融合注意力机制的学生退课行为预测

Prediction of Student Dropout Based on Attention Mechanism

近年来,大规模在线开放课程(MOOCs)发展迅速,吸引了学界的广泛关注。用户退课率极高这一问题的长期存在,使得退课行为预测成为了一个重要的研究课题。目前的退课行为预测模型过于依赖传统的机器学习算法。此外,很多研究忽视了MOOCs用户灵活修课的特点,采用时序无关的方法进行预测。针对目前该领域存在的问题,论文提出了一种融合注意力机制的时序预测模型。该模型首先利用长短期记忆网络从原始的时序数据中学习新的时序隐态表示,再使用多个一维卷积神经网络提取隐态中各类特征的时序模式,最后融合注意力机制,使模型能够通过注意力分布值强化有效特征。实验结果表明,该方法的预测能力优于其他方法。

In recent years,massive open online courses(MOOC)have developed rapidly and have attracted widespread atten⁃tion from academia.The long-term existence of the problem of extremely high user dropout rate makes the study of dropout predic⁃tion an important topic.Most existing methods still rely on traditional machine learning algorithms.Moreover,many researchers ig⁃nore the characteristics of MOOC and use time-independent methods for prediction.Different from existing works,a temporal model with an attention mechanism is proposed to solve the problem.This model first uses long short-term memory to learn new temporal hidden state representations from the original time series data and then uses multiple one-dimensional convolutional neural net⁃works to extract the temporal patterns of various features from the hidden state.Finally,the attention mechanism is integrated to en⁃able the model to select relevant features and weight them to enhance their importance.Results of the experiments show the proposed method outperforms other approaches.