{"title":"DDImage:基于图像还原的黑盒分类器自动解释方法","authors":"Mingyue Jiang, Chengjian Tang, Xiao-Yi Zhang, Yangyang Zhao, Zuohua Ding","doi":"10.1007/s10664-024-10505-0","DOIUrl":null,"url":null,"abstract":"<p>Due to the prevalent application of machine learning (ML) techniques and the intrinsic black-box nature of ML models, the need for good explanations that are sufficient and necessary towards locally interpreting a model’s prediction has been well recognized and emphasized. Existing explanation approaches, however, favor either the sufficiency or necessity. To fill this gap, in this paper, we propose an approach for generating local explanations that are both sufficient and necessary. Our approach, DDImage, automatically produces local explanations for ML-based image classifiers in a post-hoc way. The core idea behind DDImage is to discover an appropriate explanation by debugging the given input image via a series of image reductions, with respect to the sufficiency and necessity properties. Evaluation of DDImage using publicly available datasets and popular classification models reveals its effectiveness and efficiency. Compared with three state-of-the-art approaches, DDImage demonstrates a superior performance in producing small-sized explanations preserving both sufficiency and necessity, and it also shows promising stability and efficiency. We also identify the impact of segmentation granularity, reveal the performance variance for different target models, and further show that our approach is applicable across different problem domains.</p>","PeriodicalId":11525,"journal":{"name":"Empirical Software Engineering","volume":"214 1","pages":""},"PeriodicalIF":3.5000,"publicationDate":"2024-07-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"DDImage: an image reduction based approach for automatically explaining black-box classifiers\",\"authors\":\"Mingyue Jiang, Chengjian Tang, Xiao-Yi Zhang, Yangyang Zhao, Zuohua Ding\",\"doi\":\"10.1007/s10664-024-10505-0\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<p>Due to the prevalent application of machine learning (ML) techniques and the intrinsic black-box nature of ML models, the need for good explanations that are sufficient and necessary towards locally interpreting a model’s prediction has been well recognized and emphasized. Existing explanation approaches, however, favor either the sufficiency or necessity. To fill this gap, in this paper, we propose an approach for generating local explanations that are both sufficient and necessary. Our approach, DDImage, automatically produces local explanations for ML-based image classifiers in a post-hoc way. The core idea behind DDImage is to discover an appropriate explanation by debugging the given input image via a series of image reductions, with respect to the sufficiency and necessity properties. Evaluation of DDImage using publicly available datasets and popular classification models reveals its effectiveness and efficiency. Compared with three state-of-the-art approaches, DDImage demonstrates a superior performance in producing small-sized explanations preserving both sufficiency and necessity, and it also shows promising stability and efficiency. We also identify the impact of segmentation granularity, reveal the performance variance for different target models, and further show that our approach is applicable across different problem domains.</p>\",\"PeriodicalId\":11525,\"journal\":{\"name\":\"Empirical Software Engineering\",\"volume\":\"214 1\",\"pages\":\"\"},\"PeriodicalIF\":3.5000,\"publicationDate\":\"2024-07-30\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Empirical Software Engineering\",\"FirstCategoryId\":\"94\",\"ListUrlMain\":\"https://doi.org/10.1007/s10664-024-10505-0\",\"RegionNum\":2,\"RegionCategory\":\"计算机科学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q1\",\"JCRName\":\"COMPUTER SCIENCE, SOFTWARE ENGINEERING\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Empirical Software Engineering","FirstCategoryId":"94","ListUrlMain":"https://doi.org/10.1007/s10664-024-10505-0","RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"COMPUTER SCIENCE, SOFTWARE ENGINEERING","Score":null,"Total":0}
引用次数: 0
摘要
由于机器学习(ML)技术的普遍应用和 ML 模型固有的黑箱性质,人们充分认识到并强调了对模型预测进行充分和必要的良好解释的必要性。然而,现有的解释方法倾向于充分性或必要性。为了填补这一空白,我们在本文中提出了一种生成既充分又必要的本地解释的方法。我们的方法,即 DDImage,能以事后的方式为基于 ML 的图像分类器自动生成局部解释。DDImage 背后的核心理念是通过一系列图像还原来调试给定的输入图像,从而根据充分性和必要性属性发现适当的解释。使用公开可用的数据集和流行的分类模型对 DDImage 进行的评估显示了其有效性和效率。与三种最先进的方法相比,DDImage 在生成同时保留充分性和必要性的小尺寸解释方面表现出色,而且还显示出良好的稳定性和效率。我们还确定了细分粒度的影响,揭示了不同目标模型的性能差异,并进一步表明我们的方法适用于不同的问题领域。
DDImage: an image reduction based approach for automatically explaining black-box classifiers
Due to the prevalent application of machine learning (ML) techniques and the intrinsic black-box nature of ML models, the need for good explanations that are sufficient and necessary towards locally interpreting a model’s prediction has been well recognized and emphasized. Existing explanation approaches, however, favor either the sufficiency or necessity. To fill this gap, in this paper, we propose an approach for generating local explanations that are both sufficient and necessary. Our approach, DDImage, automatically produces local explanations for ML-based image classifiers in a post-hoc way. The core idea behind DDImage is to discover an appropriate explanation by debugging the given input image via a series of image reductions, with respect to the sufficiency and necessity properties. Evaluation of DDImage using publicly available datasets and popular classification models reveals its effectiveness and efficiency. Compared with three state-of-the-art approaches, DDImage demonstrates a superior performance in producing small-sized explanations preserving both sufficiency and necessity, and it also shows promising stability and efficiency. We also identify the impact of segmentation granularity, reveal the performance variance for different target models, and further show that our approach is applicable across different problem domains.
期刊介绍:
Empirical Software Engineering provides a forum for applied software engineering research with a strong empirical component, and a venue for publishing empirical results relevant to both researchers and practitioners. Empirical studies presented here usually involve the collection and analysis of data and experience that can be used to characterize, evaluate and reveal relationships between software development deliverables, practices, and technologies. Over time, it is expected that such empirical results will form a body of knowledge leading to widely accepted and well-formed theories.
The journal also offers industrial experience reports detailing the application of software technologies - processes, methods, or tools - and their effectiveness in industrial settings.
Empirical Software Engineering promotes the publication of industry-relevant research, to address the significant gap between research and practice.