Abstract
Multivariant time series (MTS) data are usually incomplete in real scenarios, and imputing the incomplete MTS is practically important to facilitate various time series mining tasks. Recently, diffusion model-based MTS imputation methods have achieved promising results by utilizing CNN or attention mechanisms for temporal features learning. However, it is hard to adaptively trade off the diverse effects of local and global temporal features by simply combining CNN and attention. To address this issue, we propose a Score-weighted Convolutional Diffusion Model (Score-CDM), whose backbone consists of a Score-weighted Convolution Module (SCM) and an Adaptive Reception Module (ARM). SCM adopts a score map to capture the global temporal features in the time domain, while ARM uses a Spectral2Time Window Block (S2TWB) to convolve the local time series data in the spectral domain. Benefiting from the time convolution properties of Fast Fourier Transformation, ARM can adaptively change the receptive field of the score map, and thus effectively balance the local and global temporal features. We conduct extensive evaluations on three real MTS datasets of different domains, and the result verifies the effectiveness of the proposed Score-CDM.
Introduction
Multivariate Time Series (MTS) data—such as traffic flow across a city or air quality readings from multiple sensors—is rarely perfect. Sensors fail and networks drop packets, leaving gaps in the data. Imputing these missing values is critical for downstream tasks like forecasting or anomaly detection. The core challenge lies in the "tug-of-war" between local and global temporal features. A traffic spike might be a local, short-term event, while morning rush hour is a global, periodic pattern. Existing models often struggle to balance these: Transformers excel at global dependencies but can overlook local nuances and are computationally expensive ($O(T^2)$), while CNNs capture local patterns well but have fixed receptive fields, making it difficult to capture long-term global trends.
Methods
Score-CDM addresses these challenges through a novel denoising function within a diffusion framework. Its architecture consists of two main components designed to harmonize feature extraction:
- Score-weighted Convolution Module (SCM): This module learns a "score map" that acts as a globally attentive convolution kernel. Unlike standard convolutions with fixed weights, SCM adaptively weights the importance of different time steps across the entire series to capture global correlations.
- Adaptive Reception Module (ARM): To handle local features, the ARM utilizes a Spectral2Time Window Block (S2TWB). Instead of convolving in the time domain, it operates in the spectral (frequency) domain using Fast Fourier Transforms (FFT).
According to the convolution theorem, multiplication in the frequency domain is equivalent to convolution in the time domain. This allows the S2TWB to adaptively change its receptive field and capture periodicities much more flexibly than a standard CNN. This approach also ensures efficiency, with a time complexity of $O(Td + T \log T)$, significantly lower than standard Transformers for long sequences.
Experimental Results
The model was evaluated on three real-world MTS benchmarks: METR-LA (traffic flow), PEMS-BAY (traffic flow), and AQI-36 (air quality), testing both point and block missing scenarios. Score-CDM consistently outperformed all baselines, including specialized diffusion models like CSDI and PriSTI.
- Accuracy: On the AQI-36 dataset, Score-CDM improved the Mean Absolute Error (MAE) by 3% to 5% over the best baseline. On the PEMS-BAY dataset, the improvement was even more pronounced at 9% to 12%.
- Robustness: In extreme cases where 95% of the data was missing, Score-CDM remained stable and accurate, while RNN-based methods like BRITS saw their error rates skyrocket.
- Global-Local Synergy: Ablation studies showed that while the global SCM has a larger impact on performance, both modules are necessary to achieve state-of-the-art results, confirming that MTS imputation requires a holistic view of time.
Conclusion
Score-CDM provides a breakthrough in temporal modeling for MTS imputation by bypassing the rigid limitations of fixed-size CNN kernels through frequency-domain operations. By adaptively balancing local fluctuations and global cycles, it offers a robust and efficient solution for incomplete data. Future work may explore integrating more complex graph-based spatial modeling and applying the S2TWB's flexible receptive field to long-term time series forecasting.
Summary updated