Variable Selection For "Forest Bats Activities" Data By Weighted Fusion Method

Variable selection in statistical modeling process is an extremely important issue, but there are some drawbacks in the traditional variable selection methods,. Especially when p is large comparing with n and the correlation between certain variables is strong, some variables which have weak impact or no impact while using it to predict are selected for the multiple regression equation, so it makes the accuracy of the estimates and predicts declined.To solve this problem, Daye and Jeng (2009) proposed Weighted fusion variable selection method, a good method to overcome some shortcomings of the traditional methods. In this paper, we use Weighted fusion variable selection method proposed by Daye and Jeng (2009) to select data variables. Calculating the sample correlation coefficient between variables and the weighted fusion of the estimates and projections ,use the weighted fusion penalty function to get the L2 norm fused lasso estimates. Through the "Forest bats activities" to analyze this data to carry out variable selection. First, analyzing of the sample correlation coefficient between the prediction variables get a lot of highly relevant variables, therefore, using the traditional methods to choose a variable is not good. Next, showing and analyzing the residual plot of the response variables and predictor variables , obtain a visual link between the response variables and the predictor variables. Re-use of the weighted fusion, sort the predictor variables according to the variable importance of it. Last,we use AIC criterion and BIC criterion to select variable,.The actual data and examples show that the weighted fusion can effectively carry out variable selection to improve prediction accuracy. So that we can facilitate the calculation of the variable selection.