Abstract: Learning performance prediction can help teachers to intervene in time by using the learning behavior data of students on the online learning platform to identify at-risk students,but it faces the problem of data imbalance,which makes it particularly difficult to accurately identify at-risk students.Addressing the issues that the mainstream deep generative model VAE,cannot guarantee the rationality of generated samples in current solution strategies,and GAN tends to introduce new errors when processing time-series data,with either over-training or under-training of the generator or discriminator leading to a decline in the quality of generated data,this paper proposes a new prediction method for student learning performance based on generative mo-dels.Firstly,the VAE based on bidirectional long short-term memory(BiLSTM) is utilized to initialize the GAN,enabling it to start training from a more stable point while also better understanding the correlation and periodic characteristics between subsequent data points in the student behavior sequence.Secondly,a multi-head attention mechanism is introduced in the discriminator part to enhance its ability to distinguish between real data and generated data,and then continue to game with the generator.Finally,the deep generative model and the classical resampling strategy SMOTE are integrated based on the idea of Blending ensemble learning,which effectively combines the advantages of data and algorithm to improve the overall generation ability of the model.A large number of experimental results on real student data sets show that the model can generate high-quality data to improve the recognition ability of the prediction model for at-risk students,and is superior to the baseline method in multiple evaluation indicators.

Key words: Learning performance prediction, Generative model, Bidirectional long short-term memory, Attention mechanism, Ensemble method