E-COMMERCE DEMAND FEATURE IDENTIFICATION AND HIGH-PRECISION FORECASTING BASED ON ENSEMBLE CLUSTERING AND REGULARIZED REGRESSION

Abstract

To address the issues of uneven distribution of e-commerce warehouse allocation demand and the difficulty in forecasting short-term fluctuations, this study proposes a comprehensive analytical framework that integrates feature extraction, preference identification, and time-series forecasting. Using hierarchical clustering algorithms to construct a clustering tree, combined with Ward’s method, the study divides the national distribution centers into four clusters with similar consumption characteristics. By utilizing the entropy weighting method to quantify the contribution of key indicators to consumption preferences, the study identifies transaction value as the core factor driving the distribution of distribution center demand. Based on this, a 62-dimensional high-dimensional feature space encompassing product attributes, user behavior, and trend characteristics was constructed, and a LASSO regression model with L1 regularization was introduced. By applying an absolute value penalty to the coefficients, this model automatically screened 32 core features. Experimental results demonstrate that this approach delivers superior performance: the R-squared value for the training set reaches 0.9999, and the residuals follow a normal distribution with a mean of 0.03 and exhibit white noise characteristics. In stability tests, the model exhibits exceptional robustness, with a coefficient of variation in R-squared of only 0.09%. This study not only reveals the regional patterns of demand for branch warehouses but also provides a decision-making basis for precise replenishment at the central warehouse over the next 14 days.