Lactation Performance Analysis Of Dairy Cows Based On Robust Additive Models

With the successful application of artificial intelligence technology in agricultural intelligent detection,diagnosis and prediction,intelligent agriculture has coming into a moment of flourish development.The analysis of lactation performance of dairy cows is one of the most important scenarios in agricultural information engineering.The machine learning method,as the core of artificial intelligence technology,is an effective way to improve the lactation performance of dairy cows.The blood indices are the important factors that directly reflects the physiological state of animals,which have been used to predict the lactation performance of dairy cows.However,due to the limited data size,large individual differences and other characteristics of cow blood parameter data,traditional learning models are difficult to effectively explore the intrinsic relationship between blood indices and lactation performance,resulting in the lack of robustness and interpretability of final results.To circumvent the above problems,we propose a novel class of interpretable machine learnings to achieve promising analysis of milk production performance.This paper collects the real-world data of 785 dairy cows from Wuhan Guangming Ecological Demonstration Dairy Farm Co.,Ltd,which are composed of 42 blood parameters(including 25 blood physiological indices and 17 biochemical indices)and 4lactation performance indices(i.e.,milk yield,milk protein percentage,milk fat percentage and somatic cell count).Firstly,the statistical properties of cow blood and lactation performance are characterized by means of ANOVA(Analysis of Variance)and Spearman correlation analysis.For instance,by analyzing the differences of blood indicators in different physiological stages of dairy cows(such as parity and lactation stage),it can provide detailed reference basis for pasture to monitor the changes of blood indicators of dairy cows in different stages.Secondly,inspired by the complex problems existing in milk cow data(such as outlier and linear inseparability),this paper proposes two novel interpretable machine learning methods to realize the robust estimation of milk cow performance and realizable inference respectively:a)We propose a novel class of robust additive classifier by integrating the pinball loss and sparsity-inducing penalty into additive model.Due to the properties of pinball loss and additive structure,our proposed method can not only reduce the perturbation of complex noise,but provide the reasonable variable selection(i.e.,mining significant blood indices).In the algorithmic optimization,we introduce the smoothing approximation technique and Nesterov’s gradient descent method to solve our method.In theoretical analysis,we establish the generalization error bound of our method.In real application,our method effectively identify high-quality cows based on blood indices,where the recall rate of high-quality cows is 98.65%,and F1 score is 75.32%.Moreover,we screen out the significant blood indices that are important to determine the lactation performance of dairy cows(e.g.,hematocrit,average platelet volume,and red blood cell distribution width).b)We propose a novel class of robust additive regression model by integrating modeinducing metric and sparsity-inducing penalty into additive model.Different from the traditional conditional mean regression(which requires zero-mean Gaussian noise),our method can reduce the perturbation of the heavy-tailed,skewed noises and outliers.In the algorithmic optimization and theoretical analysis,we solve our method by developing the half-quadratic optimization and alternating direction multiplier method.Moreover,we establish the excess risk of our method and provide the theoretical guarantee on variable selection consistency.In the prediction of lactation performance of dairy cows,our method can not only accurately estimate the milk yield,milk fat percentage,milk protein percentage and somatic cell count of dairy cows,but also select the significant blood parameters(such as the milk yield is significantly affected by hemoglobin,hematocrit,aspartate aminotransferase and other indices).In conclusion,this paper focuses on analyzing the lactation performance of dairy cows in intelligent dairy farming.To overcome the insufficiency of the application of robust and interpretable machine learning algorithms in the current agricultural engineering field,we proposes two novel robust additive,which effectively achieve the promising prediction of lactation performance and promote the common development of interpretable artificial intelligence and agricultural engineering.