A regression framework for predicting user’s next location using Call Detail Records

Abstract

With the growth of using cell phones and the increase in the diversity of smart mobile devices, a massive volume of data is generated continuously in the process of using these devices. Among these data, Call Detail Records, CDR, is highly remarkable. Since CDR contains both temporal and spatial labels, mobility analysis of CDR is one of the favorite subjects of study among the researchers. The user next location prediction is one of the main problems in the field of human mobility analysis. In this paper, we propose a regression framework to predict next locations of users of cellular operators. We propose domain-specific data processing strategies and design a deep neural network model which is based on recurrent neurons and performs regression tasks. Using this framework on real-world data, we show that the error of the prediction decreases up to 74% in comparison to the traditional location prediction models. The results of this paper can be helpful in many applications from urban planning and digital marketing to predicting the spread of pandemics.

Introduction

Location prediction is a problem in the area of human mobility analysis which has drawn the attention of data scientists and machine learning researchers to itself during the past decade. Despite the simplicity of the problem description, the solution is expected to be complicated due to its dependence on many factors such as the environment, variability in users’ habits and their locations, and different formats of the data.

Location prediction problem could be described using the available data for users’ locations and the environment in which we need to predict the next locations of these users. Such data usually contain historical records that include information about time, location, user characteristics, and the infrastructure that collects the location information.

The prediction of human mobility has a wide variety of applications and many real-life use cases. Knowledge about the places each user may traverse through would supply valuable information to use in urban planning, country infrastructure management, and public transportation organization. Smart advertising and businesses would hugely benefit from such predictions. There are also many different domains like social and cultural researches, disease epidemiology studies, criminology investigations, cellular network infrastructure administration, etc. that could benefit from the information provided from human mobility predictions [1].

With the growth of smart mobile devices that are spread geographically apart in the environment, gathering human mobility data has become much easier compared to traditional methods like surveys or census which suffer from the static or low resolution spatial and temporal information. There are various sources of data used by researches in the field of location prediction. For example, Lenormand et al. [2] has done a cross-check analysis by comparing results obtained from different sources of data (cell phones, Twitter and census) and compared the levels of correlations between these sources in three different aspects of spatial distribution of population, temporal evolution of people density, and mobility patterns of individuals.

The most common data resource that has been used in human mobility researches is mobile devices. These data might be collected from GPS-based navigation applications, social networks which work over the internet, or cellular networks infrastructure.

Cellular networks possess valuable data like call detail records (CDRs) of the users, accounting information, and infrastructure information which can play an effective role in human mobility analysis. CDR is a type of metadata which describes users’ activities in a cellular network. CDR data are commonly used for the purpose of billing users, value-added services, and network maintenance and optimizations by cellular network operators and infrastructure maintainers. However, having both spatial and temporal information about users has also made CDR a good resource for analyzing human mobility.

In this paper, we propose a framework which consists of a recurrent neural network regression model for predicting users’ next location based on the spatio-temporal information in CDR records. What distinguishes this solution from traditional models proposed for location prediction could be stated as follows:

• We propose a practical processing framework which defines modules and model to handle the data from raw CDR to accurate predictions.

• The proposed framework uses domain-specific data preparation methods for real-world data to form meaningful user trajectories which increase the performance drastically.

• The prediction module of this framework uses geographic coordinates instead of domain-independent semantic labels for locations in the process of learning and prediction. This approach resolves the sparsity issue, and copes with possible errors in the CDR data.

In order to compare the performance of our proposed framework with existing models, we implemented two baseline models and a common recurrent neural network classification model that have been widely used in the field of location prediction. We test these four models/frameworks on a real-world dataset, which includes CDR data collected from 12 users over a period of two years by one of the largest mobile phone operators in Iran, and compare the results with respect to different metrics of performance.

The remainder of this paper is organized as follows. We first provide an overview of the related works in the field of human mobility prediction in Section 2. Section 3 introduces the dataset and Section 4 formulates the problem. In Section 5, we discuss traditional models and in Fig. 6 our prediction framework is introduced. We explain the experiment results in Section 7, and finally Section 8 concludes the paper.