Analysis and prediction of student evaluation scores based on bias SVD

Students' evaluation scores for teachers are significant indicators in the teaching evaluation process of a university or online course. There is a disadvantage in all existing evaluation methods, which regard students as the same and ignore the individual differences. To solve this problem, we propose a novel teaching evaluation method which is based on Bias SVD. Firstly, we convert the evaluation scores of teachers into a matrix. Then decompose this matrix by gradient descent and the biases of students in the evaluation process are iteratively obtained. By analyzing 63,193 evaluation records from 15 schools in Beijing Institute of Technology. We find that students who tend to give high scores have corresponding high offset values. We use a sentiment lexicon in the field of education to verify this method. By calculating emotion scores for teachers, we find that biases and scoring features are considerably correlative. Finally, we filtered the really too subjective scores through a certain threshold, and then used the XGBoost model to predict scores from the filtered data. It was shown that the combination method of Bias SVD and XGBoost can improve the accuracy of the prediction experimentally.