Model combination of lasso on regularization path

作者：Wang Mei; Liao Shizhong^*

来源：Journal of Computational Information Systems, 2014, 10(2): 755-762.

DOI：10.12733/jcis9618

• Effectiveness and efficiencies
• Lasso
• Model combination
• Piecewise linearities
• Posterior probability
• Prediction performance
• Regularization parameters
• Regularization paths

Model combination can improve prediction performance of regression model by an optimal integration of all given base models. In this paper, we propose an effective method for model combination of Lasso for improved prediction performance. First, we construct the candidate model set on the Lasso regularization path. The regularization path can record the Lasso solutions for all values of the regularization parameter, moreover, its inherent piecewise linearity makes the construction of the candidate model set simple and efficient. Then, in the testing or predicting phase, we apply the minimal neighborhood method to determine the input-sensitive combination model set, and perform the Beyesian model combination. Finally, we estimate the posterior probability of each model in the combination model set using the BIC criterion, which can be computed along with the Lasso regularization path algorithm. Experimental results demonstrate the effectiveness and efficiency of the model combination method.

• 出版日期2014
• 单位天津大学