Cross-modal missing time-series imputation using dense spatio-temporal transformer nets.

Area of Science:

• Data Science
• Artificial Intelligence
• Sensor Networks

Background:

• Sensor networks frequently suffer from missing time-series data due to sampling issues or device failures.
• Existing imputation methods struggle with accuracy, particularly for complete data missing (CDM) scenarios.

Purpose of the Study:

• To develop an accurate and robust method for imputing missing time-series data.
• To address the limitations of current imputation techniques, especially in CDM cases.

Main Methods:

• Proposed a novel cross-modal imputation method utilizing dense spatio-temporal transformer networks (DSTTN).
• DSTTN embeds spatial data into time-series data using stacked spatio-temporal transformer blocks and dense connections.
• Employed cross-modal constraints and graph Laplacian regularization for model optimization in an end-to-end pipeline.

Main Results:

• DSTTN demonstrated state-of-the-art imputation performance for both random and non-random missing data.
• The method proved particularly effective in addressing the challenging complete data missing (CDM) problem.
• Experimental comparisons validated DSTTN's superiority over various baseline imputation models.

Conclusions:

• DSTTN offers a significant advancement in time-series data imputation.
• The proposed method provides a viable and effective solution for complete data missing scenarios in sensor networks.