Answers are grounded in this paper and cite their sources.
Late deliveries are financially damaging for online vendors, and the list of machine learning models proposed to predict them is growing. Yet they are mostly black boxes, and several of them report very high accuracies that are difficult to reproduce. In this paper, I propose a knowledge-guided hybrid model that combines IF-THEN production rules, organised as in an expert system, with an XGBoost classifier. The same rules explain and direct each prediction. I tested the model on the public DataCo Smart Supply Chain dataset without leakage, using only the information available at order time, and with a split by order. The major delay patterns were covered by only five rules, for example a cut-off at noon for same-day orders. The rule-based model alone had an accuracy of 0.719. This is the highest accuracy that can be obtained from pre-dispatch information in this dataset. Random Forest, XGBoost and the hybrid model reached between 0.710 and 0.715, with no significant difference between XGBoost and the hybrid model. In addition, the hybrid model gave a rule-based explanation for almost 97 out of 100 predictions. The study also shows that the gap between the shipping date and the order date reproduces the actual shipping time and makes it possible for a model to reach perfect accuracy, which can explain the very high results of previous work. So, for this task, expert knowledge brought transparency at no cost in accuracy.
No takes yet. Share an insight, caveat, or question.
Aicha Chetali (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: