基于机器学习的在线快时尚商品退货预测研究

Research on online fast fashion product return prediction based on machine learning

商品退货是影响在线快时尚运营绩效的重要因素,有效识别退货的重要因素、预测退货行为对提高在线快时尚业的运营绩效有重要意义。本研究基于真实在线快时尚商品退货数据,对比了7个应用机器学习模型的预测表现,包括决策树、随机森林、梯度提升决策树、轻量级梯度提升机、极端梯度提升和分类特征支持的梯度提升6个基础模型和一个堆叠模型,并对6个基础预测模型中的特征重要性进行比较,以识别影响退货的重要变量。通过6个评价指标来评估7个模型的退货预测的表现,结果表明影响产品退货的最重要的3个因素,即订单总花销、推荐购买价格和付款方式;分类特征支持的梯度提升模型的综合预测表现优于其他5个基础模型,更加适用于识别在线快时尚商品的退货预测;而Stacking组合堆叠模型在该数据集中并没有进一步提高预测的精度。

Merchandise returns are an important factor that affects the operational performance of online fast fashion.Effectively identifying important factors affecting returns and predicting return behavior are important to improve the operational performance of the online fast fashion industry.Based on real online fast fashion merchandise return data,this study applies machine learning models to compare the predictive performance of seven models,including this study compares the prediction performance of seven models,including six base models of Decision Tree,Random Forest,Gradient Boosting Decision Tree,LightGBM,XGBoost and CatBoost and a stacking model,and compares the importance of features in the six base prediction models to identify the important variables.The performance of the seven models for return prediction is evaluated by six evaluation metrics.The results show that the three most important factors affecting product returns,i.e.,total order spend,recommended purchase price,and payment method;the combined prediction performance of the CatBoost model outperforms the other five base models and is more suitable for identifying return predictions for online fast fashion items;and the Stacking combination model does not further improve the prediction accuracy in this dataset.