基于机器学习算法的服装直播销量预测模.

With the dramatic increase in the scales of e-commerce livestreaming the number of e-commerce livestreaming users has reached 48. 8% of the overall Internet users in China. The huge supply demand requires live e-commerce stores to improve their dispatching efficiency and reduce inventory. Therefore in order to avoid retailers' profit loss it is necessary to find a more accurate method to predict livestreaming sales. The sales prediction methods mainly include traditional statistical methods and machine learning algorithms. Due to the instability of livestreaming sales and the large number of influencing factors traditional statistical methods often fail to predict the sales accurately. To complete the index system of livestreaming sales prediction and improve the accuracy of livestreaming sales prediction this paper adopted a variety of machine learning algorithms BP neural network decision tree DT random forest RF K-nearest neighbor KNN and support vector machine SVM analyzed the influencing factors of apparel livestreaming sales predicted apparel livestreaming sales and selected the best performing algorithms. The detailed research process is as follows. Firstly 17 influencing factors of livestreaming sales were selected through literature review and nine most important influencing factors were selected by using Spearman' s correlation coefficient combined with significance. Secondly different machine learning algorithms were used to establish clothing sales prediction models and the method of 5-fold cross-validation was adopted to initially screen out three algorithms RF KNN and SVM with high and stable model fit with R² as an indicator. Finally the parameters of the three algorithms were optimized and then three prediction models were constructed. R² MAE RMSE and MAPE were used as evaluation indexes and the optimal algorithms were selected by using the method of 5-fold cross-validation to test the performance of each model. The results of the study show that the multicollinearity between the nine factors number of fans of the anchor average number of viewers of the anchor in the last 30 days average pit output of the anchor in the last 30 days product price duration of product explanation historical sales of the product in the last 30 days number of fans of the brand historical sales of the brand in the last 30 days and discounts is weak and their correlation with the livestreaming sales is significant. Therefore these nine factors can be used as influencing factors in the prediction model. Among the influencing factors the correlation among product sales in the last 30 days the duration of product explanation and livestreaming sales is the highest. In the meanwhile the prediction algorithms KNN and RF perform better with R² being greater than 0. 98 and MAPE within 30. 5% . Compared with the KNN algorithm the RF algorithm is more stable and its R² RMSE and MAE perform better than those of the KNN algorithm. But the MAPE of the KNN algorithm is smaller than that of the RF algorithm for which the possible reason is that the KNN algorithm is more accurate in predicting low sales items and the relative error is smaller. According to the result of 5-fold cross-validation the RF algorithm is more stable compared with the KNN algorithm and the possible reason is that the KNN algorithm is more suitable for the dataset with more similar data features. Therefore RF can be used as the main prediction algorithm in practical applications to ensure the stability of the overall sales trend prediction. In predicting the sales of the same brand or the same category the similarity between the data is higher and then the KNN algorithm can be considered for prediction. This paper compares the performance of various prediction algorithms on livestreaming sales prediction optimizes the parameters and improves the accuracy of livestreaming sales prediction. The prediction results can help retailers make inventory planning adjust production schedules develop marketing strategies and provide data support for product purchasing pricing and promotion. Due to the fact that only some of the easily quantifiable influencing factors are explored in this paper and the sample distribution is limited future research can expand the scope of sample selection and further improve the predictive indicator system to achieve more accurate predictions. [ABSTRACT FROM AUTHOR]