Towards Explainable Artificial Intelligence (XAI): A Data Mining Perspective

Understanding AI decisions requires examining the data, not just the algorithms: The paper argues that explainability isn't solely about making neural networks more transparent internally—it's equally important to examine what data these systems learned from, how that data was collected and processed, and how specific data points influence final decisions.

Training data quality directly shapes AI behavior and trustworthiness: The research highlights that "nuances" in training data, including data valuation and anomalies, significantly impact how AI systems make decisions. This suggests that responsible AI requires careful scrutiny of training datasets before deployment, not just after problems emerge.

XAI serves three distinct purposes for responsible development: The paper identifies three key goals—interpreting model outputs through feature attribution, understanding training data influences, and discovering domain knowledge patterns. Organizations implementing responsible AI should consider which of these purposes matters most for their specific applications and stakeholders.

Explainability methods work across different data types: Rather than being limited to one domain, XAI techniques can be applied to images, text, tabular data, and other formats. This means responsible AI practices can be systematized across diverse applications, from healthcare to finance to content moderation.

Data mining techniques are central to making AI explainable: The paper reframes XAI as fundamentally a data mining challenge, suggesting that traditional data analysis methods—examining patterns, anomalies, and relationships in datasets—are essential tools for building trustworthy AI systems.