MAGAN: Mode Information and Attention-Based GAN for Realistic Time Series Data Synthesis

The demand for efficient processing and analysis of time series data is growing across multiple fields, yet the diverse acquisition of such data is plagued by issues such as insufficient data volume, poor privacy, and uneven data distribution in related technological research. Time series data generation effectively addresses this issue, with Generative Adversarial Network(GAN) based models showing promising performance among existing methods. Nonetheless, these methods overall performance on fidelity issues(e.g., mode collapse, difficulty capturing long-term dependencies) is not particularly outstanding. In this paper, we propose a GAN framework known as Mode information and Attention-based Generative Adversarial Network(MAGAN) which transforms the metadata and sequential data from real data into mode information and temporal information. We employ a GAN based on a multilayer perceptron (MLP) for mode information, while a hierarchical attention network with attention mechanism for temporal information. In addition to fidelity, we evaluated MAGAN based on usefulness and diversity. Experimental results show that the proposed framework significantly outperforms state-of-the-art benchmarks on three typical real-world datasets.