Investigating Robustness and Interpretability of Link Prediction via Adversarial Modifications

Representing entities and relations in an embedding space is a well-studied\napproach for machine learning on relational data. Existing approaches, however,\nprimarily focus on improving accuracy and overlook other aspects such as\nrobustness and interpretability. In this paper, we propose adversarial\nmodifications for link prediction models: identifying the fact to add into or\nremove from the knowledge graph that changes the prediction for a target fact\nafter the model is retrained. Using these single modifications of the graph, we\nidentify the most influential fact for a predicted link and evaluate the\nsensitivity of the model to the addition of fake facts. We introduce an\nefficient approach to estimate the effect of such modifications by\napproximating the change in the embeddings when the knowledge graph changes. To\navoid the combinatorial search over all possible facts, we train a network to\ndecode embeddings to their corresponding graph components, allowing the use of\ngradient-based optimization to identify the adversarial modification. We use\nthese techniques to evaluate the robustness of link prediction models (by\nmeasuring sensitivity to additional facts), study interpretability through the\nfacts most responsible for predictions (by identifying the most influential\nneighbors), and detect incorrect facts in the knowledge base.\n