A Factorization Machine-Based Approach to Predict Performance Under Different Parameters in Cellular Networks

Abstract

The performance of network elements depends heavily on their parameter setting in cellular networks. Current practice of parameter setting relies largely on expert experience and self-organizing networks, which is often suboptimal. Therefore, how to find the optimal parameter combinations automatically is increasingly concerned by mobile network operators. In this article, a collaborative learning approach is proposed to meet this demand by predicting the performance of network elements under different parameter combinations before setting. The proposed approach captures the time-aware correlation between different network elements and their parameter combinations based on factorization machine to boost the prediction accuracy. Extensive experiment results demonstrate that the proposed approach outperforms the existing prediction approaches on a large-scale real-world network dataset from a metropolitan LTE network.