Scholarlink

<p>Traditional LSTM models in e-commerce inventory forecasting suffer from gradient instability and poor adaptability to rapidly changing market conditions, resulting in suboptimal performance in real-time scenarios. This paper presents a novel approach integrating two complementary neural network architectures to address these limitations. The Adaptive Feature Dynamic Adjustment Network (AFDAN) employs real-time parameter adjustment mechanisms for biases and activation functions, enabling rapid response to sudden market events and enhanced environmental adaptability. The Perception Enhanced Recursive Revision Network (PERRN) incorporates a perception enhancement mechanism that sensitively responds to immediate data changes, while its recursive revision strategy optimizes historical information processing to improve prediction. Experimental validation on the Tianchi e-commerce inventory dataset, together with a cross-dataset evaluation on the M5 Forecasting benchmark, demonstrates that the integrated AFDAN-PERRN approach achieves superior performance with a precision of 0.955, mean squared error (MSE) of 0.0495, and replenishment-decision accuracy of 0.984 on the primary dataset, while preserving comparable advantages on the external benchmark. Statistical significance tests over five independent runs further confirm that the observed improvements are not attributable to random variation. The proposed methodology represents a significant advancement in adaptive neural networks for inventory management, offering both theoretical contributions to gradient-stable learning and practical improvements in e-commerce operational efficiency.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dual-network synergy for inventory forecasting: integrating adaptive feature dynamics and recursive perception correction

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月1日

查看全文
  • Jing Long,
  • Yuan Lei

摘要

Traditional LSTM models in e-commerce inventory forecasting suffer from gradient instability and poor adaptability to rapidly changing market conditions, resulting in suboptimal performance in real-time scenarios. This paper presents a novel approach integrating two complementary neural network architectures to address these limitations. The Adaptive Feature Dynamic Adjustment Network (AFDAN) employs real-time parameter adjustment mechanisms for biases and activation functions, enabling rapid response to sudden market events and enhanced environmental adaptability. The Perception Enhanced Recursive Revision Network (PERRN) incorporates a perception enhancement mechanism that sensitively responds to immediate data changes, while its recursive revision strategy optimizes historical information processing to improve prediction. Experimental validation on the Tianchi e-commerce inventory dataset, together with a cross-dataset evaluation on the M5 Forecasting benchmark, demonstrates that the integrated AFDAN-PERRN approach achieves superior performance with a precision of 0.955, mean squared error (MSE) of 0.0495, and replenishment-decision accuracy of 0.984 on the primary dataset, while preserving comparable advantages on the external benchmark. Statistical significance tests over five independent runs further confirm that the observed improvements are not attributable to random variation. The proposed methodology represents a significant advancement in adaptive neural networks for inventory management, offering both theoretical contributions to gradient-stable learning and practical improvements in e-commerce operational efficiency.