Decision-tree induction to interpret lactation curves

Pietersma, D., Lacroix, R., Lefebvre, D. and Wade, K.M. 2002. Decision-tree induction to interpret lactation curves. Canadian Biosystems Engineering/Le génie des biosystèmes au Canada 44: 7.1 7.13. Decision-tree induction was used to learn to interpret paritygroup average lactation curves automatically in dairy farming. Three parity groups were involved consisting of cows in their first, second, or third and higher parity. A dairy-nutrition specialist analyzed 99 parity-group average lactation curves, representing 33 dairy herds, and classified these curves using predefined aspects of interpretation. For machine learning, seven main classification tasks and three secondary tasks, supporting one of the main tasks, were identified. For each task, potentially predictive attributes were created, based on the graphical and numerical information available to the specialist. Five-fold cross validation was used to estimate the classification performance, and relative operating characteristic curves were used to visualize the achieved trade-off between sensitivity and specificity. For five of the seven main classification tasks, a series of three final decision trees was induced from the entire data set with increasing sensitivity, and associated with a low, medium, and high tendency of classifying new cases as abnormal. For the other two of the main tasks, alternative trees showed very similar performance. The medium tendency trees were chosen to lead to a probability of predicting new cases as abnormal, similar to the observed prevalence of abnormal cases, given a population of cases with that prevalence. The decision trees induced for the main classification tasks performed well. For the medium tendency decision trees, the sensitivity was at least 80% and the number of truly abnormal cases (as a percentage of all cases predicted as abnormal) was at least 75%. For the secondary tasks, the performance was poor, and domain expertise was required to select a plausible tree from alternative trees generated by the induction algorithm. The decision trees, ranging from two to seven leaf nodes, were evaluated by the domain specialist and, after a few adjustments, considered as plausible. This study suggested that automatically induced decision trees are able to match the interpretation of paritygroup average lactation curves closely as performed by a domain specialist. Machine-learning assisted knowledge acquisition is expected to be especially appropriate for problem domains where specialists have difficulty expressing decision rules, such as the analysis of graphical information.