ReDiTT: Retrieval Augmented Conditional Diffusion Transformers for Asynchronous Time Series

2026-07-14 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

ReDiTT, a novel retrieval augmented conditional diffusion transformer, is introduced for asynchronous time series prediction, specifically forecasting the next inter-event time and event type. This model addresses the inherent uncertainty in future events by operating in a latent space. During both training and inference, ReDiTT retrieves structurally similar latent sequences from a memory bank, integrating them as reference conditions through cross-attention mechanisms. This retrieval-based conditioning allows the model to leverage relevant temporal dynamics and provides global structural guidance for generation. Consequently, ReDiTT significantly stabilizes long-horizon forecasting and enhances sample diversity. Experimental results across seven real-world datasets demonstrate state-of-the-art performance for both next event prediction and long-horizon forecasting. The code is publicly available.

Key takeaway

For Machine Learning Engineers developing asynchronous time series models, ReDiTT offers a robust approach to improve prediction accuracy and stability. If you are struggling with long-horizon forecasting or generating diverse event sequences, consider implementing retrieval-augmented conditional diffusion transformers. This method, by utilizing similar historical patterns, can significantly enhance your model's ability to handle inherent event uncertainty and deliver strong performance.

Key insights

ReDiTT enhances asynchronous time series prediction by integrating retrieval-augmented conditioning into a latent-space diffusion transformer.

Principles

• Retrieval augmentation improves temporal dynamics.
• Global structural guidance stabilizes generation.
• Latent space operations manage event uncertainty.

Method

ReDiTT retrieves structurally similar latent sequences from a memory bank, then incorporates them as reference conditions via cross-attention within a conditional diffusion transformer.

In practice

• Apply retrieval for long-horizon time series.
• Use cross-attention for conditional guidance.
• Enhance sample diversity in event prediction.

Topics

• ReDiTT
• Retrieval Augmented Models
• Conditional Diffusion Transformers
• Asynchronous Time Series
• Event Prediction
• Long-Horizon Forecasting

Code references

• BorealisAI/ReDiTT

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Transformer-based Diffusion models for Hydrological Time Series Probabilistic Imputation and Forecasting — Transformer-based diffusion models effectively impute and forecast sparse hydrological time series, outperforming traditional methods in complex pattern capture.
• Semantics-Enhanced Retrieval-Augmented Time Series Forecasting — Time series forecasting improves by combining numerical similarity with semantic context from textual descriptions.
• Learning Recursive Multi-Scale Representations for Irregular Multivariate Time Series Forecasting — ReIMTS forecasts irregular multivariate time series by preserving original timestamps and recursively splitting samples.
• Follow the Latent Roadmap: Navigating Revocable Decoding for Diffusion LLMs with Anchor Tokens — ASRD enhances dLLM decoding by leveraging temporally consistent "anchor tokens" to guide generation and verify candidates in the embedding space.
• From Long News to Accurate Forecast: Importance-Aware Fusion and PRM-Guided Reflection for Time Series Forecasting — Importance-aware compression and PRM-guided retrieval improve news-enhanced time series forecasting by managing context and selection.
• Spectral Retrieval-Augmented Time-Series Forecasting — SpecReTF enhances time series forecasting by integrating spectral analysis and temporal recency into its retrieval mechanism.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.