MetaSTNet: Multimodal Meta-learning for Cellular Traffic Conformal Prediction

TL;DR

MetaSTNet tackles cellular traffic forecasting under limited real-world data by leveraging sim-to-real transfer within a multimodal, spatiotemporal meta-learning framework. It employs bi-level optimization to learn meta-knowledge from simulation and quickly adapt to new tasks in the real world, while two parallel encoder–decoder streams capture short-term and periodic patterns through event-driven attention and ST-blocks that fuse traffic, textual, and image data. The approach is complemented by cross conformal prediction with a growing-window calibration to provide calibrated prediction intervals, demonstrated across Milano, Trento, and LTE datasets. The results show improved point predictions and reliable uncertainty estimates, highlighting the potential for accurate, trustworthy network-traffic forecasts when data are scarce and for applications requiring quantified confidence in predictions.

Abstract

Network traffic prediction techniques have attracted much attention since they are valuable for network congestion control and user experience improvement. While existing prediction techniques can achieve favorable performance when there is sufficient training data, it remains a great challenge to make accurate predictions when only a small amount of training data is available. To tackle this problem, we propose a deep learning model, entitled MetaSTNet, based on a multimodal meta-learning framework. It is an end-to-end network architecture that trains the model in a simulator and transfers the meta-knowledge to a real-world environment, which can quickly adapt and obtain accurate predictions on a new task with only a small amount of real-world training data. In addition, we further employ cross conformal prediction to assess the calibrated prediction intervals. Extensive experiments have been conducted on real-world datasets to illustrate the efficiency and effectiveness of MetaSTNet.

Figures (7)

• Figure 1: Splitting training samples from the target task to generate training and calibrating sets.
• Figure 2: The structure of the proposed MetaSTNet.
• Figure 3: The interval prediction results on the Milano dataset. Among four subplots, (a) and (b) display 1-hour ahead interval prediction results with metrics of coverage rate and average width length, respectively. In contrast, (c) and (d) show the 1-day ahead interval prediction results with metrics of coverage rate and average width length, respectively.
• Figure 4: The interval prediction results on the Trento dataset. Among four subplots, (a) and (b) display 1-hour ahead interval prediction results with metrics of coverage rate and average width length, respectively. In contrast, (c) and (d) show the 1-day ahead interval prediction results with metrics of coverage rate and average width length, respectively.
• Figure 5: The experimental results of 1-hour ahead interval prediction on the Milano dataset. The black line denotes ground truths and the red line represents predictions. The gradually deepening blue colors represent the prediction intervals under the confidence level of 95%, 85%, and 75%, respectively.