QiYao-I
A manifold foundation model for irregular multivariate time series forecasting. It employs a sampling-conditioned temporal manifold attention mechanism and frequency-aware dynamic variable interaction to handle irregularly sampled and asynchronously recorded time series data, supporting zero-shot and few-shot transfer.
Timeline (last 90 days)
QiYao-I is a manifold-based foundation model for irregular multivariate time series forecasting.
QiYao-I proposes a sampling-conditioned temporal manifold attention mechanism that maps real timestamps to a learnable temporal manifold feature space and injects temporal manifold biases into the attention layers.
QiYao-I's temporal manifold attention mechanism is capable of simultaneously capturing both irregular time intervals and local sampling structure.
QiYao-I proposes a frequency-aware dynamic variable interaction mechanism that selectively performs cross-variable message passing to avoid spurious cross-variable correlations when synchronized observations are absent.
In experiments on multiple real-world irregular multivariate forecasting benchmarks, QiYao-I achieves superior performance compared to both time series foundation models and end-to-end irregular forecasting models.
QiYao-I demonstrates strong generalization capability under both zero-shot and few-shot settings.
As a newly released preprint, QiYao-I still requires subsequent validation regarding larger-scale deployment, computational cost, and robustness across different domains.
All Facts (7)
QiYao-I is a manifold-based foundation model for irregular multivariate time series forecasting.
50%UnverifiedQiYao-I proposes a sampling-conditioned temporal manifold attention mechanism that maps real timestamps to a learnable temporal manifold feature space and injects temporal manifold biases into the attention layers.
50%UnverifiedQiYao-I's temporal manifold attention mechanism is capable of simultaneously capturing both irregular time intervals and local sampling structure.
50%UnverifiedQiYao-I proposes a frequency-aware dynamic variable interaction mechanism that selectively performs cross-variable message passing to avoid spurious cross-variable correlations when synchronized observations are absent.
50%UnverifiedIn experiments on multiple real-world irregular multivariate forecasting benchmarks, QiYao-I achieves superior performance compared to both time series foundation models and end-to-end irregular forecasting models.
50%UnverifiedQiYao-I demonstrates strong generalization capability under both zero-shot and few-shot settings.
50%UnverifiedAs a newly released preprint, QiYao-I still requires subsequent validation regarding larger-scale deployment, computational cost, and robustness across different domains.
50%