SemantiClean framework brings full auditability to e-commerce intent inference

SemantiClean, introduced by Liu Hung Ming, is a modular framework designed to extract structured semantic signals from e-commerce session data. Unlike conventional end-to-end predictors that optimize solely for accuracy, SemantiClean prioritizes auditability, structural governance, and reproducibility. It uses a four-layer architecture (Functional, Interaction, Systemic, Contextual) to organize 24 behavioral elements from the Online Shoppers Purchasing Intention (OSPI) dataset. Three anti-inflation mechanisms—RedundancyGroup contribution caps, TieredPenaltyCalculator bias penalties, and AdaptiveConstraintMode cold-start—enforce signal quality. This design explicitly trades marginal predictive gains for element-level transparency and defensible decision trails.

The framework includes the LLM-Integrated Semantic Inference Engine, a fully implemented two-phase architecture that leverages complete element metadata at inference time. All quantitative results in the paper are produced by this engine. Deterministic engine outputs are fully reproducible (sigma=0), while LLM-dependent results (E8, E10) vary under controlled conditions (fixed provider/model/temperature). Notably, the gender inference target remains non-functional and is excluded from all results. Compared to black-box approaches, SemantiClean offers a transparent, governable alternative for tasks like purchase intent prediction, customer segmentation, and product affinity analysis.