BioCIE

BioCIE provides post-hoc explanations for black-box machine learning models applied to biomedical text classification by extracting confident itemsets that relate biomedical concepts to class labels.

Key Features:

• Confident Itemset Mining Methodology: Employs confident itemset mining and leverages domain knowledge sources to discretize the black-box decision space into smaller subspaces for analysis.
• Semantic Relationship Extraction: Extracts semantic relationships between biomedical concepts and class labels from identified confident itemsets to approximate model behavior for individual predictions.
• Optimization of Explanation Metrics: Optimizes fidelity, interpretability, and coverage to produce concise class-wise and instance-wise explanations.
• Performance Evaluation: Shows empirical improvements over perturbation-based and decision set methods, increasing instance-wise and class-wise fidelity by 11.6% and 7.5% respectively, and improving interpretability by 8%.

Scientific Applications:

• Disease diagnosis: Supports interpretation of biomedical text classifiers used in disease diagnosis by revealing concept–label associations driving predictions.
• Drug discovery: Aids analysis of text-based models in drug discovery by exposing semantic links between biomedical entities and labels.
• Personalized medicine: Facilitates validation of classifiers in personalized medicine by generating interpretable, class-wise explanations.

Methodology:

Discretizes the decision space via confident itemset mining, leverages domain knowledge sources to extract semantic relationships between biomedical concepts and class labels, and generates explanations that approximate black-box behavior while optimizing fidelity, interpretability, and coverage.