A Reproducible Statistical-Learning Framework for Sparse Time-Indexed Transaction Data

Source abstract

Sparse time-indexed transaction data create a statistical-learning problem in which low basket occupancy, combinatorial candidate growth, temporal heterogeneity, and data-quality sensitivity complicate interpretable pattern discovery. Using a multi-year single-case dataset from an independently operated SME restaurant point-of-sale (POS) system, this study proposes a reproducible, sparsity-aware framework that separates global distributional characterization, unsupervised co-occurrence learning, statistical qualification, adjusted outcome modeling, and transparent prioritization. The empirical application comprises 31,157 baskets, 139,396 valid item lines, 710 products, and 762 active days; the 31,157 × 710 basket-product representation has 0.630% density. Multiplicity-screened association-rule learning retains 1174 directional rules. In a strictly subsequent validation period, 93.6% of discovery rules preserve lift above one, 49.3% satisfy all prespecified thresholds again, and rank correlations for support, confidence, and lift range from 0.803 to 0.849. An estimated-dispersion NB2 model with day-clustered uncertainty preserves the principal basket-breadth associations, while Gamma sensitivity analysis confirms the direction, but not invariant magnitude, of late-service basket-value effects. The methodological contribution is not a new stand-alone algorithm, but an auditable evidence architecture that explicitly separates candidate discovery, multiplicity screening, effect assessment, temporal replication, adjusted modeling, and human-supervised prioritization. Because the empirical evaluation is based on a single time-indexed case, the reported effect magnitudes remain case-specific.