MoTIF Uses MoE Attention for Interpretable Irregular Time Series

MoTIF decomposes irregular sequential data into temporal patterns using Mixture-of-Experts attention. On synthetic data, it achieves pattern recovery rates of 0.829 for cyclical patterns, significantly outperforming standard Transformers and unspecialized MoE models.