Towards Explainable Clustering: A Constrained Declarative based Approach

Clustering results need human validation, making explainability essential: The paper emphasizes that unsupervised clustering tasks require domain experts to validate results, making it critical that clustering methods can explain why objects are grouped together—not just that they are grouped.

Good explanations require both coverage and discrimination: The research proposes that cluster explanations should describe what characteristics define each cluster (coverage) while also showing what makes it distinct from other clusters (discrimination), rather than focusing on only one dimension.

Expert knowledge can be built directly into clustering algorithms: The ECS method allows domain experts to incorporate their knowledge through constraints at different levels—either about the expected clustering structure or about what constitutes a valid explanation—improving both quality and trustworthiness of results.

Explainability and clustering quality should be optimized together: Rather than treating explainability as an afterthought, the paper argues these two goals must be considered simultaneously when building clustering methods to ensure results are both accurate and understandable.

Patterns with exceptions are more realistic than rigid rules: The framework allows explanations to have parameterized thresholds and exceptions rather than requiring absolute rules, making cluster descriptions more flexible and applicable to real-world data where perfect patterns rarely exist.