Using Deep Learning to Understand Travel Demands in Different Urban Districts

Abstract

Travel demand forecast plays an important role in transportation planning. Classic models often predict people’s travel behavior based on the physical built environment in a linear fashion. Many scholars have tried to understand built environments’ predictive power on people’s travel behavior using big-data methods. However, few empirical studies have discussed how the impact might vary across time and space. To fill this research gap, this study used 2019 anonymous smartphone GPS data and built a long short-term memory (LSTM) recurrent neural network (RNN) to predict the daily travel demand to six destinations in Austin, Texas: downtown, the university, the airport, an inner-ring point-of-interest (POI) cluster, a suburban POI cluster, and an urban-fringe POI cluster. By comparing the prediction results, we found that: the model underestimated the traffic surge for the university in the fall semester and overestimated the demand for downtown on non-working days; the prediction accuracy for POI clusters was negatively related to their adjacency to downtown; and different POI clusters had cases of under- or overestimation on different occasions. This study reveals that the impact of destination attributes on people’s travel demand can vary across time and space because of their heterogeneous nature. Future research on travel behavior and built environment modeling should incorporate the temporal inconsistency to achieve better prediction accuracy.

As cities continue to grow both in population size and in the greater spatial dispersal of new developments, citizens now have more options for choosing where to live and where to travel. Meanwhile, the expansion exacerbates the pressure on already overloaded transportation networks in most auto-dependent cities in the United States. As a result, travel demand forecasting plays a vital role in transportation planning. For a long time, transportation researchers and practitioners have used classic four-step modeling to forecast future travel demand from/to an area based on local contexts, such as demographic and socioeconomic background and regional characteristics. The conceptualization of modeling travel behaviors from the built environment significantly improves transportation planners’ authority when making future-oriented decisions on significant past infrastructure investments, such as highway construction and transit system design ( 1 , 2 ).

However, there are also significant drawbacks to this approach. First, in the model, the travel demand for a destination is usually determined by the land use and employment status of a place ( 3 ). Because of the limited details provided by large-scale surveys, trip purposes are usually divided into four categories merely based on whether the origin is home and whether the destination is the workplace. This over-simplified categorization only slightly touches on the endogenous motivations of a place that people choose to visit. However, scholars have argued that the unobserved reasons people make individual travel choices are difficult to survey and dynamically account for in traditional statistical models ( 4 , 5 ). Second, large-scale surveys such as the American Community Survey (ACS) by the United States Census Bureau are too time-consuming and labor-intensive to incorporate rapidly changing travel patterns as emerging travel options come into play ( 6 ). However, many studies have proved that travel demands, regardless of modes, are changing over time ( 7 – 10 ). The loss of time effectiveness associated with using survey data in forecast models hinders transportation planners from (re)distributing limited resources in a timely and dynamic manner ( 11 ).

Benefiting from the developments in big data and artificial intelligence (AI), many researchers have proposed using concepts and techniques of big data to build intelligent transportation systems (ITS) as part of a solution to urban problems ( 12 , 13 ). Notably, the use of neural networks in the transportation field reflects a converging trend of data-rich analyses using machine learning techniques in travel behavioral research ( 14 , 15 ). However, previous studies predominantly focused on the predictive power of these and overlooked their explanatory power in the search for the relationship between people’s travel behavior and the physical environment. Besides, people’s travel choice consists of many endogenous characteristics in different built environments. The lack of such considerations will encourage researchers to abuse these models blindly and disregard their inherent limitations in querying travel behaviors from external contexts ( 4 , 16 – 18 ).

This study aims to explore the variations of prediction results incurred by various urban functionalities and different destination locations. To fulfill this goal, it uses a long short-term memory (LSTM) network to predict the daily number of trips to six destinations in Austin, Texas, and answer the following research questions:

What are the variations of prediction accuracy for different destinations?

How does the accuracy variation change over time for different destinations?

