Feature Matching Based Heterogeneous Transfer Learning for Student Performance Prediction

Improving the performance of predicting student performance by transfer learning has become a new research focus in the field of education data mining. However, most works concentrate on homogeneous transfer learning, therefore they cannot solve the feature heterogeneity between source and target domains. A heterogeneous transfer learning algorithm with feature matching based on correlation coefficients and fine-tuning (FMBCC-FT) is proposed. FMBCC-FT first normalizes the source and target domain data, then selects and matches features according to the correlation coefficients between the features and the labels, finally trains the logistic regression model with source domain data and fine-tunes with target domain labeled data. The experiments are carried out with the open datasets of edX as the source domain and the learning data from Jilin University as the target domain. The results show that the proposed method outperforms the related works and confirm its effectiveness and competitiveness.