Table 1 Implementation steps of CRM-SLIE model algorithm.

Step 1

Input data preprocessing

The discrete features and continuous features of students are embedded. Discrete features are transformed into low-dimensional embedding vectors by embedding layer, and continuous features are standardized to form comprehensive feature vectors.

Step 2

Position coding

Position coding is introduced into students’ historical learning behavior data to preserve time sequence information. Calculate the position code and splice it with the embedded vector to get the final input vector X.

Step 3

GRU module processing

Input vector X is transmitted to GRU module, and the time-series interest features in students’ historical learning records are extracted through GRU network, and the potential interest state \(\:{h}_{t}\) is updated and generated.

Step 4

Evolution of students’ interests

Input vector X is transmitted to GRU module, and the time-series interest features in students’ historical learning records are extracted through GRU network, and the potential interest state \(\:{h}_{t}^{{\prime\:}}\) is updated and generated.

Step 5

Project crossing

Second-order features of history learning courses and candidate courses are crossed, and complex interactive information between courses is extracted by inner product and Hadamard product. A new candidate course feature vector \(\:{v}_{j}\) is generated through cross features.

Step 6

output module

The interest state \(\:{h}_{t}^{{\prime\:}}\), the candidate course feature \(\:{v}_{j}\) and the user portrait feature u are spliced into a dense feature vector M. It is transmitted to the multi-layer fully connected network and processed by the PReLU activation function.

Step 7

Recommendation probability calculation

The recommendation probability of each candidate course is calculated by softmax function, and the best recommended course is selected according to the probability distribution.

Step 8

Loss calculation and optimization

Using binary cross entropy as loss function, the difference between the actual label and the predicted probability is calculated. The model parameters are optimized by back propagation algorithm to improve the recommendation accuracy.