Product Return Forecasting And Influence Factors Research For Cloth Industry In B2C Environment

Low-efficiency continues to bother the retailers in B2 C from their daily operation,and product return is one of the m ain reasons for that.Existed researches have showed various factors leading to high return ra tes during online purchasing,Little researchers have shown how the factors influence the product return rates.In this paper,an empirical reasearch is performed besed on a historical order inform ation from a German online retailer,including an exploratory analysis to data sets,analys is of the im portance of factors about product return,and product return prediction.(1)Based on a variety of visualization tools,an exploratory analysis is developed for variables in the data set.the missing values in original data set are im putated through thermal platform interpolation,the left m issing-data and the abnorm al values which are outside of the interquartile range are considered as missingness at random,and are f illed with random values inside of interquartile range.Four kinds of variables are built on basis of the original data s et,including age,deliv ery time,lifetime and basket s ize.After preprocessing the data,num erical and catego rical variables are no rmalized and o ne-hot encoding seperately.(2)The data set after processing is divide d into trian se t and test set in 7:3 seperately,and both are used to build five kind of return rate forecasting m echines,including Logistic,CART,NNET,GBM and Xgboost,and develop a com prehensive analysis of the importance for price,size,color,delivery time,age,lifetime,basket size,city and salutation in product return s.Although the rank of im portance results for each variable are various among different models,price and the basket size are the most two important variables.(3)Five index,including accuracy,true positive rate,true negative rate and F score,can be calculated by the confusion matrix.Above five indexes,along with AUC,are used to evaluate the performance of five models in product return.The comparison shows that in both train set and test set,Xgboost algorithm,GBM algorithm and classification tree model performed better th an neural network model and Logistic at th e index of accuracy,true positive rate,F score an d AUC;and Xgboost algorithm is the best perf ormed at Accuracy and AUC,the ROC curves also verify these c onclusions.However,at the index of true negative rate,neural network models performed best among five models.(4)Different scales of train sets and te st sets are used to build models,and the predicting results of models are compared under different indexes.Results show that at the index of accuracy,true positive rate,F score and AUC index evaluation,scale of data set has no apparent influence on forecasting performance on product return rate of models,while at the index of precision,scale of data set can lead to significant dif ference among models' performance.