Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Leveraging World Events to Predict E-Commerce Consumer Demand under Anomaly

Dan Kalifa, Uriel Singer, Ido Guy, Guy D. Rosin, Kira Radinsky

TL;DR

The paper tackles e-commerce demand forecasting during anomalies by incorporating world-event information through day-level embeddings learned with a transformer-based adversarial encoder. It introduces GAN-Event, which builds embeddings from the relational structure of events using Wikipedia2Vec representations and a Hausdorff distance-based reconstruction objective, then feeds these embeddings into an LSTM forecaster to predict future sales over a 30-day horizon. Empirical results across five eBay categories show GAN-Event LSTM consistently outperforming ARIMA, Prophet, Neural Prophet, and Event Neural Prophet, with the largest gains in highly anomalous periods such as 2020, and ablation studies confirm the importance of event associations and the Hausdorff-based reconstruction. The work demonstrates the practical value of leveraging externally sourced world-event semantics for robust forecasting in volatile conditions and provides a reproducible codebase and dataset combination for further research.

Abstract

Consumer demand forecasting is of high importance for many e-commerce applications, including supply chain optimization, advertisement placement, and delivery speed optimization. However, reliable time series sales forecasting for e-commerce is difficult, especially during periods with many anomalies, as can often happen during pandemics, abnormal weather, or sports events. Although many time series algorithms have been applied to the task, prediction during anomalies still remains a challenge. In this work, we hypothesize that leveraging external knowledge found in world events can help overcome the challenge of prediction under anomalies. We mine a large repository of 40 years of world events and their textual representations. Further, we present a novel methodology based on transformers to construct an embedding of a day based on the relations of the day's events. Those embeddings are then used to forecast future consumer behavior. We empirically evaluate the methods over a large e-commerce products sales dataset, extracted from eBay, one of the world's largest online marketplaces. We show over numerous categories that our method outperforms state-of-the-art baselines during anomalies.

Leveraging World Events to Predict E-Commerce Consumer Demand under Anomaly

TL;DR

The paper tackles e-commerce demand forecasting during anomalies by incorporating world-event information through day-level embeddings learned with a transformer-based adversarial encoder. It introduces GAN-Event, which builds embeddings from the relational structure of events using Wikipedia2Vec representations and a Hausdorff distance-based reconstruction objective, then feeds these embeddings into an LSTM forecaster to predict future sales over a 30-day horizon. Empirical results across five eBay categories show GAN-Event LSTM consistently outperforming ARIMA, Prophet, Neural Prophet, and Event Neural Prophet, with the largest gains in highly anomalous periods such as 2020, and ablation studies confirm the importance of event associations and the Hausdorff-based reconstruction. The work demonstrates the practical value of leveraging externally sourced world-event semantics for robust forecasting in volatile conditions and provides a reproducible codebase and dataset combination for further research.

Abstract

Consumer demand forecasting is of high importance for many e-commerce applications, including supply chain optimization, advertisement placement, and delivery speed optimization. However, reliable time series sales forecasting for e-commerce is difficult, especially during periods with many anomalies, as can often happen during pandemics, abnormal weather, or sports events. Although many time series algorithms have been applied to the task, prediction during anomalies still remains a challenge. In this work, we hypothesize that leveraging external knowledge found in world events can help overcome the challenge of prediction under anomalies. We mine a large repository of 40 years of world events and their textual representations. Further, we present a novel methodology based on transformers to construct an embedding of a day based on the relations of the day's events. Those embeddings are then used to forecast future consumer behavior. We empirically evaluate the methods over a large e-commerce products sales dataset, extracted from eBay, one of the world's largest online marketplaces. We show over numerous categories that our method outperforms state-of-the-art baselines during anomalies.
Paper Structure (33 sections, 7 equations, 2 figures, 6 tables)

This paper contains 33 sections, 7 equations, 2 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Time series forecasting
  4. Sales forecasting
  5. Prediction under anomalies
  6. Leveraging world events
  7. Proposed Method
  8. GAN-Event Generator
  9. Hausdorff Loss
  10. GAN-Event Discriminator
  11. GAN-Event LSTM
  12. Experimental Setup
  13. Datasets
  14. World-event dataset
  15. E-commerce dataset
  16. ...and 18 more sections

Figures (2)

  • Figure 1: Sales of the Football Cards category.
  • Figure 2: The GAN-Event architecture. (a) Given a set of a real day's events and their corresponding embedding vectors, the generator $G$ masks k% of the events randomly, and attempts to reconstruct the masked events based on the unmasked events. (b) The discriminator $D$ learns the strength of association between the day's events. It receives a set of event embeddings and predicts whether the set of events is from a real day or not. (c) Given a set of a real day's events, GAN-Event uses $G$ to reconstruct some of them (the events $G$ masked) and then executes $D$ to distinguish between the real day and $G$'s output. We use Hausdorff loss as an additional regulator that minimizes the distance between the reconstructed and original events.

Theorems & Definitions (1)

  • definition 1: Hausdorff Distance