Title: QuITE: Query-Based Irregular Time Series Embedding

Original: arXiv:2605.28166v3 Announce Type: replace-cross

Abstract: Although Irregular Multivariate Time Series (IMTS) are prevalent in real-world scenarios, their non-uniform sampling intervals pose significant challenges for effective modeling. Current methodologies generally fall into two categories: (i) constructing bespoke architectures that hinder the application of established Multivariate Time Series (MTS) models, or (ii) converting IMTS into regular temporal grids via interpolation, a process that can distort temporal dynamics by injecting synthetic data points. To overcome these drawbacks, we present a novel input-embedding strategy. We posit that the primary constraint is not the backbone model itself, but rather standard embedding layers that presuppose uniform sampling. Consequently, we introduce QuITE (Query-Based Irregular Time Series Embedding), a straightforward yet potent plug-and-play embedding module designed specifically for IMTS. By utilizing learnable query tokens within a single self-attention layer, QuITE aggregates irregular observations to generate latent representations compatible with existing backbones, thereby eliminating the need for artificial value creation or architectural alterations. Comprehensive evaluations on real-world benchmarks demonstrate that QuITE consistently enhances the performance of MTS models, achieving average relative improvements of up to $54.7\%$ in forecasting tasks and $15.8\%$ in classification across various datasets and backbone structures. The code is accessible at: https://github.com/Meaningfull9502/QuITE.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC