Learning probabilistic models of cellular network traffic with applications to resource management

Abstract

Given the exponential increase in broadband cellular traffic it is imperative that scalable traffic measurement and monitoring techniques be developed to aid various resource management methods. In this paper, we use a machine learning technique to learn the underlying conditional dependence and independence structure in the base station traffic loads to show how such probabilistic models can be used to reduce the traffic monitoring efforts. The broad goal is to exploit the model to develop a spatial sampling technique that estimates the loads on all the base stations based on actual measurements only on a small subset of base stations. We take special care to develop a sparse model that focuses on capturing only key dependences. Using trace data collected in a network of 400 base stations we show the effectiveness of this approach in reducing the monitoring effort. To understand the tradeoff between the accuracy and monitoring complexity better, we also study the use of this modeling approach on real applications. Two applications are studied -energy saving and opportunistic scheduling. They show that load estimation via such modeling is quite effective in reducing the monitoring burden.

The Learning and Prediction of Application-Level Traffic Data in Cellular Networks

“…However, the existing traffic prediction methods in this microscopic case still lag behind the diverse requirements of various application scenarios. Firstly, most of them still focus on the traffic of all data services [28] and seldom shed light on a specific type of services (e.g., video, web browsing, IM, etc). Secondly, the existing prediction methods usually follow the analysis results in wired broadband networks like the α-stable models 2 [29], [30] or the often accompanied selfsimilarity [21] to forecast future traffic values [15], [19], [22].…”

Use of Machine Learning for energy efficiency in present and future mobile networks

“…In [5], for example, authors suggest the use of cell-level traffic data to feed a classification system with potential applications to i) energy savings, by identifying low loaded cells as potential candidates to be switched off; and ii) spectrum planning, by identifying those cells that are likely to require additional spectrum resources. Similar applications could be derived from the work in [6], which evaluates regression mechanisms used to predict aggregated traffic from multiple cells. In [7], clustering of cells is used to improve traffic predictions.…”

Adaptive Estimation of Random Vectors with Bandit Feedback

“…For instance, using sensors to (i) monitor the temperature over a region Guestrin et al [2005] and (ii) detect contamination in a water distribution network Krause et al [2008]. Traffic monitoring in a cellular network is another application Paul et al [2014], where the underlying correlation structure plays a major role. To elaborate, the aim here is to collect traffic load measurements from a handful of base stations to form an estimate of the load on all base stations.…”

“…We assume that the underlying distribution is K-dimensional Gaussian with covariance matrix Σ: a model that has been shown to be practically viable in Paul et al [2014]. We employ the mean squared error (MSE) objective to capture the underlying correlation structure.…”