Abstract
Effective resource management in wireless networks requires both accurate traffic classification and precise future traffic prediction. Traditionally, these tasks are handled independently, leading to statistical interference and reliance on increasingly unreliable prior features. This paper proposes a joint machine learning framework that establishes a dual-feedback loop between classification and prediction. By using classification results to deploy specialized predictors and utilizing prediction errors as a posteriori features for classification, the framework significantly improves performance in both domains. Evaluations using real-world campus network traffic show a substantial reduction in Root Mean Square Error (RMSE) and improved classification robustness.
Introduction
In the evolving landscape of wireless communications, maintaining high Quality of Service (QoS) necessitates sophisticated network management. Network controllers must identify the type of traffic (classification) and forecast its volume (prediction) to allocate resources dynamically. However, traditional approaches treat these as isolated problems. This isolation results in "one-size-fits-all" predictors that struggle with the diverse statistical signatures of different traffic types, such as the steady flow of voice calls versus the bursty nature of gaming. Furthermore, the increasing use of encryption and dynamic IP assignment renders traditional classification features like port numbers less effective.
Methodology
The proposed framework introduces a dual-feedback mechanism to bridge the gap between classification and prediction, creating a synergistic relationship between the two tasks.
Dual-Feedback Mechanism
- Classification-Assisted Prediction: Instead of a single general model, the system employs a "predictor group" consisting of specialized models for Voice (VO), Video (VI), and Gaming (GM). The classifier first identifies the traffic category, which then determines which specialized predictor is activated. This ensures the model is not "confused" by irrelevant traffic patterns.
- Prediction-Assisted Classification: This novel feedback loop uses the prediction error of all available predictors as a "posteriori feature." If a specific predictor (e.g., Gaming) yields a significantly lower error than others for a given stream, this signal is fed back to the classifier to refine the traffic type identification.
The final classification decision is governed by a weighted distance formula: $$D_a = D_p + \alpha D_t$$ where $D_t$ represents the traditional classification distance, $D_p$ is the prediction error vector, and $\alpha$ is a tunable weight balancing the two components.
Dataset and Experimental Setup
The researchers validated the model using real-world data from a campus network, focusing on three business types across several popular applications: WeChat and QQ voice (VO); Douyu Live, Tencent Conference, and Kuaishou live (VI); and Honor of Kings and PUBG (GM). Data was collected in 5-minute intervals, with the first 2 minutes dedicated to training and the final 2 minutes used for testing.
Results and Discussion
The joint method (Scenario A) was compared against a traditional general predictor (Scenario C) and a scenario involving incorrect classification (Scenario B).
Performance Gains
The experimental results highlight the superiority of the specialized joint approach. As shown in the figure below, Scenario A consistently achieved the lowest RMSE across various machine learning algorithms, including BP Neural Networks, GWO-ELM (Grey Wolf Optimized Extreme Learning Machine), and Random Forests.
Specifically, Scenario A reached an RMSE of approximately 0.15 (roughly 85% accuracy), a significant improvement over the general predictor's RMSE of 0.33. Conversely, Scenario B (incorrect classification) resulted in an RMSE exceeding 1.0, underscoring the critical role of accurate classification in the prediction pipeline.
The Role of Prediction Feedback
A key insight of this research is that prediction error serves as a proxy for identity. Every traffic type possesses a unique statistical signature; a model's failure to predict a stream provides valuable information for classification. The weight $\alpha$ plays a crucial role: if it is too low, the feedback benefit is lost; if it is too high, transient prediction spikes can cause misclassification.
Conclusion
This paper demonstrates that breaking the silos between traffic classification and prediction leads to a "virtuous cycle" of performance improvements. While the system shows sensitivity to initial classification accuracy and requires manual tuning of the weight parameter $\alpha$, it offers a robust path forward for intelligent network management. Future work will focus on integrating this framework into the Media Access Control (MAC) layer to evaluate its impact on real-time throughput and resource allocation in live simulations.
Summary updated