Various attribution methods have been developed to explain deep neural networks (DNNs) by inferring the attribution/importance/contribution score of each input variable to the final output. However, existing attribution methods are often built upon different heuristics. There remains a lack of a unified theoretical understanding of why these methods are effective and how they are related. Furthermore, there is still no universally accepted criterion to compare whether one attribution method is preferable over another. In this paper, we resort to Taylor interactions and for the first time, we discover that fourteen existing attribution methods, which define attributions based on fully different heuristics, actually share the same core mechanism. Specifically, we prove that attribution scores of input variables estimated by the fourteen attribution methods can all be mathematically reformulated as a weighted allocation of two typical types of effects, i.e. , independent effects of each input variable and interaction effects between input variables. The essential difference among these attribution methods lies in the weights of allocating different effects. Inspired by these insights, we propose three principles for fairly allocating the effects, which serve as new criteria to evaluate the faithfulness of attribution methods. In summary, this study can be considered as a new unified perspective to revisit fourteen attribution methods, which theoretically clarifies essential similarities and differences among these methods. Besides, the proposed new principles enable people to make a direct and fair comparison among different methods under the unified perspective.
In graph modeling, scarcity of labeled data is a challenging issue. To address this issue, state-of-the-art graph models learn the representation of graph data via contrastive learning. Those models usually use data augmentation techniques to generate positive pairs for contrastive learning, which aims to maximize the similarity of positive data pairs while minimizing the similarity of negative data pairs. However, samples with the same labels may be separately mapped in the feature space. To solve this problem, we introduce a novel model called Clustering Contrastive Graph Neural Network (CC-GNN), which develops a new kind of grouped contrastive learning that maximizes the similarity of positive data groups and minimizes the similarity of negative groups. That is, contrastive learning is defined on a group level rather than on an instant level. We assert that parameters learned by this kind of contrastive learning will lead to better performance of graph neural networks for downstream classification tasks. We combined the clustering contrastive learning technique with three baseline GNN models for graph classification. We found that the performance of these models was significantly improved, which strongly supports our assertion. We also testified the models for node classification on three popular citation networks. Finally, we conducted an ablation study to analyze how the clustering contrastive learning influence the performance of a graph model.