Publication authored by Amnir Hadachi, Mozhgan Pourmoradnasseri, Kaveh Khoshkhah. You can find our publication here link.

Highlights

• Novel use of mobile data for commuting patterns.
• Advanced Hidden Markov Model for trajectory analysis.
• Detailed urban and intercity commuting insights.

1. Problem Statement

Our Research focuses on exploring the potential of cellular network mobile data to extract extensive movement and commuting patterns within the population. Our approach comprises two main levels: one involving the utilization of the Hidden Markov Model (HMM)^[1] and OD-matrix to construct comprehensive mobility patterns and flows^[2], and the other level concentrating on estimating departure and arrival times for commuting journeys in the OD-matrix, alongside classifying movement status (Stay or Move)^[3]. As a result, the paper is structured to emphasize the description of the data used in this research and the challenges encountered due to its inherent nature. The second section covers related work and similar projects, while the third section provides a detailed overview of our proposed approach. Lastly, the final section presents a clear summary of the results, followed by an in-depth discussion.

2. Passive Mobile Network Data

The literature discusses various types of mobile data for extracting human mobility patterns^[4], often dependent on telecommunication providers and their equipment. Predominantly, Call Details Records (CDR) and Visitor Location Registry (VLR), used individually or in combination, are the most commonly employed.

CDR data comprises registered records documenting mobile phones’ interactions within the network or with other devices. It includes mobile identity, cell-ID, transaction start time/duration, and event types (e.g., calls, SMS). These records are generated during phone use.

Anonymized user ID	cell-ID	Timestamp & Event
G3Z03R8269	W8575	1525757266 & par-c
G2H99K9882	F3268	1531421050 & detach-c
G8J84W8462	I7520	1526482568 & loc-up-c

Table 1: An example of CDR and VLR dataset attributes.

VLR, located in the mobile communication network’s database, contains subscribers’ exact locations within the service area. It holds information similar to CDR but includes additional details about handovers and phone activities. VLR data, suitable for real-time applications, is deleted upon subscriber departure.

Mobile operator-collected data usually includes cell IDs and timestamps, representing temporal information about event occurrences. This data is combined with cell plan data containing coverage area shapes and signal strength to analyze mobility. This combination transforms it into spatiotemporal data, used to construct user trajectories characterized by sparseness in space and time, As shown in Fig. 1.

Fig. 1: Representation of a path derived from Call Details Records (CDR) data - The polygons depict coverage areas, while the arrows indicate the sequential occurrence of events.

A dataset in a particular study comprised over 600 million anonymized CDR and VLR events from around 300,000 users in Estonia in May 2018. Supplementing event records with cell-plan data, which includes coverage area shapes, enables analysis of the spatiotemporal aspect. However, challenges persist in extracting spatial aspects due to obstacles like radio fluctuations, weather changes, and false displacements in the mobile network, leading to considerable noise in positioning users.

3. Methodology

Call detail record (CDR) data from mobile networks is the methodology’s core. We employed a Hidden Markov Model (HMM) to reconstruct and analyze the trajectories derived from CDR data. This approach facilitated the extraction of origin-destination matrices at various geographical levels, providing a detailed view of commuting patterns across Estonia. The model’s ability to handle large-scale data and its innovative use of mobile network information are particularly noteworthy.

The primary goal of our research is to comprehend intercity commuting behaviors at the national and intra-city levels. To achieve this, we utilize mobile data to derive the Origin-Destination (OD) matrix depicting travel patterns among 77 municipalities in Estonia, followed by the OD matrix for travel between the 15 counties in the country. Additionally, we aim to delve into daily commuting behaviors among various districts within Tallinn, the capital of Estonia.

To accomplish these objectives, our study involves employing a Hidden Markov Model (HMM), where, in the initial phase, the hidden states of the model represent municipalities and counties at the national scale. Subsequently, in the latter phase, the model’s hidden states correspond to the various districts within Tallinn. After acquiring the parameters of the HMM, the Vertabi algorithm is utilized in each phase to allocate the maximum likelihood sequence of hidden states (such as municipalities or city districts) to a sequence of cell-IDs derived from mobile data.

This approach enables us to infer and analyze commuting patterns, delineating travel flows between different geographical units—municipalities, counties, and city districts—through the application of HMM and data sequence analysis methodologies.

Fig. 2: Architecture of the Methodology Adopted.

4. Experimentation

The experimentation involved analyzing data from over 300,000 users in Estonia during May 2018. The study focused on three key aspects: large-scale commuting across the country, commuting within the most populated counties with the highest mobility, and specific commuting patterns in major cities. A notable part of the analysis was the construction of an OD matrix that visualizes commuting flows within and across counties.

