Clustering and modeling of wireless backhaul data traffic

Abstract

To handle the future demands of mobile broadband, knowledge of user behavior is of great importance. Knowing when and where user demand will be high allows for better planning of spectrum and computational resources. To aid this, this thesis investigates if the behavior of traffic throughput of wireless backhaul links from a European operator can be segmented into different clusters. We use both the k-means clustering algorithm and t-distributed stochastic neighbor embedding to attain the clusters. No distinct patterns emerge from the data, which instead appears to be uniformly spaced without clear boundaries. To aid simulation of wireless links in future studies we also model the traffic behavior of urban links. The correlation between the model parameters as well as the error terms are calculated as a function of the geographical distance between the links. This helps decide whether links in proximity behave similarly or not. We find that the traffic on urban links has a similar shape but the model parameters are not correlated with respect to the distance between them. The parameters can be sampled from given distributions to generate synthetic traffic data.