用于加速预训练知识追踪模型的深层特征提取器

An accelerated deep feature extractor for pre-trained knowledge tracing models

通过对知识追踪进行预训练的嵌入与微调,可以显著提高知识追踪任务的准确性,但现有的微调方法通常采用全微调的方式,即调整预训练模型的所有参数,效率较低。本文提出了一种新的高效微调方法,称为快速适配训练方法,通过提取深层特征来调整模型,以适应不同的下游任务。与全微调方法相比,快速适配训练方法在ASSIST 09和EdNet数据集上能够实现5.2倍和3.6倍的训练时间加速,并且同时提高了0.49%和0.19%的预测精度。

The accuracy of the knowledge tracking task can be significantly improved by embedding and fine-tuning the pre-training for knowledge tracking,but the existing fine-tuning methods usually use full fine-tuning,i.e.,adjusting all parameters of the pre-trained model,which is less efficient.In this paper,we propose a new efficient fine-tuning method,called the fast adaptation training method,to adjust the model to different downstream tasks by extracting deep features.Compared with the full fine-tuning method,the fast adaptation training method is able to achieve 5.2 times and 3.6 times training time speedups on the ASSIST 09 and EdNet datasets,and improves the prediction accuracy by both 0.49%and 0.19%.