The presented study contributes twofold to the literature. First, travel-related data passively collected from GPS-equipped smart devices, compared with those from active methods such as travel surveys and sensors, have more advantages in data volume and frequency and are more informative in planning urban transport systems ( 19 ). This study empirically proves the applicability of big urban data in gaining useful knowledge to guide planning practice. Second, most empirical studies querying travel behavior from built environments in the past have addressed the importance of land use, such as urban parks and mixed-use downtown, in promoting people’s travel activities ( 20 – 22 ). Nevertheless, limited evidence is provided to indicate how this promotional relationship might shift over space and time because of factors embedded in the destinations’ nature. To this end, this study uses a hybrid LSTM network with both dynamic predictors and a static predictor indicating the urban functionality and the regional location to predict the daily travel demand for different destinations. By analyzing the prediction performance, this study provides proof of heterogeneous impacts of destination factors when predicting people’s travel behavior at different times.

This paper is organized as follows. In the next section, we review the literature on predictive studies in transportation using neural networks. Then, we introduce the data and measurements as well as the LSTM network model architecture design. After that, we discuss the forecast results and evaluate the model performance for different destinations at different times. In the last section, we recapitulate the key results and summarize the corresponding implications for transportation research and practice in the future.

Literature Review

Innovative AI algorithms such as artificial neural networks (ANNs) in deep learning have emerged as new modeling alternatives in transportation. Compared with classical statistical models, ANN is often regarded as a more flexible prediction model for handling complex datasets ( 16 ). Mo and Su proposed an ANN model to predict real-time transit passenger volumes using time-series records and weather data. They argued that the ANN model could self-correct stochastic problems when forecasting people’s travel behavior ( 23 ). Çetiner et al. ( 24 ) built a neural network to predict short-term traffic volume in Istanbul, Turkey, and found out that time features and last observed volume were effective in the prediction.

In recent years, the recurrent neural network (RNN) has been getting more attention in transportation because of its better performance in predicting sequential events. Compared with feedforward ANNs, in which learning iterations are independent, the learning architecture of an RNN is recurrent and successive in that RNN models pass predicted outputs from previous learning to the next, sequentially ( 25 ). Deep neural networks such as RNNs can model complicated nonlinear relationships in an unsupervised fashion and reinforce the prediction accuracy through continuous learning ( 26 ).

Long short-term memory recurrent neural network, or LSTM RNN in short, is a popular RNN model that has been used by transportation scholars in recent years. Unlike convolutional RNNs, the LSTM RNN is a powerful model that learns long-term dependencies of sequential events by generating memory units ( 27 , 28 ). Khan et al. ( 29 ) analyzed the performance of a simple RNN, a gated recurrent unit (GRU) RNN, and an LSTM RNN in predicting the hourly traffic volume and annual average daily traffic, and found that the LSTM RNN was the champion in prediction accuracy. Tian and Pan ( 30 ) compared LSTM RNN with other machine learning models in predicting short-term traffic flow and found out that the approach achieved the highest prediction accuracy and generalizability. However, a limitation of their experiment was that their model only used traffic data. Variation caused by externalities was not evaluated in their comparison. Zhao et al. ( 31 ) designed a spatial–temporal LSTM network by establishing a learning relationship from observation at a location at a previous time point to the prediction at another location at the next time point. By testing the performance of the 2D LSTM model in predicting short-term vehicle traffic using sensor data, they concluded that the prediction was more accurate than many supervised models. Zhao et al. was an improvement of Tian and Pan’s study by incorporating spatial information in the modeling. However, both Tian and Pan and Zhao et al. used data actively collected for experiments, which could still be time-consuming, and their methodologies might be difficult to generalize in different contexts. As an effort to test the model applicability on passive data, Orsini et al. ( 32 ) used WiFi trace data in Bologna Airport, Italy. They proved that the LSTM network still can provide satisfactory results in predicting passenger behavior using passive data.

Recently, more studies have been introduced to academia in which transportation forecasters fused prediction models to achieve better prediction results. Ma et al. ( 33 ) first deployed a convolutional neural network (CNN) to capture the correlation between inter- and intra-day traffic flow patterns and then fed the output to an LSTM network to improve the prediction accuracy. Similarly, Yao et al. ( 34 ) proposed a combined model in which a CNN handled the spatial dependency, and an LSTM handled the temporal dynamics. A common limitation is that they both considered spatial dependency in a small geographical area (six adjacent traffic detectors in Ma et al. and a 3 × 3 convolution panel in Yao et al.). As a result, even if the prediction results are better than simpler models, whether they are equally competent in larger geography is questionable.

In this study, we built a hybrid LSTM network with both dynamic and static predictors to examine the influence of the destination attributes on predicting people’s daily travel demand for different destinations in the city of Austin, Texas.

