Abstract: To address the issues of initial centroid sensitivity and local optimal traps arising from clustering high-dimensional consumption data, this paper proposes a hybrid clustering method integrating Self-Organizing Map (SOM) neural network and K-means, tailored to the demand of cigarette customer segmentation. The silhouette coefficient method is adopted to objectively determine the optimal number of clusters. The SOM network conducts nonlinear dimensionality reduction to extract data topological structures and optimize the initial centroids of K-means. Meanwhile, the classic RFM (Recency, Frequency, Monetary) model is extended to a four-dimensional RFMI (Recency, Frequency, Monetary, Item) segmentation indicator system to improve the fitting degree of consumer segmentation. The experimental results are as follows: 1) Based on the RFMI sales terminal data of cigarette products collected from a certain region within a period of 2024, the proposed SOM+K-means hybrid model achieves a silhouette coefficient of 0.442, with a running time of 2.77 s and memory consumption of 0.458 MB. Its overall performance is significantly superior to other clustering algorithms such as DBSCAN. 2) The RFMI data are divided into five consumer segments via the SOM+K-means hybrid model, and the group characteristics of each category are analyzed. This method provides reliable technical support for precise consumer segmentation and lean cigarette supply allocation in the tobacco industry.