基于机器学习和夏普利加法解释(SHAP)模型的饲料原料价格可解释预测  

Explainable prediction of feed raw material prices based on machine learning and SHAP

作　　者：吴展 王春晓[1] WU Zhan;WANG Chun-xiao

机构地区：[1]上海海洋大学经济管理学院,上海201306

出　　处：《饲料研究》2023年第23期178-181,共4页Feed Research

基　　金：国家现代农业产业技术体系(项目编号:CARS-47)。

摘　　要：文章旨在评估机器学习模型的性能,提出一种饲料原料价格可解释预测的框架。选取豆粕为饲料产品原材料的代表品种,基于2006年1月至2023年4月的豆粕期货月度结算价数据,采用反向传播(BP)神经网络、梯度提升决策树(GBDT)和极限梯度提升(XGBoost)等3种机器学习算法进行训练测试,使用贝叶斯优化算法调整各模型参数,选择性能最优模型结合SHAP模型解析预测结果。结果显示,贝叶斯优化的极限梯度提升算法(BO-XGBoost)模型的预测性能显著优于其他基准模型,其测试集的平均绝对百分比误差(MAPE)和决定系数(R2)分别为0.03和0.892,模型精度较高;滞后一期豆油期货结算价对豆粕价格具有显著正向影响。研究表明,该模型具有良好的应用前景,可为饲料相关企业管理者决策和有关部门制定政策提供一定参考。The purpose of the study is to evaluate the performance of machine learning models and to propose an interpretable prediction framework for feed material prices.Soybean meal was selected as the representative raw material of feed products.Based on the monthly settlement price data of soybean meal futures from January 2006 to April 2023,BP neural network,GBDT and XGBoost machine learning algorithms were used to conduct training tests,and then Bayesian optimization algorithm was used to adjust the parameters of each model.Finally,the optimal model and SHAP model are selected to analyze the prediction results.The prediction performance of the BO-XGBoost model proposed in this study is significantly better than that of other benchmark models.The MAPE and R2 of the forecast set are 0.03 and 0.892,indicating a high accuracy of the model.The research shows that the model has a good application prospect,and can provide some reference for the decision making of feed-related enterprise managers and relevant departments.

关 键 词：机器学习 SHAP模型 贝叶斯优化 可解释预测 饲料原料价格 

分 类 号：S8-9[农业科学—畜牧兽医]