Perceptron-based Parallel Compression Method with High Compression Ratio under Point-wise Relative Error Bound

As the amount of global data is growing exponentially, lossy compression technology has become an inevitable trend in artificial intelligence and big data applications. However, existing compression algorithms are difficult to further improve the compression ratio while ensuring compression performance. Therefore, this study proposes a data prediction compression framework based on a lightweight perceptron model. Our work mainly includes three aspects of innovation: (1) introducing positive and negative factors to achieve effective conversion from relative domain to absolute domain and eliminate additional symbol storage;(2) building a lightweight four-layer perceptron network to improve prediction accuracy;(3) modifying the data selection rules of the prediction surface to achieve parallel compression within the compression block and improve the compression speed. In the comparative experiment with SZ2.1's PW\_REL, our method achieved a compression ratio improvement of up to 17.78% on multiple test sets.