Methods

Study Area and Data

The study chose Austin, Texas, as the study area and selected the Census block group (CBG henceforth) as the geographic unit of analysis. We chose six places with distinct urban functions (i.e., downtown, university, airport, and POI clusters) as target destinations. Moreover, we selected three clusters in different regional locations, namely inner-ring, suburban, and urban-fringe. Since the POI cluster on the urban fringe overlapped with two CBGs, we combined trips to both as the total trips to the cluster. In total, there were six target destinations in the final modeling: downtown, the University of Texas at Austin (university henceforth), Austin-Bergstrom International Airport (airport henceforth), one inner-ring POI cluster (the Barton Creek Square shopping mall), one suburban (the Domain), and one urban-fringe POI cluster (the Lakeline Mall). The locations are shown in Figure 1. It is worth noting that there could be journeys to these destination CBGs that were not relevant to these facilities. However, given the dominant attractiveness of these places, we relaxed the assumption by treating trips to the corresponding CBGs as the trips to these facilities. The destination type with six possible values was the static predictor in the LSTM model.

Figure 1.

Study area and target destinations.

The travel-related data in this study is the time-series GPS data published by a private data company called SafeGraph. The data was collected from anonymous GPS pins of personal mobile devices daily. The key in the data set is the Federal Information Processing Standards (FIPS) code of the origin CBG, which is defined as the “home” of a device if the device has been located during the nighttime over a 6-week time window. The sizes of the student device group (i.e., devices seen to have location changes during the period) and the population device group (i.e., all devices regardless of activities) are shown in Appendix A1 and A2. For the purpose of modeling the intra-city travel demand, we only selected the travel records for which the origin CBGs were in the study area. The column-of-interest in this study is the destination CBGs, which indicates the number of devices whose “home” was the origin CBG that visited the destination on a particular day in the past. A visit to the destination is loosely defined as a device stopping there for more than 1 min during the period. By parsing the raw dataset and regrouping it on destination CBGs, we retrieved the number of visits to target destinations on each day from January to December in 2019. Besides GPS data, we also incorporated weather data from the National Weather Service Forecast Office to capture the effect of daily temperature and precipitation on people’s travel demand ( 35 ). At this stage, we retrieved the dynamic data needed for the prediction model.

Measurements

Target Variable

This study selected the daily number of trips as the target variable indicating the travel demand. The histogram in Figure 2 shows the distribution of the daily number of trips to all destinations of interest in Austin. In general, the daily travel demand to all destinations ranged from 0 to 3500 with two pinnacles at around 1000 and 2000.

Figure 2.

Histograms of daily number of trips to six target destinations in 2019, Austin, Texas.

Table 1 provides an overview of descriptive statistics for the target variable. According to the dataset, there were over 2.5 million trips made by Austin residents to the six destinations. In particular, the most popular destination was downtown, taking up about 30% of the total trips. Then, the proportions of the suburban POI trips, the urban-fringe POI trips, and the university trips were somewhere between 17% and 18%. The last two destinations were the airport and the inner-ring POI, where the proportions of the total number of trips were around 12% and 6%, respectively. Notably, the largest variation of the daily number of trips was identified at the university, while the smallest was found at the inner-ring POI.

Table 1.

Descriptive Statistics of Daily Number of Trips in 2019, Austin, Texas

Destination	Sum	Mean	SD	Min.	Max.
Airport	306,889	841	151	471	1384
Downtown	741,272	2031	411	684	3174
Fringe POI	432,405	592	184	174	1268
Inner POI	157,323	431	123	80	1044
Suburban POI	457,696	1254	183	192	1777
University	434,851	1191	859	59	3332
Overall	2,530,436	990	642	59	3332

Note: SD = standard deviation; Min. = minimum; Max. = maximum; POI = point-of-interest.

Predicting Variables

To forecast travel demand on a particular day for a destination, we chose to use a hybrid LSTM model that took the three dynamic indicators, namely the travel demand and two weather data on the previous day, and one static destination indicator as the input predictors. In particular, travel demand on the previous day was the observed number of trips on that day. To indicate the weather condition of a day, we used the average temperature in degrees Fahrenheit and the rainfall amount in inches as the two weather indicators. Appendix A3 demonstrates an uneven distribution of travel demand in different weather conditions, indicating that the influence of weather conditions on people’s travel was significant. