Socially Responsible Ai: Theories And Practices (by Team-IRA)

Description

In the current era, people and society have grown increasingly reliant on artificial intelligence (AI) technologies. AI has the potential to drive us towards a future in which all of humanity flourishes. It also comes with substantial risks for oppression and calamity. In response, researchers and organizations have been working to publish principles and develop AI regulations for the responsible use of AI in consequential application domains. However, these theoretically formulated principles and regulations also need to be turned into actionable algorithms to materialize AI for good. This book introduces a unified perspective of Socially Responsible AI to help bridge conceptual AI principles to responsible AI practice. It begins with an interdisciplinary definition of socially responsible AI and the AI responsibility pyramid. Existing efforts seeking to materialize the mainstream responsible AI principles are then presented. The book also discusses how to leverage advanced AI techniques to address the challenging societal issues through Protecting, Informing, and Preventing, and concludes with open problems and challenges. This book serves as a convenient entry point for researchers, practitioners, and students to understand the problems and challenges of socially responsible AI, and to identify how their areas of expertise can contribute to making AI socially responsible. Contents Preface About the Authors Acknowledgments 1. Defining Socially Responsible AI 1.1. Why NOW 1.2. What is Socially Responsible AI 1.3. The AI Responsibility Pyramid 1.4. Socially Responsible AI Algorithms 1.5. What Could Go Wrong? 1.5.1. Formalization 1.5.2. Measuring Errors 1.5.3. Biases 1.5.4. Data Misuse 1.5.5. Dependence versus Causality 1.6. Concluding Remarks 1.6.1. Summary 1.6.2. Additional Readings 2. Theories in Socially Responsible AI 2.1. Fairness 2.1.1. Different Fairness Notions 2.1.1.1. Group fairness 2.1.1.2. Individual fairness 2.1.2. Mitigating Unwanted Bias 2.1.2.1. Pre-processing approaches 2.1.2.2. In-processing approaches 2.1.2.3. Post-processing approaches 2.1.3. Discussion 2.2. Interpretability 2.2.1. Different Forms of Explanations 2.2.2. Taxonomy of AI Interpretability 2.2.3. Techniques for AI Interpretability 2.2.3.1. Pre-model interpretability 2.2.3.2. In-model interpretability 2.2.3.3. Post hoc local interpretability 2.2.3.4. Post hoc global interpretability 2.2.4. Discussion 2.3. Privacy 2.3.1. Traditional Privacy Models 2.3.1.1. PPDP via syntactic anonymity 2.3.1.2. PPDM via differential privacy 2.3.2. Privacy for Social Graphs 2.3.3. Graph Anonymization 2.3.3.1. k-anonymity-based graph anonymization 2.3.3.2. Differential privacy-based graph anonymization 2.3.4. Discussion 2.4. Distribution Shift 2.4.1. Different Types of Distribution Shifts 2.4.2. Mitigating Distribution Shift via Domain Adaptation 2.4.2.1. Label shift 2.4.2.2. Covariate shift 2.4.2.3. Concept shift & conditional shift 2.4.3. Mitigating Distribution Shift via Domain Generalization 2.4.3.1. Label shift 2.4.3.2. Covariate shift 2.4.3.3. Concept shift, conditional shift, and other distribution shifts 2.4.4. Discussion 2.5. Concluding Remarks 2.5.1. Summary 2.5.2. Additional Readings 3. Practices of Socially Responsible AI 3.1. Protecting 3.1.1. A Multi-Modal Approach for Cyberbullying Detection 3.1.1.1. Challenges 3.1.1.2. The approach: XBully 3.1.2. A Deep Learning Approach for Social Bot Detection 3.1.2.1. Account-level detection 3.1.2.2. Tweet-level detection 3.1.3. A Privacy-Preserving Graph Convolutional Network with Partially Observed Sensitive Attributes 3.1.3.1. Problem definition 3.1.3.2. The approach: DP-GCN 3.2. Informing 3.2.1. An Approach for Explainable Fake News Detection 3.2.1.1. Problem definition 3.2.1.2. The approach: dEFEND 3.2.2. Causal Understanding of Fake News Dissemination on Social Media 3.2.2.1. Problem definition 3.2.2.2. A causal approach 3.3. Preventing 3.3.1. Mitigating Gender Bias in Word Embeddings 3.3.1.1. The problem 3.3.1.2. The approach: Hard debiasing 3.3.2. Debiasing Cyberbullying Detection 3.3.2.1. Problem definition 3.3.2.2. A non-compromising approach 3.4. Concluding Remarks 3.4.1. Summary 3.4.2. Additional Readings 4. Challenges of Socially Responsible AI 4.1. Causality and Socially Responsible AI 4.1.1. Causal Inference 101 4.1.2. Causality-based Fairness Notions and Bias Mitigation 4.1.2.1. Causal fairness 4.1.2.2. Causality for bias mitigation 4.1.2.3. Counterfactual data augmentation 4.1.3. Causality and Interpretability 4.1.3.1. Model-based causal interpretability 4.1.3.2. Counterfactual explanations 4.1.3.3. Partial dependence plots 4.2. How Context Can Help 4.2.1. A Sequential Bias Mitigation Approach 4.2.1.1. The approach 4.2.2. A Multidisciplinary Approach for Context-Specific Interpretability 4.3. The Trade-offs: Can’t We have Them All? 4.3.1. The Fairness–Utility Trade-off 4.3.1.1. 4-way balanced dataset 4.3.1.2. Group-balanced dataset 4.3.1.3. Outcome-balanced dataset 4.3.1.4. Imbalanced dataset 4.3.2. The Interpretability–Utility Trade-off 4.3.3. The Privacy–Utility Trade-off 4.3.4. Trade-offs among Fairness, Interpretability, and Privacy 4.4. Concluding Remarks 4.4.1. Summary 4.4.2. Additional Readings Bibliography Index