Hallucination Detection and Mitigation with Diffusion in Multi-Variate Time-Series Foundation Models
This addresses the problem of hallucination in multi-variate time-series foundation models for improving their adoption and safe usage, representing an incremental advancement by adapting concepts from NLP to a new domain.
The paper tackles the lack of hallucination definitions and methods for multi-variate time-series foundation models by proposing new definitions and detection/mitigation techniques using diffusion models, finding that existing models hallucinate up to 59.5% as much as a baseline and reducing this by up to 47.7% with their method.
Foundation models for natural language processing have many coherent definitions of hallucination and methods for its detection and mitigation. However, analogous definitions and methods do not exist for multi-variate time-series (MVTS) foundation models. We propose new definitions for MVTS hallucination, along with new detection and mitigation methods using a diffusion model to estimate hallucination levels. We derive relational datasets from popular time-series datasets to benchmark these relational hallucination levels. Using these definitions and models, we find that open-source pre-trained MVTS imputation foundation models relationally hallucinate on average up to 59.5% as much as a weak baseline. The proposed mitigation method reduces this by up to 47.7% for these models. The definition and methods may improve adoption and safe usage of MVTS foundation models.