Towards a Unified Method for Network Dynamic via Adversarial Weighted Link Prediction

TL;DR

The paper tackles the problem of dynamic network control by reframing network dynamics as weighted temporal link prediction over graph snapshots represented by adjacency matrices $A$. It introduces HQ-TLP, an adversarial framework that combines GANs with graph convolutional networks and GRUs to generate high-quality next-step weighted snapshots from history, addressing the challenges of wide value ranges and sparsity in edge weights. The approach is demonstrated on three network scenarios with six baselines, reporting improvements in RMSE, EW-KL, and MR, and highlights its potential for enabling precise, prediction-driven network control in data centers, wireless mesh, and mobile ad-hoc networks. It also outlines future work on dynamic node sets, integration with coarse control methods, theoretical error bounds, and adaptive snapshot sampling to further enhance applicability and robustness.

Abstract

Network dynamic (e.g., traffic burst in data center networks and channel fading in cellular WiFi networks) has a great impact on the performance of communication networks (e.g., throughput, capacity, delay, and jitter). This article proposes a unified prediction-based method to handle the dynamic of various network systems. From the view of graph deep learning, I generally formulate the dynamic prediction of networks as a temporal link prediction task and analyze the possible challenges of the prediction of weighted networks, where link weights have the wide-value-range and sparsity issues. Inspired by the high-resolution video frame prediction with generative adversarial network (GAN), I try to adopt adversarial learning to generate high-quality predicted snapshots for network dynamic, which is expected to support the precise and fine-grained network control. A novel high-quality temporal link prediction (HQ-TLP) model with GAN is then developed to illustrate the potential of my basic idea. Extensive experiments for various application scenarios further demonstrate the powerful capability of HQ-TLP.

Figures (5)

• Figure 1: Illustration of the unified strategy to model system behaviors in a time step by constructing a weighted graph snapshot.
• Figure 2: Illustration of VFP and TLP, as well as the comparison between (a) low-quality prediction via error reconstruction and (b) high-quality prediction via adversarial learning, with the figure adapted from qin2023highqin2023temporal.
• Figure 3: The architecture of HQ-TLP, with a generative network $G$ and a discriminative network $D$.
• Figure 4: Algorithm to train HQ-TLP and generate the prediction result with traffic prediction of DCN as an example.
• Figure 5: Evaluation experiment results with ARMSE, AEW-KL, and AMR of the last 50 snapshots on test sets.