Heterogeneity-Aware Active Federated Learning for Demand Forecasting in Online-Retail Supply Chains (HFWP)

Online-retail catalogs are deeply heterogeneous: a perishable item such as fresh herbs and a durable, promotion-driven item such as wireless earbuds differ in nature, category, price, life-cycle, and event sensitivity. A common objection to unified demand-forecasting systems is that one model cannot serve such different products. We argue that heterogeneity should not be assumed away but encoded as a structured conditioning input to a single shared model. We present HFWP, a privacy-preserving Active Federated Learning framework for demand forecasting across an online-retail supply chain. Each order is mapped to a unified representation of product attributes, temporal signals, region, and quantified event context, including localized Iranian-calendar events; a shared backbone with conditioning heads then produces item-specific forecasts. Retailers collaborate without sharing raw data through selective parameter encryption, dual ElGamal encryption, and homomorphic aggregation, while active learning concentrates training on the hardest heterogeneous cases. On a Superstore benchmark with ۵۰ clients, HFWP reaches ۹۹.۳% accuracy, close to a non-private centralized model (۹۹.۹%) and about ۱۵ points above local-only models (۸۴.۶%), and accelerates convergence up to ۲۷.۶ times. Heterogeneity is thus an input, not a barrier.

کلیدواژه ها:

federated learning ، demand forecasting ، supply chain ، product heterogeneity ، active learning ، homomorphic encryption ، privacy preservation ، online retail