To explore a more holistic view of urban human mobility, a few researchers have previously attempted to build models driven by a fusion of multi-source datasets. Yet, owing to limited understanding of the intrinsic characteristics and relationships underlying multi-source datasets, most of such interventions naively merge these datasets, producing sub-optimal prediction models. To address this issue, we propose a three-step methodological framework to capture urban mobility from an integrative point of view. The novelty of our framework is three-fold: (i) a systematic characterization of the multi-source data to leverage hidden relationships when integrating the data; (ii) integrating data with contextual information of the urban environments to improve explainability; and (iii) conforming the model building to incremental learning paradigm to adapt to changing patterns of mobility data. Using the New York City (NYC) case study, extensive analyses show salient relationships within datasets which could form a basis for optimal data fusion and subsquently improving the mobility modelling pipeline in line with the proposed three-step methodological framework.
Human mobility patterns in urban settings are indicative of various phenomena such as travel demand,
trafic congestion, resource allocation, CO2 emission, and infectious disease spread. Therefore,
understanding such patterns is crucial to applications such as transportation management, urban planning,
resource optimization, and epidemiology among others. In recent years, the pervasive use of sensing
technologies has produced an enormous amount of human mobility data that researchers have used
to answer critical questions pertaining to when, why, and how people move from one place to the
other [
Urban mobility is multi-faceted in nature, mainly due to diferent modes of transport. As such, we
have recently seen a growing shift from single source to multi-source data-based models for human
mobility modelling in urban settings [
To optimize multi-view mobility modeling, we propose an integrative three-step methodological framework, as shown in Figure 2. The framework combines multi-source mobility data for a more representative picture of travel behaviours, then integrates it with existing domain-knowledge to minimize data requirement while enhancing explainability, and finally conforms the mobility modeling to incremental learning, adapting to ever-changing urban mobility conditions. Each component is briefly overviewed in the following subsequent sections.
In the context of multi-view learning, data fusion captures the comprehensive view of urban mobility
patterns, and also difuses the limitations that come along with relying solely on single-source data.
However, most existing works make general assumption of the multi-source data when it is being
fused. For example in [
It is widely known that data-driven models are only as good as the quantity and quality of data they are
trained on. Given issues relating to incompleteness, inconsistencies, sparsity and uncertainity, it is often
dificult to acquire large amounts of quality mobility data. However, integrating empirical data with
domain-specific knowledge often validates the data, enhancing model reliability and interpretability,
which is still an open challenge in human mobility research. To the best of our knowledge, very few
mobility modeling approaches exist in literature today, which incorporate domain-specific knowledge.
In MobTCast [
Urban mobility is highly dynamic, with context and patterns that typically change from time to time. Conventional data-driven models, though they have shown considerable performance, they often struggle to keep at per with evolving urban environments, where mobility patterns keep on changing due to changing policy, transportation systems, road networks, weather conditions, and events. As an emerging approach, Incremental Learning (IL) which enhances adaptability of models by learning continuously from new incoming data streams, is poised to address these challenges sufered by static models. In our framework, the knowledge-enhanced mobility model will be tuned to continuously update on only new incoming data, as and when it becomes available, while retaining previously learned knowledge.
In the previous section, we give an overview of the proposed integrative three-step methodological framework. Note however that in the present work, the focus is to preliminarily use the case of New York City (NYC) to analyze the underlying characteristics of multi-source data, setting up an optimal foundation for efective data fusion and subsquent multi-view mobility modeling in the line of the proposed framework.