4.1 Large-scale commuting patterns in Estonia.

In this research, our primary objective is to uncover extensive commuting patterns across Estonia using our proposed method. Consequently, we extracted a total of 1,723,866 trips occurring in May 2018, encompassing travel between 77 distinct urban and rural municipalities within Estonia. Among these journeys, 1,405,552 trips originated and concluded within the same county, while 318,314 trips occurred between cities of different counties. The visual representation of the extracted Origin-Destination (OD) matrix from our dataset reveals notable trends. Predominantly, a significant portion of the trips within each county is observed between cities within the same county. Additionally, a concentration of trips is evident in Harju county, particularly notable due to the presence of Tallinn, Estonia’s capital. Apart from intra-county trips and those to Harju county, a substantial number of trips are directed towards neighboring counties.

As anticipated, most trips occur between cities within Harju County, amounting to 955,654 trips. This aligns with statistical data indicating that over 40% of the country’s population resides in Harju County, according to Estonia’s statistics database (Stat2018ee), while 56% of enterprises are located there (Katrin2018es). Following Harju County, Tartu County emerges as the second-ranked county, attributed to its prominence in educational activities and the services industry. To gain a comprehensive understanding of mobility patterns among all counties, our focus shifts to depicting three specific trip types: trips from one county to others, trips within a county, and trips from other counties to a specific county.

The visualized results in Fig. 3 depict the fraction of trips originating and terminating within each county, originating from each county and concluding in other counties, or inbound trips from all other counties to a particular county. This analysis reveals two noteworthy observations. Firstly, major counties exhibiting high internal mobility include Harju County (with the largest population and Tallinn as its capital), Ida-Viru County, Tartu County, and Pärnu County. Secondly, symmetry is apparent in the number of trips in counties concerning inbound and outbound travel.

Fig. 3: Trips between counties of Estonia.

Furthermore, our investigation into inbound and outbound mobility includes extracting top trip destinations among Estonian cities using our trip extraction methods Fig. 4. The distribution obtained underscores the significance of each county’s capital city in internal and external mobility. The analysis of extensive commuting patterns indicates a concentration of mobility within Harju county, particularly around Tallinn city, and in Tartu county, centered around Tartu city. Consequently, our subsequent section will delve into depicting mobility within Harju and Tartu Counties, focusing on Tallinn and Tartu as gravitational centers, respectively.

Fig. 4: The arrangement of primary trip destinations among Estonian cities acquired through our methodologies for trip extraction.

4.2 Commuting in Harjumaa County

Our study focused on understanding commuting patterns, particularly within Harju County in Estonia. We discovered that Tallinn is a major commuting destination by analyzing over 350,000 trips between different municipalities and Tallinn over a month. The majority of movements within Harju County revolve around Tallinn, with noticeable commuting activity between neighboring municipalities. Viimsi, Jõelähtme, and Rae are significant contributors to commuting towards Tallinn. Additionally, our analysis revealed distinct commuting patterns during weekdays and weekends, with industrial cities like Tallinn having more activity during workdays while other municipalities show increased movement on weekends. These trends align with economic activities within the county, with Tallinn’s diverse activities reflecting its status as the capital city of Estonia.

Fig. 5: Trips to Tallinn from the other municipalities of Harju County.

4.3 Commuting in Tartumaa County

Tartu County, Estonia’s second most populated region, houses Tartu, the country’s second-largest city. Analyzing over 90,000 trips within the county, we observed that commuting patterns primarily center around Tartu city, with neighboring municipalities significantly contributing to overall commuting activity.

Visualizing commuting distribution towards Tartu city Fig. 6, Tartu County and Kambja stand out as major contributors, followed by Luunja, Elva, Nõo, Kastre, and Peipsiääre. The high population in Tartu city naturally leads to substantial commuting activities. Tartu city displays consistent mobility during both weekdays and weekends, akin to Kambja municipality, while Elva and Peipsiääre municipalities are more active during weekends.

Fig. 6: Trips to Tartu from the other municipalities of Harju County.

Economic activities largely center around Tartu city, influencing Elva, Kambja, and Tartu municipalities. These trends are noticeable in commuting patterns, highlighting the connection between the county’s economic activities and commuting behaviors.

5. Conclusion

This research is a testament to the potential of mobile data in urban planning and transportation management. The study offers valuable insights into urban and intercity commuting complexities by converting cellular network data into meaningful mobility patterns. The findings promise to enhance sustainable urban development and address transportation challenges in modern cities. For more details and additional insights discovered and analyzed, check our paper.

---

1. Rabiner, Lawrence R. “A tutorial on hidden Markov models and selected applications in speech recognition.” Proceedings of the IEEE 77.2 (1989): 257-286. ↩︎

2. Järv, Olle, et al. “Mobile phones in a traffic flow: A geographical perspective to evening rush hour traffic analysis using call detail records.” PloS one 7.11 (2012): e49171. ↩︎

3. A, Hadachi, et al. “Exploring a new model for mobile positioning based on CDR data of the cellular networks.” arXiv preprint arXiv:1902.09399. ↩︎

4. Jiang, Shan, Joseph Ferreira, and Marta C. Gonzalez. “Activity-based human mobility patterns inferred from mobile phone data: A case study of Singapore.” IEEE Transactions on Big Data 3.2 (2017): 208-219. ↩︎