Introduction
Irregular multivariate time series (IMTS) are a common challenge in real-world sensor networks, such as ICU monitoring and weather stations. These datasets are characterized by uneven sampling intervals and asynchronous recording across different variables. Historically, researchers used Canonical Pre-Alignment (CPA) to align all variables onto a shared temporal grid via zero-padding. While effective for alignment, CPA causes a "length explosion" that leads to prohibitive memory and compute costs, prompting many modern models to abandon it. This paper introduces KAFNet, an architecture that revitalizes CPA by using smart compression and frequency-domain attention to achieve state-of-the-art performance with high efficiency.
The KAFNet Architecture
KAFNet is designed to maintain the structural benefits of CPA while neutralizing its computational bloat. The model processes data through a specialized pipeline that includes pre-convolution for smoothing and Temporal Kernel Aggregation (TKA) to compress the sequence.
The architecture's core innovations include:
- Temporal Kernel Aggregation (TKA): This module places learnable Gaussian kernels across a normalized timeline to pool the bloated CPA sequence into fixed temporal "buckets." This drastically reduces sequence length while preserving irregular timing information.
- Frequency Linear Attention (FLA): By moving inter-variate attention to the frequency domain via FFT, KAFNet captures global correlations with linear $O(N)$ complexity, avoiding the quadratic costs of standard attention mechanisms.
Experimental Results
KAFNet was evaluated on four standard IMTS benchmarks: PhysioNet, MIMIC, Human Activity, and USHCN. The results demonstrate that KAFNet consistently outperforms a wide range of baselines, including ODE-based models and recent graph neural networks like tPatchGNN.
Key performance metrics highlight the model's efficiency:
- Accuracy: Achieved state-of-the-art Mean Squared Error (MSE) and Mean Absolute Error (MAE) across all datasets.
- Parameter Efficiency: Required 7.2x fewer parameters than leading graph-based alternatives.
- Speed: Achieved an 8.4x acceleration in both training and inference times.
- Compute: Operated at a fraction of the FLOPs required by competitors, making it highly suitable for resource-constrained deployments.
Conclusion
KAFNet proves that the abandonment of Canonical Pre-Alignment was premature. By engineering a way to handle CPA's length explosion through Gaussian kernel aggregation and frequency-domain processing, KAFNet provides a highly efficient and accurate framework for irregular time series forecasting. Future work will expand the model's capabilities to tasks such as classification, anomaly detection, and missing value interpolation.
Summary updated