In this work, we characterize and compare mobility flows extracted from NYC’s taxi and Citi bike trip
records. NYC is the largest city in the United States, known for its fast pace, dynamism and cultural
diversity. It continues to be a global hub for finance, culture, commerce, and innovation. It covers a
total area of 784 km2 with an approximate population of 8.5 million people as of 2023. The city is
composed of five boroughs: Manhattan, Brooklyn, Queens, Bronx, and Staten Island, each with its own
unique character, and it is demarcated into 263 zones, as shown in Figure 3. The city is covered with an
extensive and a well-integrated transport system that is comprised of public transit, commuter rails,
taxis, and bicycles. A number of researchers have utilized NYC as a case study to investigate diferent
aspects of urban mobility. For example, in [
In this section, the three NYC datasets used in this study, are summarized. The Yellow taxi and Green taxi datasets are provided by the Taxi & Limousine Commission of NYC, while the bike-sharing dataset is provided by Citi bike NYC. For all datasets, each record is a distinctive trip extracted, with an origin, destination, duration, distance and fare among other trip attributes. After pre-processing of the datasets, a total of 10,047,135 trips were successfully extracted from the Yellow taxi, 1,080,844 trips from the Green taxi, and 1,000,000 trips from the NYC Citi bikes. To ensure temporal compatibility, the collection period was the same for all the three datasets, running from April 1, 2017 to April 30, 2017.
We characterize multi-source datasets based on three meta-metrics: (1) correlation; and (2) representativeness. We argue that examining and leveraging these three meta-metrics will form a basis for optimal data fusion and subsquently improving the multi-view mobility modelling pipeline. Firstly, we extract trips – which is the basic unit of mobility for this study. With all the three datasets having pickup and drop-of attributes, extracting origins and destinations (OD) is simple and straightforward. However, to reduce the size and address errors, irrelevant attributes and errant records were intuitively eliminated right away before any form of feature engineering was done. For example, all records with trip duration < 5 minutes and/or trip distance > 30 miles, were automatically dropped for all datasets.
We examine mobility (travel demand) to evaluate whether there exist temporal correlation between Yellow taxi, Green taxi, and Citi bikes trip patterns in NYC. Owing to significant diference in scale among the datasets , we observe in Figure 4 that both Green taxi and Citi bike data temporally lag the Yellow taxi data, but there exist similar mobility trends worth further investigation. To provide fair and meaningful comparison, we normalize the data to capture the hourly variations as a proportion of total mobility lfow for each dataset in Figure 5. We then use three diferent measures (Cosine-similarity, Pearson Correlation, and Spearman Rank Correlation), to explore similarities in their temporal distribution. These measures have been widely used in previous studies to quantify the similarity between two vectors in a multi-dimensional space, depending on the scale and context of the data. In our context, we opted for them because they quantify similarity in patterns regardless the diference in scale.
Taxi and Bike-sharing mode choices provide NYC residents with certain levels of convenience, timeliness, and flexibility. However, they co-exist and complement other modes in an integrated transport system. It is thus crucial to examine if travel demand extracted from the three datasets reflect their actual respective mobility activity in NYC.
We perform a spatial coverage analysis to ensure that all significant zones are truly represented. For each dataset, we compute travel demand as the proportion of total outgoing trip counts at the zone level. We then use a choropleth map to visualize and compare the spatial distribution of these values, for the three datasets. We then use oficial statistics from the 2017 citywide mobility survey of NYC, to benchmark patterns derived from our datasets. Particularly we compare the trip proportions extracted from the datasets to the relative trip proportions extracted from the 2017 citywide mobility survey of NYC. And consequently we perform statistical tests on some aspects of the observed trips in comparison with surveyed trips.
This section summarizes the results obtained along with detailed discussions.
Figure 5 depicts the temporal distribution of normalized travel demand as extracted from Yellow taxi, Green taxi, and Cite bike trip records. Considerable relationships can be observed among the three datasets within the time dimension. For instance, we can see a similar drop in demand from midnight up until 5 a.m., across all the datasets. We also see a shoot-up in demand at 8 a.m. and another one at 6 p.m. To probe these relationships further, Table 1 summarizes results of three well-known measures of similarity against pairs of the three datasets.
