Feature-based explainable AI
The increasing complexity of modern machine learning models poses significant challenges for their transparency and accountability. Explainable artificial intelligence (XAI) has emerged as an important field that seeks to make black-box models understandable to human users. Among its various approaches, feature-based explanation methods play a central role, as they relate a model's prediction back to its inputs. Feature-based explanation methods such as attribution techniques, most prominently Shapley values, quantify the contribution of individual features on the model output and are grounded in strong theoretical foundations. Yet, with the advent of advanced deep learning models such as multi-modal architectures, explanations limited to single features often fall short of capturing the complex mechanisms driving predictions. Many predictive patterns arise from interactions among features. Shapley interactions generalize the classical attribution framework to higher-orders offering a flexible trade-off between faithfulness and complexity of the explanations. However, the use of Shapley interactions is limited by severe computational challenges, rendering them impractical without new algorithmic solutions. At the same time, most existing XAI techniques implicitly assume static models trained once on fixed data. In many real-world scenarios, however, data arrives sequentially in the form of continuous data streams and models are trained incrementally. In such dynamic settings, traditional explanations quickly become outdated and misaligned with current model behavior. Explanation methods must therefore update alongside incremental models in data stream environments. This thesis advances feature-based explainability along both dimensions. The first set of contributions moves from attribution to interaction. It introduces SHAP-IQ, a model-agnostic method for approximating any-order Shapley interactions. It further develops model-specific algorithms, namely TreeSHAP-IQ for tree-based models and GraphSHAP-IQ for graph neural networks, which compute exact Shapley interactions by leveraging structural properties of the underlying models. These contributions are consolidated in the unified software library shapiq, providing a common platform for interaction-based explainability. The second set of contributions shifts from static to dynamic explanations. It presents iSAGE and iPDP, which incrementally compute and maintain global feature importance scores and feature effect estimates for models operating on evolving data streams. Together, this thesis' contributions provide both theoretical advances and practical tools for generating faithful, flexible, and efficient feature-based explanations. By extending the scope of XAI from attribution to interaction, and from static to dynamic, this dissertation pushes the frontier of explainability for complex and continuously evolving machine learning models.