KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
XAI (Explainable AI)
Set of techniques and methods aimed at making the decisions of artificial intelligence systems understandable by humans, essential for trust and acceptability.
Interpretability
Ability of a model to present its decisions in a way understandable to humans, distinguished from transparency which concerns the understanding of the internal mechanism.
Post-hoc explanations
Explanation methods applied after model training without modifying its architecture, allowing to explain the predictions of black box models.
SHAP (SHapley Additive exPlanations)
Theoretical approach based on game theory to assign the importance of each feature in a model's prediction in an additive and coherent manner.
LIME (Local Interpretable Model-agnostic Explanations)
Local explanation technique that approximates the behavior of a complex model by a simple and interpretable model in the neighborhood of a specific prediction.
Influence factors
Specific elements (items, attributes, past behaviors) that have directly contributed to the generation of a particular recommendation in a system.
Counter-explanations
Explanations that justify why certain items were not recommended, helping users understand the limitations and exclusion criteria of the system.
Causal justifications
Explanations based on cause-effect relationships between user actions and generated recommendations, rather than simple correlations.
Knowledge-based approaches
Recommendation methods using ontologies or knowledge graphs to generate semantically rich and contextually relevant explanations.
Explanation visualizations
Interactive graphical representations that transform algorithmic justifications into intuitive visual elements to facilitate user understanding.
Explanation personalization
Adaptation of the content, style, and level of detail of explanations according to each user's profile, preferences, and expertise.
Qualitative evaluations of explanations
Evaluation methods based on user studies, interviews, and content analysis to measure the relevance and perceived usefulness of explanations.
Explanatory feedback
Mechanism allowing users to react to the provided explanations, thereby refining future recommendations and the quality of justifications.
Explanatory complexity
Measure of the cognitive difficulty required to understand an explanation, evaluating the trade-off between technical accuracy and user accessibility.
Algorithmic transparency
Principle of revealing the underlying mechanisms, data, and logic of a recommendation system to ensure its traceability and auditability.
Algorithmic trust
Level of credibility and reliability perceived by users towards a system, directly influenced by the quality and relevance of the provided explanations.
Intrinsic explanations
Models designed from their conception to be interpretable, natively integrating explanation capabilities unlike post-hoc approaches.
Explanatory association rules
Sets of logical rules (IF-THEN) that justify recommendations by showing the discovered relationships between behaviors and items.
Explanatory bias
Systematic distortions in generated explanations that may over-represent certain factors or minimize others, affecting the fair perception of the system.