Near-infrared spectrum discriminant analysis based on information extraction by using the elastic net

Elastic net method combines the merits of ridge regression and Lasso method. It reduces model prediction error by variable selection while not over-shrinking regression coefficients. In this paper, we take advantages of the elastic net´s good properties of variable selection and simultaneous parameter estimation to select the important principal components, then establish discriminant model and apply it to near-infrared spectroscopy quantitative analysis. In the real data set analysis, 103 rhubarb samples were randomly split into two groups, one is viewed as training set which contains 35 samples, another group is considered as testing set which contains 68 samples. All of the samples´ protein contents are measured by the national standard Kjeldahl method and the data were called chemical values. In order to testify feasibility and stability of the method, the training set and testing set were conducted random split and analyzed for ten times, respectively. According to these predicting results, the maximum number of false positives was 10, the minimum number of false positives is 5, and average false positive rate is 11.76%. These results showed a significant improvement compared to the results which derived by using ordinary principal component method directly.