ReDiTT: Retrieval-Augmented Diffusion Transformer for Time Series

Predicting asynchronous time series, which involves forecasting the next inter-event time and event type, is challenging due to the inherent uncertainty of future events. Existing models often struggle with long-horizon forecasting and generating diverse, realistic samples.Researchers have introduced ReDiTT, a novel retrieval-augmented conditional diffusion transformer designed to address these limitations. ReDiTT operates in a latent space and enhances its predictions by retrieving structurally similar latent sequences from a memory bank during both training and inference. These retrieved sequences serve as reference conditions, providing global structural guidance through cross-attention mechanisms.This retrieval-based conditioning significantly stabilizes long-horizon forecasting and improves the diversity of generated samples. Experimental results across seven real-world datasets demonstrate that ReDiTT achieves state-of-the-art performance in next event prediction and long-horizon forecasting, making it a powerful tool for various time series applications.