Scholarlink

<p>Forecasting demand for new fashion items poses a persistent challenge in retail analytics, primarily due to the unavailability of historical sales data. This study presents a hybrid deep learning architecture tailored to address the cold-start problem by combining visual similarity analysis with correction techniques for censored sales data. The proposed approach utilizes multimodal inputs, including high-resolution product images and structured metadata (e.g., category, fabric, and price), in addition to sales and inventory records that are affected by stockout-related censoring. The architecture is composed of two integrated pipelines. The first applies the Expectation–Maximization (EM) algorithm to adjust sales records, estimating true latent demand by accounting for censored entries. The second pipeline extracts deep visual features through a modified ResNet-50 architecture, followed by compression via an autoencoder to generate compact, semantically meaningful representations. These representations are clustered to identify visually similar products, enabling demand estimation through similarity-weighted averaging and localized regression within each cluster. Prediction uncertainty is captured using a bootstrap-based resampling technique. This work offers several notable contributions: the fusion of EM-based demand correction with deep visual learning, the introduction of an adaptive similarity-based forecasting strategy, and the integration of uncertainty estimation for cold-start scenarios. Empirical evaluation on a real-world multimodal fashion dataset demonstrates that the proposed model significantly outperforms several baseline approaches—including those relying solely on metadata, visual features, or global regressors—with results achieving RMSE of 10.95, MAE of 8.54, MAPE of 16.5%, Precision@10 of 73.9%, and Recall@10 of 70.7%. These outcomes underscore the effectiveness and practical value of the architecture in predicting demand for newly launched fashion products.</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

Deep learning for new fashion product demand prediction: integrating visual similarity and demand correction in cold-start scenarios

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月8日

查看全文
  • Mona Adel,
  • Amgad M. Mohammed,
  • Asmaa H. Elsaid,
  • Amira Abdelatey

摘要

Forecasting demand for new fashion items poses a persistent challenge in retail analytics, primarily due to the unavailability of historical sales data. This study presents a hybrid deep learning architecture tailored to address the cold-start problem by combining visual similarity analysis with correction techniques for censored sales data. The proposed approach utilizes multimodal inputs, including high-resolution product images and structured metadata (e.g., category, fabric, and price), in addition to sales and inventory records that are affected by stockout-related censoring. The architecture is composed of two integrated pipelines. The first applies the Expectation–Maximization (EM) algorithm to adjust sales records, estimating true latent demand by accounting for censored entries. The second pipeline extracts deep visual features through a modified ResNet-50 architecture, followed by compression via an autoencoder to generate compact, semantically meaningful representations. These representations are clustered to identify visually similar products, enabling demand estimation through similarity-weighted averaging and localized regression within each cluster. Prediction uncertainty is captured using a bootstrap-based resampling technique. This work offers several notable contributions: the fusion of EM-based demand correction with deep visual learning, the introduction of an adaptive similarity-based forecasting strategy, and the integration of uncertainty estimation for cold-start scenarios. Empirical evaluation on a real-world multimodal fashion dataset demonstrates that the proposed model significantly outperforms several baseline approaches—including those relying solely on metadata, visual features, or global regressors—with results achieving RMSE of 10.95, MAE of 8.54, MAPE of 16.5%, Precision@10 of 73.9%, and Recall@10 of 70.7%. These outcomes underscore the effectiveness and practical value of the architecture in predicting demand for newly launched fashion products.