ConceptTS: LLM-Guided Concept Bottlenecks for Interpretable Multivariate Time-Series Forecasting
ConceptTS turns LLM-proposed concepts into supervised bottlenecks for time-series forecasts.
The framework has the language model propose task-relevant concepts and executable labeling rules, avoiding manual concept annotation. It separates those concepts across historical context, local forecast intervals, and the full forecast horizon. A shared decoder builds the forecast from predicted concept activations, allowing concept-level interventions. On the Beijing Multi-Site Air Quality dataset, the authors report accuracy competitive with strong black-box baselines and semantically meaningful activations. ArXiv · AI/CL/LG's note
The framework has the language model propose task-relevant concepts and executable labeling rules, avoiding manual concept annotation. It separates those concepts across historical context, local forecast intervals, and the full forecast horizon. A shared decoder builds the forecast from predicted concept activations, allowing concept-level interventions. On the Beijing Multi-Site Air Quality dataset, the authors report accuracy competitive with strong black-box baselines and semantically meaningful activations. ArXiv · AI/CL/LG's note
score 4