Research On The Prediction Approach Of University Student Scholarship And Grant Based On LGB-LR Model

With the spread of network information technology and the construction of smart campus,colleges have developed various information platforms successively.Among them,the campus card is an indispensable tool for school teachers and students in daily life,and the massive data generated every day is of great value to the research of college student management.The assessment of scholarship and grant is an important part of the college management,as well as an important indicator to help colleges for poverty identification and academic early warning.However,many colleges only regarded it as a routine task,and failing to proceed from the perspective of student management,then ignoring the value hidden in the rich campus resources.The method of college student management based on campus data has value in theoretical research and practical application.This article combines machine learning algorithms to explore and research the consumption data of campus cards,and provides new ideas for college administrators to accurately identify poor students and predict students’academic risks.This article mainly had done the following work:(1)Collected the data for research,and processed the data acquired,like consumption data of campus card and data of scholarship and grant,for obtain the data that meets the requirements of the model.(2)In order to solve the feature engineering problem of logistic regression model,constructed a prediction model based on LightGBM and logistic regression,and used Sklearn to help logistic regression achieve multi-classification by the OVR packaging function,then set the parameters and evaluation indicators of the models.Then selected six public data sets from UCI(Breast Cancer Wisconsin,Heart Disease and Diabetes are binary data sets,then Seeds,Wine and Wall-Following Robot Navigation are multi-class data sets),trained on the six public data sets with LightGBM model,LR model and LGB-LR model respectively,and used the five-fold cross-validation method for model verification.According to the evaluation index,came to a conclusion that the performance of LGB-LR has been improved on the basis of LightGBM and LR.(3)In order to verify the predictive performance of the logistic regression model with different feature selection methods,five models were used for comparative experiments,they are LR with~L1penalty,LR with PCA,LR,LightGBM and LGB-LR.Then selected the optimal parameters of the models by Bayesian,and used ten-fold cross-validation for model verification.The experimental results showed that,for the poverty identification,the accuracy,F1 score and AUC of LGB-LR are 89.54%,87.49%and 0.87,respectively,higher than the other models,except that the F1 score is slightly lower than the LightGBM.For the academic prediction,the accuracy and macro-F1 of the LGB-LR are 77.72%and73.82%,are slightly lower than the LightGBM model,but better than the other three models.In summary,the prediction effect of the LGB-LR model is relatively ideal,especially suitable for dealing with two-classification problem.In multi-classification problem,the performance of the logistic regression based on the LightGBM is higher than other feature selections,indicating that the LGB-LR has a certain reference value for the accurate identification of poverty and the assessment of academic risk.Of course,the logistic regression model still has some limitations on the multi-classification problem.So increasing the diversity of data sources and improving the prediction accuracy of the multi-classification models are the key work of future research.