Information

Identifying employees who may need additional professional-learning support could help organisations allocate training resources, but model comparisons in this area often omit trivial baselines and class-specific metrics. We compare seven classifiers (decision tree, random forest, logistic regression, k-nearest neighbours, XGBoost, LightGBM, CatBoost) with a rule-based and a majority-class baseline under one protocol: stratified hold-out, the same tuning procedure, paired bootstrap confidence intervals, a descriptive repeated cross-validation check, and SHAP and LIME explanations. Experiment 1 predicts a KPI label of undocumented origin for 17,415 employees (35.9% positive); it is an illustrative comparison using the label as distributed, not validated evidence about KPI attainment. Experiment 2 predicts promotion for 54,808 employees (8.5% positive), a related task from the same public source and not an external validation. In Experiment 1, learning models reached macro-F1 of 0.628 to 0.663, against 0.590 for the rule-based baseline; a depth-5 decision tree was 0.005 below the reference random forest (95% bootstrap CI −0.016 to +0.005) but was not shown non-inferior at a 0.01 margin. In Experiment 2, boosting models were best (macro-F1 0.734 to 0.739), the decision tree was clearly worse (−0.030), and only about 35% of promoted employees were identified. Associations with training indicators are not causal, and no effect on decisions was tested.

In content-based recommendation, a second-stage (L2) reranker decides which retrieved items fill the first screen. Decision models, which answer a typed question about a text with a probability, are an alternative to trained cross-encoders, but whether they separate exact matches from close substitutes is unknown. We cast a decision model as a pointwise reranker that asks whether each candidate exactly matches the request. Under a pre-registered protocol, a gain must hold on both retrieved and fully labelled lists. On Amazon ESCI (1000 requests) and Wayfair WANDS (299), the commercial Jev model, used zero-shot, improves P@10 over embedding order by 0.046 to 0.054 and beats the 0.6B Qwen3-Reranker by 0.02 to 0.03 on ESCI, but not significantly on WANDS. Larger trained cross-encoders (Qwen3-Reranker-4B and 8B, bge-reranker-v2-m3) erase most of this lead: Jev is ahead only on fully labelled ESCI lists. A prompted Qwen3-8B stays behind Jev everywhere. The open decision model CLM-v0.1-8B ranks at random level when scoring a whole list but matches embedding order one product at a time. Jev costs about as much as the 4B model and 13 times bge-reranker-v2-m3. A decision model can serve as an L2 reranker but not demonstrably better than trained cross-encoders.

Localizing the source of degradation in a communication network is difficult when symptoms come from sources of unequal reliability, are partly missing and are mixed with false alarms. We propose a reasoning model in which every piece of evidence is specified as a Z-number of strength and reliability. Observed symptoms support the candidates that can explain them through the topology and through temporal Z-rules evaluated automatically from time stamps; unexplained alarms and expected but unobserved symptoms act as negative evidence; all evidence is discounted by the reliability of its source. In 47 simulated configurations, the model identified the true source in 85.9% of incidents in the default setting, against 82.9% for Dempster–Shafer fusion with the same negative evidence and 73.4% for codebook correlation, and it was the most accurate method in most stress tests, but not with correlated false-alarm bursts or upstream symptoms. On an independent generator of telecom alarm storms on three real backbones, it was within four points of the evidential baseline and about 10–12 points more accurate than the codebook. A crisp version was equally accurate; we show why the benefit lies in the structure and linguistic specification of the evidence rather than in fuzzy arithmetic.