Record Details

Prediction of milk production using penalized regression techniques in cattle

KrishiKosh

View Archive Info
 
 
Field Value
 
Title Prediction of milk production using penalized regression techniques in cattle
 
Creator Hemant Kumar
 
Contributor Hooda, B.K.
 
Subject Yields, Selection, Livestock, Animal husbandry, Cultivation, Lactation, Physical control, Productivity, Biological phenomena, Layering
 
Description Multiple linear regression models (MLR) have been widely used in dairy sciences to predict lifetime milk production in cattle on the basis of lactation traits. MLR often gives unsatisfactory results in the presences of high multicollinearity among the explanatory variables. Choice of functional form and selection of xplanatory variables is also important for getting a parsimonious and useful model for explaining any phenomenon. In the presence of multicollinearity and model mis-specification ordinary least square estimators of regression parameters generally have low bias and large variances resulting poor predictive performance. Keeping in view the presence of multicollinearity in mind shrinkage and penalized regression techniques have been used along with the artificial neural network for prediction of lifetime milk production on the basis of lactation traits. In the present study lactation traits such as previous lactation yield, age at first calving, lactation length, calving interval, service period, and dry period have been used for prediction of lifetime milk yield in crossbred cattle data. The lifetime milk production has been defined as total amount of milk produced by cattle from initiation of first lactation till the completion of third lactation. Small eigen values of correlation matrix of predictor variables, high value of variance inflation factor and higher condition index indicated presence of multicollinearity in crossbred cattle data. Consequently biased and penalized regression models have been adopted to take care of multicollinearity among the predictors. In addition to ridge regression the relatively recent techniques of penalized regression called LASSO and elastic net given by Tibshirani (1996) and Zou and Hastie (2005) respectively were also applied for developing prediction model for lifetime milk production and selection of principal lactation traits. On the basis of AIC and BIC values LASSO and elastic net outperformed the ridge regression and elastic net techniques was found most satisfactory. Forward selection, backward elimination, LASSO and elastic net were used for selection of best subset of lactation traits for prediction of lifetime milk production. It was observed that seven variables out of eleven were selected by LASSO and six by elastic net using optimal value of regularization parameters. The optimum value of regularization parameters was computed using 10- fold cross validation. The number of traits in best subset was found to six for backward elimination and four for forward selection method. On the basis of adjusted R2, AIC and BIC values and simplicity of the model it was concluded that subset selected by LASSO techniques having just two significant traits was best. Evaluation of predicting performance of multiple regression, ridge regression, LASSO, elastic net and ANNs models has been done by dividing the sample under study into two sets, by taking 90% observations in training set and 10% observations on test set. Coefficient of determination, root mean square error, mean absolute error, mean absolute percentage error and Theil’s U-statistics were computed for the test set, and based on these performance measures elastic net was found most satisfactory techniques for prediction of lifetime milk yield using lactation traits in crossbred cattle.
 
Date 2016-02-27T13:06:26Z
2016-02-27T13:06:26Z
2013
 
Type Thesis
 
Identifier http://krishikosh.egranth.ac.in/handle/1/64463
 
Language en
 
Format application/pdf
 
Publisher CCSHAU