S2TGAN: A GAN-Based Mobile Video Streaming Generation Model for Stalling Detection

Due to the complex and dynamic nature of the wireless network environment, data-driven model often requires a significant amount and diversity of training data, leading to a challenging difficulty of stalling detection with high performance, especially at the client side. To address this issue, a novel Generative Adversarial Network (GAN) based data generation model Seq2Seq-TCN GAN (S2TGAN) is proposed in this paper for both chunk-based and stream-based video streaming sequences in order to further improve the performance of stalling detection of video on demand (VoD) and live video streaming (LVS). Resorting to Seq2Seq, the generator is constructed to generate video streaming sequences under the control of stalling information, and the discriminator adopts Temporal Convolutional Network (TCN) to extract more discriminative temporal features of video streaming sequences. Meanwhile, Iman-Conover method is used to incorporate domain knowledge to guide the generation of video streaming and improve the fidelity of synthetic data. A LVS dataset collected from HuYa and DouYu is carefully prepared to verify the generation performance of proposed model, as well as a public RequetDataSet. Extensive experimental results indicate that the proposed model can outperform the other state-of-the-arts in terms of maximum mean discrepancy (MMD) and Train on synthetic, Test on Real (TSTR) evaluation methods of stalling detection.