作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• •    

基于可信度引导的轻量化交通预测课程知识蒸馏

  • 发布日期:2026-11-19

Confidence-Based Curriculum Knowledge Distillation for Lightweight Traffic Forecasting

  • Published:2026-11-19

摘要: 针对轻量化交通预测中学生模型预测精度受限,以及现有知识蒸馏方法难以兼顾教师知识可靠性差异与不同预测步学习难度的问题,提出一种基于可信度引导的轻量化交通预测课程知识蒸馏方法(CCKD)。首先,将教师预测误差分解为节点级与预测步级误差,经归一化构建复合可信度,以表征教师知识的可靠性。其次,设计可信度引导的双路径蒸馏机制,利用连续可信度及其互补权重分别调节预测值蒸馏与趋势蒸馏,使学生在学习可靠数值知识的同时保持预测步间的变化趋势一致性。同时,构建预测步动态软课程机制,综合学生预测误差、师生预测差距及教师可信度调整蒸馏权重,并随训练进程逐步提高长期预测步的基础权重。推理阶段仅保留轻量学生模型,不引入额外的蒸馏计算开销。在METR-LA和PEMS-BAY数据集上开展12步(60 min)交通速度预测实验,主要方法采用3个随机种子独立训练。以Graph WaveNet为教师、轻量图卷积网络(GCN)为学生时,CCKD的12步平均绝对误差(MAE)分别达到3.37 ± 0.01和1.72 ± 0.01,均优于学生独立训练与普通知识蒸馏。消融实验验证了该方法的有效性,跨模型实验进一步考察了方法在不同师生结构下的适用性。在METR-LA数据集、批大小为64的测试条件下,学生模型的参数量和推理时间较教师分别减少约91%和52.74%。在10%~40%的随机观测缺失条件下,CCKD的平均MAE均低于普通知识蒸馏。实验结果表明,CCKD能够在保持学生模型轻量化优势的同时提高多步交通预测精度,并在所测试的随机观测缺失场景下表现出一定的鲁棒性。

Abstract: Lightweight student models for traffic forecasting often have limited predictive accuracy. Existing knowledge distillation methods struggle to jointly address variations in teacher knowledge reliability and differences in learning difficulty across forecast steps. To address these limitations, we propose confidence-based curriculum knowledge distillation (CCKD) for lightweight traffic forecasting. Teacher prediction errors are first decomposed into node-level and forecast-step-level components, which are normalized to construct a composite confidence score representing teacher knowledge reliability. A confidence-guided dual-path distillation mechanism uses continuous confidence scores and their complementary weights to regulate prediction-value distillation and trend distillation, respectively. This enables the student to learn from reliable prediction values while maintaining consistency with the teacher in trends across forecast steps. A dynamic soft curriculum over forecast steps adjusts distillation weights based on student prediction errors, discrepancies between teacher and student predictions, and teacher confidence scores. The baseline weights assigned to longer-horizon forecasts are gradually increased as training progresses. Only the lightweight student model is retained during inference, avoiding additional distillation overhead. We evaluated CCKD on METR-LA and PEMS-BAY for 12-step traffic speed forecasting over a 60-min horizon. The main methods were independently trained using three random seeds. Graph WaveNet served as the teacher, and a lightweight graph convolutional network (GCN) served as the student. CCKD achieved mean absolute errors (MAEs), averaged over 12 forecast steps, of 3.37 ± 0.01 and 1.72 ± 0.01 on METR-LA and PEMS-BAY, respectively. Both values were lower than those obtained with independently trained student models and standard knowledge distillation. Ablation studies confirmed the method’s effectiveness, and experiments with different teacher and student architectures further examined its applicability. On METR-LA with a batch size of 64, the student used approximately 91% fewer parameters and required 52.74% less inference time than the teacher. When 10–40% of observations were randomly missing, CCKD consistently achieved lower average MAEs than standard knowledge distillation. These results indicate that CCKD improves multi-step traffic forecasting accuracy while preserving the computational advantages of lightweight student models. It also exhibits a degree of robustness to randomly missing observations under the tested conditions.