M-SAITS: A Dual-Stage Time Series Imputation Network via Decoupled Large-Kernel Convolution and Diagonally-Masked Attention

Missing value imputation in multivariate time series is a critical challenge in the field of data mining. Although Transformer-based methods excel in modeling long-range dependencies, their inherent point-wise attention mechanisms often lack explicit modeling of local inductive biases in time series, making it difficult to effectively capture local smoothness and evolutionary trends. Furthermore, existing feature embedding strategies struggle to fully decouple the internal temporal evolution of variables from complex cross-variable dependencies. To address these limitations, this paper proposes a novel dual-stage imputation framework named M-SAITS. This framework innovatively introduces a decoupled feature encoder based on large-kernel depthwise convolutions. By utilizing an extended effective receptive field, it explicitly enhances the model's perception of local trends. Additionally, it employs a grouped convolution structure to achieve decoupled modeling of intra-variable temporal patterns and inter-variable interaction features. On this basis, combined with a Diagonally-Masked Self-Attention mechanism, the framework physically blocks information leakage paths while achieving lossless global context aggregation. Relying on a "Preliminary Inference–Iterative Refinement" cascade strategy and a masked weighted joint optimization objective, the model achieves high-fidelity data reconstruction. Extensive experiments on multiple benchmark datasets, such as Electricity and Air Quality, demonstrate that this method significantly outperforms existing state-of-the-art models across multiple evaluation metrics. Notably, in high-dimensional electricity data imputation tasks, M-SAITS achieves substantial performance improvements over baseline models such as CSDI and Transformer, with the Mean Absolute Error significantly reduced (up to approximately 60% under low missing rates).