Cell-to-cell activity prediction for smart cities

In this paper, we analyze data from a large mobile phone provider in Europe, pertaining to time series of aggregate communication volume A
_i,j
(t) > 0 between cells i and j, for all pairs of cells in a city over a month. We develop a methodology for predicting the future (in particular whether two cells will talk to each other A
_i,j
(t) > 0) based on past activity. Our data set is sparse, with 80% of the values being zero, which makes prediction challenging. We formulate the problem as binary classification and, using decision trees and random forests, we are able to achieve 85% accuracy. By giving higher weight to false positives, which cost more to network operators, than false negatives, we improved recall from 40% to 94%. We briefly outline potential applications of this prediction capability to improve network planning, green small cells, and understanding urban ecology, all of which can inform policies and urban planning.