RAID: Semantic Graph Diffusion for True Cold-Start... | AI Research

RAID: Semantic Graph Diffusion for True Cold-Start and Cross-Lingual Forecasting
Time-series forecasting models often struggle when they encounter a "cold-start" scenario—a situation where a new item, such as a newly launched product, has no historical sales data to analyze. While modern foundation models are powerful, they typically require a history window to make predictions, causing them to fail when that data is missing. This paper introduces RAID (Retrieval-Augmented Iterative Diffusion), a framework designed to solve this by replacing history-based learning with metadata-driven insights. Instead of relying on past observations, RAID uses textual descriptions to understand an item's potential demand, allowing it to make accurate forecasts for items that have never been seen before.

How RAID Works

RAID operates by mapping textual metadata—such as product titles, categories, and descriptions—into a shared "semantic space" using a pre-trained multilingual model. By doing this, the system can identify relationships between items based on their descriptions rather than their sales history. For example, if a new running shoe is introduced, the model identifies similar existing shoes and uses their demand patterns as a guide. It constructs an "inductive retrieval graph" that connects these items, allowing the model to propagate demand signals from established products to new ones.

Generating Forecasts

The forecasting process in RAID is split into two main phases. First, it creates a "base forecast" by aggregating the normalized patterns of semantically similar neighbors. This provides a solid starting point that avoids the common pitfall of overestimating or underestimating demand for niche items. Second, it uses a "Gated Diffusion Transformer" to refine this base prediction. This module models the residual uncertainty—the difference between the base prediction and the actual outcome—to produce a final, probabilistic forecast. By focusing on this residual, the model becomes more efficient, requiring significantly fewer refinement steps than traditional diffusion models.

Efficiency and Performance

A major advantage of the RAID framework is its speed. Because it decouples the memory-intensive retrieval process from the generative diffusion process, it achieves much lower inference latency compared to standard transformer-based models. In testing, RAID outperformed existing foundation models in strict cold-start scenarios, where it was able to provide accurate predictions despite having zero historical data. Furthermore, because the model uses a multilingual embedding space, it demonstrates strong "zero-shot" cross-lingual transfer. This means a model trained on English product descriptions can effectively forecast demand for items described in other languages, such as Spanish or Russian, without needing additional training data in those languages.

Key Takeaways

RAID represents a shift in how we approach forecasting for new items. By grounding predictions in static metadata rather than temporal history, it overcomes the "cold-start" failure mode that plagues many current foundation models. The framework’s ability to maintain high accuracy while reducing computational costs makes it a practical solution for real-world applications where new items are frequently added to a catalog. While traditional models remain strong when historical data is abundant, RAID provides a robust, scalable alternative for scenarios where that data is entirely absent.