li 2

              Abstract                Despite their practical success, most GNNs are deployed as
    This paper presents Gem, a model-agnostic ap-       black boxes, lacking explicit declarative knowledge repre-
    proach for providing interpretable explanations      sentations. Therefore, they have difficulty in generating the
    for any GNNs on various graph learning tasks.   ...