A viable data driven method for the assessment of the productivity level of dairy cows in future lactations

The challenge of increasing the economic and environmental sustainability of the dairy cattle sector involves several aspects and among these, the milk production throughout the career of a cow, is perhaps the parameter that all farmers would like to know for a more efficient planning of entries and exits. In fact, if on one hand numerous researchers are studying the problem selecting most efficient animals, on the other hand, few studies have focused on the definition of tools forecasting the productivity of dairy cows in the future lactations starting from the data collected in past lactations. This aspect has a particular importance in the first years of a cow since, as well known, first lactation usually has lower production than subsequent lactations. For a farmer it is important to know, as soon as possible, if a specific animal will have on a long term, lactations with high, medium or low milk productivity. The current availability of large dataset collected by automatic milking systems or by electronic milking parlors, paves the way for application of big data approaches based on machine learning algorithms with classification learner representing one of the most promising data-driven tools. In this study, firstly two supervised learning methods, i.e., Super Vector Machine and K-Nearest Neighbors, have been applied to a large dataset of 720 complete lactations, with the object to train machine learning tools able to classify and separate first and second lactation. The two classification algorithms have been applied to the raw dataset and after the application of a dimensionality reduction method. Four different dimensionality reduction methods (i.e., ISOMAP, UMAP, MDS and t-SNE) have been tested to evaluate the most efficient for this application. Finally, the same two classification algorithms have been used for the attribution of the productivity level of the second lactation starting from data of the first lactation. The two classification methods reached very encouraging accuracy values, ranging from 70% to 73%, indicating that selected predictors despite their simplicity look very promising and entail for the definition of enhanced future models. In fact, the method is particularly interesting for practical applications, as it represents a viable support-to-decision tool for selecting the most productive animals. Interpretive summary The need to increase the sustainability of the animal production sector makes essential to search for new tools to support farmers and vets, and in this the modern numerical techniques based on artificial intelligence are becoming increasingly popular. Especially in the dairy sector, the increase in sustainability is closely linked to the selection of the cows capable of guaranteeing high milk production. In this work a new approach based on machine learning models is proposed for the identification of the most productive animals and for the assessment of the long-term milk productivity class of a specific cow.