Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

F-FOMAML: GNN-Enhanced Meta-Learning for Peak Period Demand Forecasting with Proxy Data

Zexing Xu, Linjun Zhang, Sitan Yang, Rasoul Etesami, Hanghang Tong, Huan Zhang, Jiawei Han

TL;DR

This work tackles peak-period demand forecasting under severe data scarcity by combining GNN-based proxy data with a graph-augmented meta-learning framework called F-FOMAML. By extracting task embeddings through a GNN forecaster and modulating a meta-learned model with FiLM layers, the approach achieves rapid adaptation to new peak tasks while leveraging related tasks for improved generalization. Theoretical analysis provides excess-risk bounds that justify the bias-variance trade-off induced by proxy data, and empirical results on vending-machine and JD.com datasets demonstrate substantial MAE improvements over strong baselines, including notable gains over GNN-only benchmarks. The method offers a scalable, domain-agnostic blueprint for data-scarce forecasting in retail and beyond, with potential applications in settings like real-time promotions and cold-start scenarios.

Abstract

Demand prediction is a crucial task for e-commerce and physical retail businesses, especially during high-stake sales events. However, the limited availability of historical data from these peak periods poses a significant challenge for traditional forecasting methods. In this paper, we propose a novel approach that leverages strategically chosen proxy data reflective of potential sales patterns from similar entities during non-peak periods, enriched by features learned from a graph neural networks (GNNs)-based forecasting model, to predict demand during peak events. We formulate the demand prediction as a meta-learning problem and develop the Feature-based First-Order Model-Agnostic Meta-Learning (F-FOMAML) algorithm that leverages proxy data from non-peak periods and GNN-generated relational metadata to learn feature-specific layer parameters, thereby adapting to demand forecasts for peak events. Theoretically, we show that by considering domain similarities through task-specific metadata, our model achieves improved generalization, where the excess risk decreases as the number of training tasks increases. Empirical evaluations on large-scale industrial datasets demonstrate the superiority of our approach. Compared to existing state-of-the-art models, our method demonstrates a notable improvement in demand prediction accuracy, reducing the Mean Absolute Error by 26.24% on an internal vending machine dataset and by 1.04% on the publicly accessible JD.com dataset.

F-FOMAML: GNN-Enhanced Meta-Learning for Peak Period Demand Forecasting with Proxy Data

TL;DR

This work tackles peak-period demand forecasting under severe data scarcity by combining GNN-based proxy data with a graph-augmented meta-learning framework called F-FOMAML. By extracting task embeddings through a GNN forecaster and modulating a meta-learned model with FiLM layers, the approach achieves rapid adaptation to new peak tasks while leveraging related tasks for improved generalization. Theoretical analysis provides excess-risk bounds that justify the bias-variance trade-off induced by proxy data, and empirical results on vending-machine and JD.com datasets demonstrate substantial MAE improvements over strong baselines, including notable gains over GNN-only benchmarks. The method offers a scalable, domain-agnostic blueprint for data-scarce forecasting in retail and beyond, with potential applications in settings like real-time promotions and cold-start scenarios.

Abstract

Demand prediction is a crucial task for e-commerce and physical retail businesses, especially during high-stake sales events. However, the limited availability of historical data from these peak periods poses a significant challenge for traditional forecasting methods. In this paper, we propose a novel approach that leverages strategically chosen proxy data reflective of potential sales patterns from similar entities during non-peak periods, enriched by features learned from a graph neural networks (GNNs)-based forecasting model, to predict demand during peak events. We formulate the demand prediction as a meta-learning problem and develop the Feature-based First-Order Model-Agnostic Meta-Learning (F-FOMAML) algorithm that leverages proxy data from non-peak periods and GNN-generated relational metadata to learn feature-specific layer parameters, thereby adapting to demand forecasts for peak events. Theoretically, we show that by considering domain similarities through task-specific metadata, our model achieves improved generalization, where the excess risk decreases as the number of training tasks increases. Empirical evaluations on large-scale industrial datasets demonstrate the superiority of our approach. Compared to existing state-of-the-art models, our method demonstrates a notable improvement in demand prediction accuracy, reducing the Mean Absolute Error by 26.24% on an internal vending machine dataset and by 1.04% on the publicly accessible JD.com dataset.
Paper Structure (42 sections, 1 theorem, 21 equations, 4 figures, 5 tables, 1 algorithm)

This paper contains 42 sections, 1 theorem, 21 equations, 4 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Prediction with Limited Data
  4. Meta-Learning for Demand Prediction
  5. Meta-Learning Methods for Few-Shot Learning
  6. Graph Neural Networks for Time Series Forecasting
  7. Problem Formulation
  8. Task Definition
  9. Methodology
  10. Proxy Data Selection
  11. Graph Construction for Proxy Data
  12. GNN-enhanced Representation Learning
  13. Input Product Features
  14. Product Embedding Generation via Forecasting
  15. Edge Relationship Determination
  16. ...and 27 more sections

Key Result

Theorem 1

Consider the data generative model, the algorithm $\widehat{g}_{\widetilde{t}}$, and the assumptions above. Suppose we have $n_d\gtrsim n$ for all $t\in\mathcal{D}^{tr}$. Define the excess risk for the test domain $\widetilde{t}$ by $R(\widehat{g}_{\widetilde{t}})=\mathop{\mathrm{\mathbb{E}}}\limits In particular, if $h$ is properly chosen such that $h\asymp (\frac{C(\mathcal{H})/n}{T})^{\frac{1}{

Figures (4)

  • Figure 1: Pipeline of the GNN-enhanced F-FOMAML for demand forecasting.
  • Figure 2: Evaluation Performance (MSE values) throughout training epochs over MAML, MLP and our proposed methods, where $k$ is set to be $5$.
  • Figure 3: Evaluation of metrics versus different values of k for the k-shot proxy data selection.
  • Figure 4: Training performance (MSE values) throughout training epochs over MAML, MLP and our proposed methods, where $k$ is set to be $5$.

Theorems & Definitions (3)

  • Remark 1
  • Theorem 1
  • proof