Predictive Modelling For E-Commerce Logistics Using Minimum Redundancy Feature Selection and Deepsmote with Non-Linear Analysis

Abstract

In the fast-growing e-commerce industry, ensuring consumer satisfaction depend on efficient logistics. Predictive modelling in major part determines how best to optimise numerous logistical operations like demand forecasting, inventory control, and delivery routing. Still, the imbalanced, high-dimensional character of logistics data causes great challenges. To address these difficulties, we propose a predictive modelling framework combining non-linear analysis techniques with Deep Synthetic Minority Over-sampling Technique (DeepSMote) coupled with minimum redundancy maximum relevance (mRMR) feature selection. This paper tackles the problem of achieving correct predictions in e-commerce logistics coming from duplicated attributes and imbalanced datasets producing biassed models. Starting mRMS, the suggested method reduces dimensionality by selecting highly significant but minimally duplicated features, hence improving the model's efficiency. Deep SMote thus addresses class imbalance by generating synthetic samples for minority classes using deep learning techniques. Then complex correlations in the data are obtained via non-linear analysis—more notably, kernel-based methods. The results reveal that using numerous logistics criteria, our approach significantly increases prediction accuracy. For demand forecasting, for instance, our model utilising traditional methods got a Mean Absolute Error (MAE) of 2.3% against 5.7%. The Root Mean Square Error (RMSE) changed in the delivery time predicted from 4.8 hours to 1.9 hours. Moreover implying better accuracy and recall, the F1-score of the model for identifying high-risk delivery routes improved from 0.72 to 0.88. These results suggest that mRMR and DeepSMote combined with non-linear analysis can significantly improve predictive modelling for e-commerce logistics, therefore enabling more dependable and effective operations.