Beyond Multimodal Fusion: Graph-Smoothed Latent Embeddings and Dynamic Gradient Boosting for E-Commerce Product Price Prediction

Abstract

Predicting reliable transaction benchmarks over complex real-world digital market inventories remains difficult due to multi-modal data discrepancies and severe target heteroscedasticity. This paper presents an integrated system framework engineered to compute precise catalog item costs by simultaneously ingest descriptive textual documentation and corresponding merchandise iconography. The system first projects visual elements and dynamic data text fields into shared high-dimensional vector spaces using a deep dual-modal architectural backbone. To eliminate non-aligned information anomalies and feature dispersion patterns, properties are regularized through a multi-layer graph convolutional network execution scheme over symmetric nearest neighbor network frameworks. Target value scales are adjusted to normal tracking variations through specialized logarithmic mappings to bound relative variances. High-dimensional graph representations are combined side-by-side with localized multimodal vectors to construct robust unified properties arrays. Final numerical target metrics are generated through out-offold multi-seed gradient-boosting ensemble regression models. Empirical validations conducted across large-scale market data arrays containing more than one hundred and fifty thousand individual samples demonstrate a symmetric mean absolute percentage error (SMAPE) profile of 53% without any negative value prediction abnormalities.