Toward Classroom Evaluation Indicator-System Construction for Generative AI-Enabled Teaching: High-Dimensional Indicator Selection via Generative Rebalancing and Asymmetric Heuristic Optimization

Generative AI-enabled teaching increasingly requires evaluation mechanisms that can process large candidate indicator sets without being dominated by uneven category frequencies or redundant variables. Existing filter methods are computationally inexpensive but mainly measure marginal relevance, whereas wrapper optimizers can capture feature interactions at the cost of a rapidly expanding combinatorial search space. Standard PSO can additionally be sensitive to random initialization and fixed or weakly adaptive search controls, which may reduce diversity or promote premature convergence. To address this computational trade-off, this study combines a conditional generative adversarial network (CGAN) for class-conditioned rebalancing, a two-stage Pearson correlation coefficient and improved particle swarm optimization (PCC–IPSO) procedure for indicator/feature selection, and a parallel-pooling one-dimensional convolutional neural network with a bidirectional gated recurrent unit (1D-CNN–BiGRU) for downstream validation. IPSO uses Logistic–Sine chaotic initialization, nonlinear inertia control, asymmetric cognitive/social learning-factor schedules, and adaptive t-distribution mutation. Because no classroom dataset is introduced, the computational mechanism is examined on the five-class NSL-KDD benchmark as a high-dimensional, strongly imbalanced test bed. PCC reduces the 41 benchmark variables to 31, and the archived IPSO output contains 16 selected variables. The retained experimental record reports 99.06% aggregate accuracy for the integrated benchmark pipeline and favorable single-run convergence behavior on several optimization functions. However, the archived record does not preserve an independently verifiable final split identifier, repeated-seed statistics, synthetic-sample distance metrics, or computational-cost measurements; therefore these results are reported as recorded benchmark outcomes rather than as statistically validated superiority or official-KDDTest+ performance. Within these limits, the study illustrates how generative rebalancing and deliberately asymmetric heuristic search can be combined for reduced indicator selection under distribution asymmetry.