EMP: Enhanced Multi-modal Prediction for fashion sales using Fourier Mapping and ERP-based contrastive learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis
• E-commerce Analytics

Background:

• Predicting sales for new fashion products is challenging due to a lack of historical data and complex, multi-modal input features.
• Existing forecasting models often depend on external data sources, which can be unreliable or unavailable, limiting their applicability.
• The inherent variability and multi-modal nature of fashion product data necessitate advanced forecasting techniques.

Purpose of the Study:

• To introduce the Enhanced Multi-modal Prediction (EMP) model, a novel framework for zero-shot time series forecasting specifically for fashion products.
• To develop a model that accurately predicts short-term (12-week) sales trajectories for new fashion items without historical sales data.
• To enhance forecasting accuracy by leveraging internal reference item sales and a unique Fourier Mapping technique.

Main Methods:

• The EMP model employs a two-stage architecture: ERP-Aware Contrastive Learning for embedding space construction and a Transformer-based encoder-decoder for sales prediction.
• Edit Distance with Real Penalty (ERP) is utilized to measure time series similarity and retrieve relevant reference items.
• Fourier Mapping is introduced to generate scale-aware embeddings that preserve sales magnitude and capture both low- and high-frequency patterns.

Main Results:

• The EMP model demonstrates superior forecasting accuracy on both private and benchmark datasets compared to existing methods.
• ERP-Aware Contrastive Learning effectively creates a meaningful embedding space for reference item retrieval.
• Fourier Mapping enhances the model's ability to capture sales dynamics and provides theoretical advantages in preserving input variations.

Conclusions:

• The proposed Enhanced Multi-modal Prediction (EMP) model offers an effective solution for zero-shot fashion sales forecasting.
• Leveraging internal reference item sales and Fourier Mapping significantly improves prediction accuracy and reduces reliance on external data.
• The model's ability to handle multi-modal inputs and preserve fine-grained sales variations validates its potential for practical application in the fashion industry.