서브메뉴
검색
Explaining Unintuitive Feature Importance Explanations
Explaining Unintuitive Feature Importance Explanations
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202103209
- ISBN
- 9798291558157
- DDC
- 020
- 저자명
- Qu, Jiaming.
- 서명/저자
- Explaining Unintuitive Feature Importance Explanations
- 발행사항
- [Sl] : The University of North Carolina at Chapel Hill, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 166 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-02, Section: A.
- 주기사항
- Advisor: Wang, Yue.
- 학위논문주기
- Thesis (Ph.D.)--The University of North Carolina at Chapel Hill, 2025.
- 초록/해제
- 요약Explainable AI (XAI) research has proposed a variety of approaches for explaining machine learning model predictions to end-users. Feature importance explanation, which highlights input features that are most influential to the output, is a popular and effective approach. Studies have developed different algorithms to identity important features, and also empirically shown that such explanations can improve end-users' decision-making performance and understanding of the AI system in certain tasks. Despite these promising advancements, a critical gap remains open: features deemed predictive by algorithms can appear unintuitive to humans. For instance, prior studies found that the word ``problems'' predicts positive sentiment in product reviews, and the word ``Chicago'' is strongly associated with deceptive hotel reviews. However, most XAI research on feature importance explanations does not further explain why a feature is predictive. Therefore, this dissertation explores an underexplored area in XAI research: explaining unintuitive feature importance explanations. Using deception detection as a case study, this dissertation investigated two primary research questions.The first research question focused on developing a computational approach to explaining predictive but unintuitive words in deception detection. While previous research has used contextual information to explain unintuitive words in sentiment analysis, those in deception detection often represent some underlying phenomena that are too complex to be captured by local context. To this end, I leveraged a large language model (LLM) to conjecture the phenomena associated with words predictive of a review being genuine or deceptive, and then conducted algorithmic evaluations to validate the technical reliability of this LLM-based approach. The evaluation results confirmed that the LLM used in this dissertation (GPT-4o) generated non-hallucinated and generalizable phenomena that effectively explained why a word is predictive of genuine or deceptive reviews.The second research question focused on evaluating how unintuitive words and LLM-generated explanations influence participants in a decision-making task. To address this research question, I conducted a crowdsourced user study (N=220). In the study, participants first interacted with an AI system and judged a sequence of eight Chicago hotel reviews. Participants were randomly assigned to one of five interface conditions (i.e., a between-subjects design). The interface conditions varied in terms of the AI assistance features made available to participants. Then, participants independently completed a task of detecting deceptive hotel reviews from other cities. The study results found that showing predictive words without explaining why they are predictive was no better than not showing them at all, while explaining why these words are predictive significantly enhanced participants' learning of the task, appropriate reliance on AI assistance, and perceptions of the AI system.In summary, this dissertation makes three key contributions: (1) it sheds light on the critical issue of unintuitive feature importance explanations and the urgent need for making them understandable to end-users; (2) it introduces a conjecture-then-validate pipeline for explaining unintuitive words through LLMs; (3) it provides empirical insights into how unintuitive words and LLM-generated explanations influence end-users during decision-making. This dissertation offers an important design implication for future XAI research: machine-generated explanations should be aligned with human intuition and common sense to facilitate effective human-AI interaction.
- 일반주제명
- Information science
- 키워드
- Explainable AI
- 키워드
- User studies
- 기타저자
- The University of North Carolina at Chapel Hill Information and Library Science
- 기본자료저록
- Dissertations Abstracts International. 87-02A.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017357333
■00520260202103209
■006m o d
■007cr#unu||||||||
■020 ▼a9798291558157
■035 ▼a(MiAaPQ)AAI32000696
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a020
■1001 ▼aQu, Jiaming.
■24510▼aExplaining Unintuitive Feature Importance Explanations
■260 ▼a[Sl]▼bThe University of North Carolina at Chapel Hill▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a166 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-02, Section: A.
■500 ▼aAdvisor: Wang, Yue.
■5021 ▼aThesis (Ph.D.)--The University of North Carolina at Chapel Hill, 2025.
■520 ▼aExplainable AI (XAI) research has proposed a variety of approaches for explaining machine learning model predictions to end-users. Feature importance explanation, which highlights input features that are most influential to the output, is a popular and effective approach. Studies have developed different algorithms to identity important features, and also empirically shown that such explanations can improve end-users' decision-making performance and understanding of the AI system in certain tasks. Despite these promising advancements, a critical gap remains open: features deemed predictive by algorithms can appear unintuitive to humans. For instance, prior studies found that the word ``problems'' predicts positive sentiment in product reviews, and the word ``Chicago'' is strongly associated with deceptive hotel reviews. However, most XAI research on feature importance explanations does not further explain why a feature is predictive. Therefore, this dissertation explores an underexplored area in XAI research: explaining unintuitive feature importance explanations. Using deception detection as a case study, this dissertation investigated two primary research questions.The first research question focused on developing a computational approach to explaining predictive but unintuitive words in deception detection. While previous research has used contextual information to explain unintuitive words in sentiment analysis, those in deception detection often represent some underlying phenomena that are too complex to be captured by local context. To this end, I leveraged a large language model (LLM) to conjecture the phenomena associated with words predictive of a review being genuine or deceptive, and then conducted algorithmic evaluations to validate the technical reliability of this LLM-based approach. The evaluation results confirmed that the LLM used in this dissertation (GPT-4o) generated non-hallucinated and generalizable phenomena that effectively explained why a word is predictive of genuine or deceptive reviews.The second research question focused on evaluating how unintuitive words and LLM-generated explanations influence participants in a decision-making task. To address this research question, I conducted a crowdsourced user study (N=220). In the study, participants first interacted with an AI system and judged a sequence of eight Chicago hotel reviews. Participants were randomly assigned to one of five interface conditions (i.e., a between-subjects design). The interface conditions varied in terms of the AI assistance features made available to participants. Then, participants independently completed a task of detecting deceptive hotel reviews from other cities. The study results found that showing predictive words without explaining why they are predictive was no better than not showing them at all, while explaining why these words are predictive significantly enhanced participants' learning of the task, appropriate reliance on AI assistance, and perceptions of the AI system.In summary, this dissertation makes three key contributions: (1) it sheds light on the critical issue of unintuitive feature importance explanations and the urgent need for making them understandable to end-users; (2) it introduces a conjecture-then-validate pipeline for explaining unintuitive words through LLMs; (3) it provides empirical insights into how unintuitive words and LLM-generated explanations influence end-users during decision-making. This dissertation offers an important design implication for future XAI research: machine-generated explanations should be aligned with human intuition and common sense to facilitate effective human-AI interaction.
■590 ▼aSchool code: 0153.
■650 4▼aInformation science
■653 ▼aExplainable AI
■653 ▼aHuman-AI interaction
■653 ▼aLarge language models
■653 ▼aUnintuitive features
■653 ▼aUser studies
■690 ▼a0723
■71020▼aThe University of North Carolina at Chapel Hill▼bInformation and Library Science.
■7730 ▼tDissertations Abstracts International▼g87-02A.
■790 ▼a0153
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17357333▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


