[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-119446-en":3,"doc-seo-119446-105":30,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":4,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":27,"seo_description":14,"update_tm":28,"read_time":29},119446,1374391974468,"Eden","https://ap-avatar.wpscdn.com/davatar_29158cc5080c5b710cf443261637dec0",8,"Research & Report","Towards Robust and Actionable Explanations in Machine Learning Systems","Organizations increasingly rely on machine learning systems while collecting massive real-world data, yet the resulting models can be hard for humans to interpret and may behave unexpectedly. This dissertation advances explainability research by emphasizing two key properties of explanations: robustness and actionability, evaluated across different users and use cases. Four investigations study influence functions for identifying training contributions to biases, reformulate their objectives for robustness to atypical data, analyze explanation consistency under model indeterminacy, and extend explanations to include human directives for large language models using a markup-like syntax.","Towards Robust and Actionable Explanations in Machine Learning  \nSystems  \nby  \nMarc-Etienne Brunet  \nA thesis submitted in conformity with the requirements for the degree of Doctor of Philosophy  \nDepartment of Computer Science  \nUniversity of Toronto  \n© Copyright 2025 by Marc-Etienne Brunet  \nTowards Robust and Actionable Explanations in Machine Learning Systems  \nMarc-Etienne Brunet  \nDoctor of Philosophy  \nDepartment of Computer Science  \nUniversity of Toronto  \n2025  \nAbstract  \nIn the last few decades, organizations have collected vast quantities of data about the world. The capabilities of machine learning systems have also increased dramatically. These two intertwined developments have opened the door for automation on an unprecedented scale, but many critics have expressed concern. Machine learning has the potential to improve the efficiency of many existing processes, but it also poses a substantial technological risk. The nature of most modern machine learning algorithms makes it very difficult for a human to understand what has been learned by a model. As such, any software built on these models has the potential to behave in unexpected ways. The field of explainability research has grown in response to the opaque nature of modern machine learning algorithms. Explanations, whether produced post-hoc or inherent to a model’s design, aim to bridge the gap between what a model has learned and what human users can understand. In this dissertation, we stress the importance of two properties of explanations: robustness and actionability, whose characteristics vary depending on the target user and use case. We present four investigations which contribute to the body of explainability research. In the first, we show how the method of influence functions can be used to identify documents which have contributed to the biases learned by a language model. In the second, we present a reformulation of the influence function objective. We show how this reformulation enables the method to produce explanations which are more robust to atypical data points. In the third, we explore the effects of model indeterminacy (underspecification and the Rashomon effect) on the consistency of explanations. In the fourth, we expand on the traditional definition of an explanation to include directives from a human to a large language model. We propose a Markup-like syntax to aid in the design of in-context learning prompt templates. Through theoretical analysis and experimentation, we explore the benefits and limits of the methods we propose. We discuss the implications of our findings, both in the context of each individual investigation, and in relation to the robustness and actionability of explanations.  \nThis thesis is dedicated to Nora.  \nAcknowledgements  \nAs I’m sure is the case for many doctoral candidates, my experience as an undergrad was egoinflating. As I approached graduation, I remember feeling a confidence in my abilities that bordered on cockiness, intoxicated by a naive interpretation of my own success. Now, as I move towards the completion of my PhD, the overwhelming feeling is one of humility. In the context of the COVID-19 pandemic and new fatherhood, the doctoral process has been quite challenging for me. Along the way, I have been forced to confront my personal limits, pushing through some—accepting others. Iam proud of the work in this thesis, and it would not have been possible without enormous support.  \nFirst and foremost, I want to thank my supervisors, Rich Zemel and Ashton Anderson. I will forever be indebted to their intellectual generosity. They offered me a truly privileged graduate school experience. I felt cared for. Not just through funding and access to snazzy computing resources, but because of the ample time that they carved out for our conversations. Their joint support as I worked through some mental health struggles was also profound; I would not have made it this far ingrad school without their empathy ","cbCaijGK9sJqXkcR","https://ap.wps.com/l/cbCaijGK9sJqXkcR","pdf",8019039,1,142,"English","en",105,"# Abstract\n# Introduction and Motivation\n## Explainability: robustness and actionability\n# Four Investigations\n## Influence functions for bias attribution\n## Robust objective reformulation for atypical points\n## Consistency under model indeterminacy\n## Human-directed explanations and markup-like prompt syntax\n# Theoretical and Experimental Analysis\n## Benefits and limits of proposed methods\n# Implications and Conclusions","[{\"question\":\"Why is explainability necessary for modern machine learning systems?\",\"answer\":\"Most modern algorithms are difficult to interpret, so humans may not understand what a model has learned. This can cause unexpected behavior in software built on these models.\"},{\"question\":\"What two properties does the dissertation emphasize for explanations?\",\"answer\":\"The dissertation stresses robustness and actionability. Their characteristics vary depending on the target user and the specific use case.\"},{\"question\":\"How do the investigations extend influence functions for better explanations?\",\"answer\":\"The thesis shows how influence functions can identify documents contributing to biases learned by a language model, then introduces an objective reformulation to make resulting explanations more robust to atypical data points.\"}]","Towards Robust and Actionable Explanations in Machine Learning Systems | PDF",1785724324,358,{"code":4,"msg":31,"data":32},"ok",{"site_id":24,"language":23,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":28},"towards-robust-and-actionable-explanations-in-machine-learning-systems","",{"@graph":36,"@context":85},[37,54,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,51],{"item":41,"name":42,"@type":43,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":47},"https://docshare.wps.com/document/","Document",2,{"item":49,"name":12,"@type":43,"position":50},"https://docshare.wps.com/document/research-report/",3,{"item":52,"name":13,"@type":43,"position":53},"https://docshare.wps.com/document/towards-robust-and-actionable-explanations-in-machine-learning-systems/119446/",4,{"url":52,"name":13,"@type":55,"author":56,"headline":13,"publisher":58,"fileFormat":61,"inLanguage":23,"description":14,"dateModified":62,"datePublished":62,"encodingFormat":61,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":57},"Person",{"url":41,"name":59,"@type":60},"DocShare","Organization","application/pdf","2026-08-03",true,{"@type":65,"interactionType":66,"userInteractionCount":4},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"Why is explainability necessary for modern machine learning systems?","Question",{"text":75,"@type":76},"Most modern algorithms are difficult to interpret, so humans may not understand what a model has learned. This can cause unexpected behavior in software built on these models.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"What two properties does the dissertation emphasize for explanations?",{"text":80,"@type":76},"The dissertation stresses robustness and actionability. Their characteristics vary depending on the target user and the specific use case.",{"name":82,"@type":73,"acceptedAnswer":83},"How do the investigations extend influence functions for better explanations?",{"text":84,"@type":76},"The thesis shows how influence functions can identify documents contributing to biases learned by a language model, then introduces an objective reformulation to make resulting explanations more robust to atypical data points.","https://schema.org",{"og:url":52,"og:type":87,"og:title":13,"og:site_name":59,"og:description":14},"article",{"robots":89,"canonical":52},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":20,"doc_module":4,"doc_module_name":46,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":53,"doc_module":4,"doc_module_name":46,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":46,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":46,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":46,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":46,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":46,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":46,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]