Determination of factors affecting dairy cattle: a case study of Ardahan province using data mining algorithms

ABSTRACT This study was conducted to compare predictive performances of different data-mining algorithms for determining factors influencing the average daily milk yield at dairy cattle enterprises of Ardahan province, located in the Eastern Anatolia region of Turkey. The algorithms employed in the present study were Classification and Regression Tree (CART), Chi-Square Automatic Interaction Detector (CHAID), Exhaustive Chi-Square Automatic Interaction Detector (Exhaustive CHAID), Multivariate Adaptive Regression Splines (MARS), and Multilayer Perceptron (MLP). The MARS algorithm outperformed the other algorithms in the study. Visual results of CART revealed that the culture-breed cows with a lactation length greater than 237.500 days had the highest milk yield (10.64 kg/day). Culture-breed cows calving earlier than the 4th month gave the highest yield of approximately 10 kg/day in the regression tree of CHAID. The Exhaustive CHAID results were almost the same as the structure of the CHAID. The use of MARS may provide an opportunity to detect factors affecting milk production (breed, feed supply, type of milking, mastitis control, cow year group, and lactation length) and their interactions. Moreover, the MARS algorithm may be useful in making an accurate decision about increasing milk yield per cow.

摘要 本研究旨在对比不同数据挖掘算法的预测性能，以明确土耳其东安纳托利亚地区阿尔达汉省奶牛场中影响奶牛日均产奶量的关键因素。本研究所采用的算法包括分类与回归树（Classification and Regression Tree, CART）、卡方自动交互检测（Chi-Square Automatic Interaction Detector, CHAID）、穷尽式卡方自动交互检测（Exhaustive Chi-Square Automatic Interaction Detector, Exhaustive CHAID）、多元自适应回归样条（Multivariate Adaptive Regression Splines, MARS）以及多层感知器（Multilayer Perceptron, MLP）。本研究中，多元自适应回归样条（MARS）算法的综合表现优于其余所有受试算法。

分类与回归树（CART）的可视化分析结果显示，泌乳时长超过237.500天的培育品种奶牛日均产奶量最高，可达10.64 kg。在卡方自动交互检测（CHAID）的回归树分析中，于第4月龄前产犊的培育品种奶牛日均产奶量最高，约为10 kg。穷尽式卡方自动交互检测（Exhaustive CHAID）的分析结果与标准CHAID的模型结构基本一致。多元自适应回归样条（MARS）算法可用于识别影响奶牛产奶量的各类因素（品种、饲料供应、挤奶方式、乳腺炎防控、奶牛年度分组以及泌乳时长）及其交互作用。此外，该算法还有助于为提升单头奶牛日均产奶量制定精准决策。