Analysis of Public Tenders with Rule-Based Systems to Improve Decision-Making

We evaluate supervised classification models on European Union public tenders, pairing random forest and gradient boosting with explainable artificial intelligence (XAI) and a rule-based system. Using TED data processed with Apache Spark, we modeled the main activity of the contracting authority among 21 classes (Experiment A) and whether an SME wins the contract (Experiment B). Random forest outperforms gradient boosting and a decision tree baseline in every fold of both experiments (one-sided Wilcoxon test, p = 0.001). Its F1-macro in Experiment A is 0.3550, far above the majority-class baseline (0.0228), yet only three of the 21 classes exceed F1 = 0.6, and the score falls to 0.2203 on later years and 0.1443 on countries absent from training. Experiment B is easier: F1-macro reaches 0.5863 (AUC = 0.7100) against a baseline of 0.4500 and 0.5515 when the model sees only the features that are known before bidding. That restricted model, however, has no discriminative ability on unseen countries (AUC = 0.51). National markets explain much of this. In the seven largest, F1-macro averages 0.59 when the model is trained within the country and 0.50 when the country is left out. The if–then rules derived from LIME have low fidelity (R2 = 0.43 and 0.16), and, on held-out data, a third of the Experiment B rules are no more precise than an empty rule; TreeSHAP supports the global feature ranking but not these local explanations. Any use for SME decision support would be limited to markets that are represented in training and would first need validation with end users.