Self-supervised generative learning for sequential data prediction

Ke Xu

doi:10.1007/s10489-023-04578-5

Self-supervised generative learning for sequential data prediction

2023-04-20

4

PRE

AI

DOI：10.1007/s10489-023-04578-5

Original

Original request for help

Save

Abstract

En 中文

For many real world applications, such as stock price prediction and video frame synthesis, sequential data prediction is a fundamental and challenging task. Considering the temporal features of sequential data, existing approaches generally adopt recurrent neural network and its variants for the prediction. However, for sequences with complex structure, these approaches cannot guarantee to obtain promising results. In this paper, to address the above issue, we formulate sequential data prediction as a self-supervised generative learning problem. Concretely, we design a generator to learn the distribution of the sequential data and generate the predicted values, as well as a discriminator to judge whether or not the input sequential data are real or fake. Based on this proposed framework and the adversarial learning scheme, we develop the corresponding networks for vector inputs and high-order tensor inputs, respectively, which are respectively named vector generative network (VGN) and high-order tensor generative network (HTGN). Extensive experiments on five stock price prediction datasets and two video frame prediction datasets demonstrate the effectiveness of our framework, and its advantages over the state-of-the-art approaches. Our main code and the used data have been shared at https://github.com/xsavagek/SSGL.

Keywords:

Sequential data prediction

Self-supervised learning

Generative learning

Vector Generative Network (VGN)

High-order Tensor Generative Network (HTGN)

Self-supervised generative learning for sequential data prediction

Abstract

Journal

Organization