Cooperative Explanations of Graph Neural Networks

Fang, Junfeng; Wang, Xiang; Zhang, An; Liu, Zemin; He, Xiangnan; Chua, Tat-Seng

doi:10.1145/3539597.3570378

Cited by 13 publications

(12 citation statements)

References 16 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The existing GIB framework could work for simple synthetic datasets by relying on the implicit knowledge associated with the class and assuming a large decision margin between the two or more classes. However, in more practical scenarios like MUTAG, the existing approximation may be heavily affected by the distribution shifting problem [11,26].…”

Section: Generalized Gibmentioning

confidence: 99%

“…Examples of such methods include GNNExplainer [50], which determines the importance of nodes and edges through perturbation, and PGExplainer [24], which trains a graph generator to incorporate global information. Recent studies in the field [11,30] also contribute to the development of these methods. Post-hoc explainability methods can be classified under a label-preserving framework, where the explanation is a substructure of the original graph and preserves the information about the predicted label.…”

Section: Introductionmentioning

confidence: 99%

“…[28] addresses the distribution shifting issue in image explanation via a module that quantifies affinity between perturbed data and original dataset distribution. In the graph domain, while a recent work [11] attempts to address distribution shifting by annealing the size constraint coefficient at the start of the explanation process, the distribution shifting issue still persists throughout the explanation process.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

MixupExplainer: Generalizing Explanations for Graph Neural Networks with Data Augmentation

Zhang

Luo

Wang³

2023

Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

View full text Add to dashboard Cite

Graph Neural Networks (GNNs) have received increasing attention due to their ability to learn from graph-structured data. However, their predictions are often not interpretable. Post-hoc instance-level explanation methods have been proposed to understand GNN predictions. These methods seek to discover substructures that explain the prediction behavior of a trained GNN. In this paper, we shed light on the existence of the distribution shifting issue in existing methods, which affects explanation quality, particularly in applications on real-life datasets with tight decision boundaries. To address this issue, we introduce a generalized Graph Information Bottleneck (GIB) form that includes a label-independent graph variable, which is equivalent to the vanilla GIB. Driven by the generalized GIB, we propose a graph mixup method, MixupExplainer, with a theoretical guarantee to resolve the distribution shifting issue. We conduct extensive experiments on both synthetic and real-world datasets to validate the effectiveness of our proposed mixup approach over existing approaches. We also provide a detailed analysis of how our proposed approach alleviates the distribution shifting issue. CCS CONCEPTS• Computing methodologies → Neural networks; Artificial intelligence; • Human-centered computing → Human computer interaction (HCI).

show abstract

Section: Generalized Gibmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

MixupExplainer: Generalizing Explanations for Graph Neural Networks with Data Augmentation

Zhang

Luo

Wang³

2023

Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

View full text Add to dashboard Cite

show abstract

“…Several pioneering studies have attempted to address this distributional challenge (Fang et al, 2023d;Zhang et al, 2023b;a). For example, CGE regards the GNN model as a transparent, fully accessible system.…”

Section: Preprint (A) Explanation Process (B) Ood Problemmentioning

confidence: 99%

“…For example, CGE regards the GNN model as a transparent, fully accessible system. It considers the GNN model as a teacher network and employs an additional "student" network to predict the labels of explanation subgraphs (Fang et al, 2023d). As another example, MixupExplainer (Zhang et al, 2023b) generates a mixed graph for the explanation by blending it with a non-explanatory subgraph from a different input graph.…”

Section: Preprint (A) Explanation Process (B) Ood Problemmentioning

confidence: 99%

hECA: the cell-centric assembly of a cell atlas

Chen

Luo

Gao

et al. 2021

Preprint

View full text Add to dashboard Cite

The significance of building atlases of human cells as references for future biological and medical studies of human in health or disease has been well recognized. Comparing to the rapidly accumulation of single-cell data, there has been fewer published work on the information structure to assemble cell atlases, or on methods for using reference atlases once they are ready. Most existing cell atlas work organize single-cell gene expression data as a collection of individual files, allowing users to download selected data sheets, or to annotate query cells using models pretrained with the collected data. These features are useful as the basic use of cell atlases. More comprehensive uses of global cell atlases can be developed once data of cells from multiple organs across different studies can be assembled into one orchestrated data repository rather than a collection of data files. For this purpose, we presented a unified giant table or uGT to store and organize single-cell data from multiple studies into a single huge data repository, and a unified hierarchical annotation framework or uHAF to annotate cells from uncoordinated studies. Based on these technologies, we developed a system that enables users to design complex rules to recruit from the atlas cells that meet certain conditions, such as with desired expression range of a gene or multiple genes and with required organ, tissue or developmental origins, across multiple datasets that were otherwise unconnected. The conditions can be expressed as sophisticated logic criteria to pinpoint specific cells that cannot be easily spotted in traditional in vivo or in vitro cell sorting or in traditional searching in published data. We call this technology as in data cell sorting from cell atlases. With the increasing coverage of the cell atlas, this in data experiment paradigm will facilitate scientists to conduct investigations in the data space beyond the restrictions in traditional in vivo and in vitro experiments. In the current work, we collected scRNA-seq data of more than 1 million human cells from scattered studies and assembled them as a human Ensemble Cell Atlas or hECA using the proposed information structure, and provided comprehensive tools for in data experiments based on the atlas. Case examples on agile construction of atlases of particular cell types and on studying the side effects of targeted immune therapy drugs showed that in data cell sorting is an efficient and effective way for comprehensive discoveries. hECA provides a powerful platform for assembling massive scattered single-cell data into a unified atlas, and can serve as a prototype for building future cell atlases.

show abstract