Shajalal, Md
Preferred name
Shajalal, Md
ORCID
Now showing 1 - 1 of 1
- Some of the metrics are blocked by yourconsent settings
Publication Access status: Open Access , Towards Human-Centered Actionable Explainable AI-enabled Systems(2025)Recently, the applications of complex artificial Intelligence (AI) models have increased exponentially in almost every sector due to the enormous advancement of computing power and the availability of high-quality annotated data for training complex machine learning (ML) models. Generally, AI models are very complex in structure, and they often need to learn thousands, even millions, of parameters in the training phase. Though the predictions are accurate, due to the complex decision-making process, the predictions from such AI models are not understandable to users. Hence, AI systems lack explainability, transparency, and trustworthiness in making the decision explainable to the users. AI models with such complexity are often referred to as black-box models. To interpret the black-box AI models, recently, there has been a high interest in the AI research community concentrating on extracting facts and rationale to explain the reasons behind the prediction and overall models' decision-making priorities. The field that practices interpreting complex AI models and explaining the predictions to uncover the reasons behind particular predictions is known as eXplainable Artificial Intelligence (XAI). Improvements in interpreting ML models have been evident in this decade. However, the current explainability techniques are helpful for AI practitioners in the way that they can employ explanations to debug and eventually improve the models' performance. However, the primary objective of XAI is to help laypeople understand the predictions by providing human-centric explanations, which will eventually increase transparency and trust and lead to faster AI adoption in real-world applications. A significant gap exists in achieving human-centered explainability for AI systems due to associated challenges, including user experience variability, context sensitivity, bias and data deficiency, and actionability. This dissertation aims to advocate the human-understandable explainability of AI-enabled systems and introduces explainable models in different real-world application scenarios. We strive to answer research questions, including i) How can we achieve high-performance ML models addressing technical challenges, including data imbalance, data inadequacy, and model bias? ii) How do explanations vary across different application contexts? iii) What underlying facts and rationale should be considered when explaining prediction for a given context? and iv) How could we achieve actionable explanations? We adopted an exploratory, experimental approach to answering these questions by conducting a wide range of experiments, introducing explainability techniques, and demonstrating explanations in three application areas: smart home, business, and natural language processing (NLP). After carefully selecting application scenarios considering the mentioned questions, this thesis proposed multiple high-performance ML models for energy demand forecasting, occupants' thermal comfort preference modeling, product backorder prediction, multi-class patent classification, and fake review identification. Then, it introduced explainability to provide comprehensible, actionable explanations so that users and stakeholders could understand the predictions and take necessary action accordingly. The results from a wide range of experiments demonstrated high performance compared to state-of-the-art methods. They provided explanations that capture relevant facts and rationale to make users understand the proposed ML models' predictions and overall priorities. The technical and empirical evaluation of the generated explanations for explainable AI-enabled systems highlighted what information needs to be considered and how they should be represented in explanations. The broad contribution of this thesis is three-fold: i) We achieved high-performance ML models in different application areas addressing the challenges, including data inadequacy and extreme imbalance; ii) With a wide range of experiments, this thesis gives a holistic conclusion on what facts and rationale should be employed in generating explanations for a given application context; iii) Lastly, this dissertation highlighted how we can achieve actionable explanations so that users can take necessary actions to earn more system efficiency in the given application context.Source Type:Doctoral Thesis11 10
