A deep fusion framework for end-to-end multi-product inventory optimization in e-commerce scenarios

Inventory management has become extremely difficult in the e-commerce industry due to the variety of products, the complexity and unpredictability of demand, and the frequency of discounts. According to previous research,it is simple to arrive at the local optimal solution using the traditional two-step method, which first predicts needs and then optimizes inventory. This method ignores the relationship between items and the impact of promotions. To address the challenges of multi-period and multi-product inventory management in the e-commerce scenario, this paper proposes an end-to-end integrated optimization model based on deep learning. This model first uses the Dynamic Programming or Rolling Horizon Control with Integer Linear Programming to determine the ideal order quantity within each ordering cycle by leveraging historical data, which is then used as a training label. Subsequently, a Multi-Product Inventory Optimization Neural Network model (MPIONN) is designed. This model integrates modules such as Long Short-Term Memory networks, attention mechanisms, and Graph Convolutional Networks to model the time dependence of demand, promotional activity information, and multi-product correlations respectively. Finally, the optimal order quantity is output through multi-source information fusion. By introducing an end-to-end differentiable mapping between features learning and inventory optimization, we eliminate error accumulation across sequential stages. This study validates the proposed method using a large-scale e-commerce dataset. The experimental results show that the algorithm proposed in this paper can significantly reduce the total inventory cost, effectively balance the inventory holding cost and the stock-out cost, and outperforms existing integrated optimization methods and two-stage methods in terms of performance.