Semantic Similarity-Based Aggregator (SIMBA): A ZeroPreprocessing Framework for Clustering Mixed-Type Tabular Data via Cell-Level Sentence Embeddings

SIMBA (Semantic Similarity-Based Aggregator) is a zero-preprocessing framework for clustering mixed-type tabular data. Instead of one-hot encoding or label encoding, SIMBA embeds each table cell individually as column_name: value through a pretrained sentence encoder (all-MiniLM-L6-v2), so that categorical, ordinal, and numeric values are represented in a semantic space where conceptually related values are geometrically close. A two-level self-supervised weighting scheme — global column importance and local per-row unusualness — sharpens the resulting representation before clustering with HDBSCAN, which discovers the number of clusters automatically. Across IBM HR Employee Attrition, UCI Adult Income, and UCI Car Evaluation, SIMBA produces clusters that are both statistically stronger (Silhouette gains up to 60× over K-Modes) and semantically interpretable, validated through a five-test reliability suite including bootstrap stability, label purity, decision-tree interpretability, and a semantic coherence check computed directly in the encoder's native embedding space.