From the table 1, we can see that travel demand across all pairs of datasets generally has a positive correlation in the temporal space. Particularly, travel demand patterns from Yellow taxi & Green taxi exhibit a very strong alignment across time, for all the three measures utilized, whereas the correlation between green taxi & cite bikes is relatively weak but still considerably above average.
To validate this representativeness, we statistically compare our trip data with the 2017 citywide mobility survey of NYC. Figure 7 illustrates variations in the mode composition as extracted from the trip data and survey data. We observe that among the three modes, Green taxi exhibits a strong fit, when trip data is compared against survey data. This is further collaborated by the CDF distribution of their respective trip durations as shown in Figure 8. We can observe that though Yellow taxi and Citi bikes both have considerable similar distribution (p-value=>0.05), Green taxi once again exhibit remarkable values of Kolmogorov-Smirnov (KS) test, which demonstrating a strong representativeness of the Green taxi records against the actual mode split as by the 2017 city-wide mobility survey of NYC.
In this study, we proposed a three-step methodological framework for modelling urban mobility from an integration of multi-source data. The preliminary step of our framework involves the systematic characterization of multi-source data as a foundation for a weight-based data fusion. Using NYC as a case study, we characterized three mobility datasets based on two meta-metrics: (1) correlation; and (2) representativeness. Our findings revealed that there exist strong correlation between datasets, which can be leveraged in data fusion. For instance, we see a strong temporal correlation between Yellow taxi & Green taxi, in terms of relative travel demand, which could be used to infer missing data in one dataset using the other. It could also be used in weighted average fusion of the data, where weights are based on the correlation. Also revealed via our findings is the level of representativeness between the respective mobility datasets and the actual urban scenarios. For example we see in our case study how the Green taxi has the most alignment with actual sample data surveyed from the population in the 2017 city wide mobility survey of NYC. This could also be leveraged in the weight scheme to ensure the most representative and reliable dataset contributes most to the data fusion. It should be noted that this work does not attempt to examine all the three steps of the proposed framework, but rather to give its overview, and the characterization of the multi-source data to be used as ingredients in the subsequent steps of the framework. One of the limitations of this work however, results from the fact that currently, we only focus on data sources from NYC. If matching multi-source data from other mega cities like Shenzhen and Beijing was readily available, a comparison of mobility trends from diferent urban cities against the two meta-metrics would further validate the characterization of multi-source data. Another limitation stems from the age of data used (2017), given major mobility shifts that have occurred with the coming Covid-19 pandemic. While very recent mobility data is available at the NYC open data portal, for the interest of measuring representativeness of the data, we wanted to avoid a time diference of data collections between trip records data and the mobility survey data, given that the city-wide mobility survey data available is of 2017. In our future work, while acknowledging and working around the above limitations, we plan to leverage the findings in this preliminary work, to design an efective weighting scheme for the data fusion. We shall then improvise efective ways to incorporate contextual knowledge in the fused data, and finally employ incremental learning to build an eficient mobility prediction model in line with the proposed methodological framework for diferent application contexts. [17] E. Rajabi, S. Nowaczyk, S. Pashami, M. Bergquist, G. S. Ebby, S. Wajid, A knowledge-based ai framework for mobility as a service, Sustainability 15 (2023) 2717. [18] X. Hao, R. Jiang, J. Deng, X. Song, The impact of covid-19 on human mobility: A case study on new york, in: 2022 IEEE International Conference on Big Data (Big Data), IEEE, 2022, pp. 4365–4374. [19] A. A. Rajput, Q. Li, X. Gao, A. Mostafavi, Revealing critical characteristics of mobility patterns in new york city during the onset of covid-19 pandemic, Frontiers in Built Environment 7 (2022) 654409. [20] N. Dong, J. Zhang, X. Liu, P. Xu, Y. Wu, H. Wu, Association of human mobility with road crashes for pandemic-ready safer mobility: A new york city case study, Accident Analysis & Prevention 165 (2022) 106478.