A Gain-Budgeted Business-Constrained Forest for Directional Consistency and Reference-Threshold Alignment in Communication QoE Assessment

Communication quality-of-experience (QoE) assessment may require accurate predictions together with a record of how business priors influence the model. We propose a gain-budgeted business-constrained forest (GBBCF). At each node, the method limits relative split-gain loss, filters candidates according to the direction of weighted child means, and uses lexicographic selection to favor proximity to a reference point. If the feasible set is empty, the node becomes a leaf. The guarantee applies to local split gain for fixed node samples and candidate sets; it implies neither global monotonicity nor a bound on prediction error. In repeated nested cross-validation on seven project-supplied tasks, macro-average RMSE was 4.711 for GBBCF, 4.675 for random forest, and 4.371 for monotonic XGBoost. GBBCF and random forest had similar RMSE point estimates, but equivalence was not tested. Monotonic XGBoost had lower RMSE on all seven tasks (Holm-adjusted p = 0.0469). In controlled mechanism comparisons, alignment reduced node-unweighted and coverage-weighted reference distances by 6.7% and 10.8%. Grouped validation on three public subjective datasets assessed predictive performance and directional behavior separately within each dataset. On LIVE-Viasat, source-clustered analyses at overlapping low-QoE boundaries did not establish a screening advantage over every comparator. An optional gate uses only training data to assess predictive support for reference locations. Its high activation rate for permuted synthetic references limits its ability to discriminate among threshold locations. The method is intended for offline auditing of node decisions; the validity of engineering thresholds requires independent evidence.