Collaborative Research: III: Small: Graph-Oriented Usable Interpretation

University Of Virginia Main Campus, Charlottesville VA

Investigators

Abstract

Interpretation holds great promise in gaining the trust of end-users by understanding how machine learning models work. In graph-based machine learning, although various interpretation methods have been proposed, the potential of interpretation has not been fully unleashed to make it a really useful tool. For example, existing interpretation methods can identify the important graph components (e.g., subgraph patterns and node features) given a model prediction, but they are not well equipped to shed light on other critical model properties, especially trustworthiness (e.g., fairness and robustness) that is crucial in many real-world applications. In addition, although the interpretation of graph models provides friendly visualization to humans for understanding, it remains nascent how the interpretation will inform the design of better models. To bridge the gap, this project takes a paradigm shift from traditional interpretation methods development, aiming to improve the usability of interpretation in graph learning system deployment, model training and data preparation. The results of this project will boost the overall value of interpretation in graph-based information systems. Furthermore, this research will play an integral part in educating and training undergraduate and PhD students. It will also be tightly integrated with multiple courses related to data mining and machine learning. This project aims to systematically explore usable interpretation in three different stages of a graph learning pipeline in backward order, ranging from system diagnosis, model improvement, back to data refinement. The project approaches interpretability through a novel perspective, which goes beyond conventional paradigms of simply understanding model predictions, towards explaining higher-level model properties and exploring how models could actually benefit from interpretation. First, it develops post-hoc interpretation tools to diagnose trustworthiness of graph learning models in various aspects, including fairness, robustness, and causality. Second, it develops interpretation-guided training algorithms and textual generative modules to comprehensively improve graph learning models in terms of effectiveness, robustness, and interactivity. Third, it utilizes interpretation to refine graph data from two complementary directions, including graph augmentation via a counterfactual Mixup strategy and graph compression via data distillation, which provide the fundamental basis of effective and efficient graph learning. The project will also result in the dissemination of shared data and open-source software to broader data mining and graph machine learning communities. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.