STD-Explain: Generalizing explanations for spatio-temporal graph convolutional networks based on spatio-temporal decoupled perturbation

Spatio-temporal graph convolutional networks utilize an alternating combination of one-dimensional ordinary convolution and graph convolution to extract spatio-temporal features. This alternation intertwines temporal and spatial features closely, leading to a tight coupling between them. The presenc...

Celý popis

Uloženo v:

Podrobná bibliografie
Vydáno v:	Neurocomputing (Amsterdam) Ročník 657; s. 131539
Hlavní autoři:	Li, Yanshan, Shi, Ting, He, Suixuan, Chen, Zhiyuan, Zhang, Li, Yu, Rui, Xie, Weixin
Médium:	Journal Article
Jazyk:	angličtina
Vydáno:	Elsevier B.V 07.12.2025
Témata:	Explainability algorithm Human skeleton action recognition Spatio temporal graph convolutional networks Spatio-temporal decoupling Human skeleton action recognition Spatio-temporal decoupling Explainability algorithm Spatio temporal graph convolutional networks
ISSN:	0925-2312
On-line přístup:	Získat plný text
Tagy:	Přidat tag Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!

Popis
Shrnutí:	Spatio-temporal graph convolutional networks utilize an alternating combination of one-dimensional ordinary convolution and graph convolution to extract spatio-temporal features. This alternation intertwines temporal and spatial features closely, leading to a tight coupling between them. The presence of spatio-temporal coupling complicates the analysis of spatio-temporal data, posing challenges for existing explainability algorithms to effectively separate and interpret these intertwined features. Therefore, we propose STD-Explain, an explainable algorithm based on spatio-temporal decoupled perturbation, which employs a two-stage perturbation approach considering subgraph and node-level explanations. Firstly, targeting the spatio-temporal coupling issue in spatio-temporal graph convolutional networks, the algorithm proposes a temporal perturbation algorithm based on Slice Graph and a spatial perturbation algorithm aimed at important subgraph node features. Secondly, to avoid introducing additional semantic information when extracting temporal subgraphs, we propose a method for generating temporal subgraphs in spatio-temporal decoupling, slicing human skeleton sequences with discrete masks to ensure each subsequence maintains spatial structure integrity without introducing additional edges. Furthermore, to ensure the maximum correlation between the interpreted subgraphs and model predictions, we propose a temporal important subgraph discrimination strategy to select the most relevant subgraphs to model predictions. Experimental results demonstrate that STD-Explain performs well in qualitative and quantitative analysis.
ISSN:	0925-2312
DOI:	10.1016/j.neucom.2025.131539