World Scientific Reports

This paper focuses on the study of student behavior-based grade prediction. This research investigates grade prediction based on student behavior. After evaluating student behavior data from campus events, an attention-based CNN-LSTM student grade prediction model is developed. Initially, we use Convolutional Neural Network (CNN) to extract deep student behavior features and maximum pooling approach to pick the significant features in student behavior features; the retrieved features are then used as the input of Long and Short-Term Memory networks. A temporal attention mechanism is included at the output of LSTM in order to allocate attention weights for different weekly student behavior characteristics in a reasonable manner, hence improving the model's performance in terms of prediction. Experiments revealed that the student performance prediction model suggested in this paper had excellent performance.

At present, deep learning has been successfully applied in the field of medical diagnosis, and it can effectively improve the diagnostic accuracy by using deep learning to predict brain tumor images. Based on traditional convolution neural networks tend to ignore the problem of key location in brain tumor image information, this paper proposes a door control channel attention conversion unit, door control channel attention switching unit by using the relationship between a small number of parameters to change the channel, under the condition of without any increase in computational cost more significantly improve classification accuracy. Also, the Sequeeze and Excitation block is introduced for the communication between the channels. Finally, a multipath attentional network model (FSnet) for brain tumor image classification is constructed based on ResNeXt network model. Experimental tests on brain MRI data sets show that FSnet has better classification performance than traditional convolutional neural networks.

This paper is mainly based on the prediction of English grades, aiming at the data of students' English grades and personal identity during school, and establishes a fuzzy neural network prediction model based on the spike mechanism and genetic algorithm. First of all, in the aspect of parameter training of neural network optimization design, this paper adopts genetic algorithm as the learning algorithm of fuzzy neural network, so that the network can achieve the global optimum. Then, the spiking mechanism proposed based on the cerebral cortex information transmission mode and the Spiking neuron accumulation trigger (Integrate-and-Fire, IF) model is used to complete the structure growth and pruning of the fuzzy neural network, realize the dynamic adjustment of the network structure during the training process, and improve the performance of fuzzy neural network in English grade prediction application. Finally, after testing, it is found that the fuzzy neural network prediction model based on the spike mechanism and genetic algorithm proposed in this paper has a good performance.

With the rapid development of education for overseas students in China, psychological problems caused by cross-cultural factors have become increasingly prominent. Based on the psychological questionnaire data, in order to make full use of the relationship information between students, this paper proposes a psychological crisis individual identification model based on bipartite graph convolution network model (B-GCN) to make up for the shortcomings of traditional identification methods. Because the GCN model can not realize inductive learning, this paper improves the GCN model based on the structural characteristics of psychological tests, and proposes a bipartite graph neural network model. The model can classify the students who have never appeared in the graph structure, so as to realize the early warning of psychological crisis. Experiments show that the proposed B-GCN model has a good performance.

This report takes learning behavior and psychological support as the factors to predict the students' academic performance. A total of 13 factors were selected at the initial stage. Firstly, PCA and MDT algorithms were used for data preprocessing, and we used CNN to extract deep student behavior features. Then, the maximum pooling approach is used to pick the significant features; next, the selected features are used as the input of LSTM. A temporal attention mechanism is introduced at the output of LSTM to allocate attention weights for different weekly student behavior characteristics. This paper divides students' academic performance into several score intervals and makes a multi-classification prediction. This study uses 39 algorithms based on six types to predict the student's academic performance. The result shows that: (1) The average prediction accuracy based on the deep learning model is the highest (85.46%) in the six types of models; (2) The accuracy of the LSTM algorithm (88.36%) outperformed all the benchmark models. (3) The attention mechanism introduced can effectively improve the model's prediction accuracy; the proposed model in this paper outperformed all the benchmark models. Furthermore, the statistical analysis found that (1) Students who have academic confidence (confidence in teachers, courses, and themselves) may obtain higher academic performance; (2) A clear learning motivation is essential to final academic performance; (3) Lack of confidence in others in the study group may ultimately improve students' academic performance by encouraging them to work harder.