InstructTime++: Time Series Classification with Multimodal Language Modeling via Implicit Feature Enhancement
Researchers have introduced InstructTime++, a novel framework for time series classification that reframes the task as a multimodal generative process. This approach treats numerical sequences, textual context, and instructions as inputs, generating class labels as textual outputs via language models. To enhance cross-modal alignment, InstructTime++ incorporates a discretization module for temporal tokens, an alignment projection layer, and a generative self-supervised pre-training strategy. The enhanced version, InstructTime++, further integrates implicit feature modeling by mining statistical patterns and using vision-language models for textual descriptions, leading to superior performance on benchmark datasets. AI
IMPACT This research could improve the accuracy and interpretability of time series classification models by leveraging multimodal language modeling.