Decision tree technique applied to pig farming datasets

Abstract The decision tree technique is an effective instrument for making large datasets accessible and different sow herd data comparable. This technique can be used to improve the detected differences and weak points in farm management. The calculation of the decision tree based on the C4.5-algorithm, which constructs trees in a top down recursive strategy. It calculates the ranking of the attributes within the tree with the gain ratio criterion. In datasets from two Northern German sow herds, the decision tree technique was used to generate such trees and exemplarily classify the binary farmer decision regarding replacing or not replacing a sow with a gilt. The datasets were accumulated for farm A between 1994 and 2001 and for farm B within a time span from 1984 to 1999. The datasets consisted of 14,897 and 21,818 observations from an average of 386, respectively, 484 sows per year. The C4.5-decision tree algorithm generated decision trees, which showed different sizes ranging from 15 to 55 nodes. In relation to the sow herd performance, the threshold values of the attributes at the branches varied between the trees. The sensitivity, the kappa value and the error rate were the evaluation parameters for estimating the algorithms performance in classifying the present unbalanced datasets. For both datasets the results reached a sensitivity value of 39.2–47.3% and the kappa value showed a rate between 44.9% and 53.9%. The error rate varied between 14.2% and 19.2%. Additionally, the datasets were modified in order to identify the reasons for partially false classification. The datasets were reduced by excluding the replacement decisions, such as aggressive sows, which are not reflected in the sow reproduction performance. Henceforth, the sensitivity reached values of between 69.4% and 87.5% and the error rate (10.1–15.0%) decreased.