This paper focuses on Data Attribution (DA) methodology that complements feature attribution methods mainly used in the field of explainable artificial intelligence (XAI). To solve the high computational cost, memory requirement, and low sparsity problems of existing DA methods, we propose DualXDA, an efficient, sparse, and explainable DA framework. DualXDA consists of two approaches, DualDA and XDA. DualDA provides fast and natural sparse data attribution by utilizing the Support Vector Machine theory. XDA leverages the advantages of existing feature attribution methods to explain why training samples are important for predicting test samples. DualDA achieves explanation time up to 4,100,000x faster than existing influence function methods and up to 11,000x faster than existing most efficient approximation methods, while maintaining high attribution quality and performing well on various subsequent tasks. In conclusion, DualXDA is expected to open a new era of responsible AI systems by enabling transparent and efficient analysis of large-scale neural network architectures.
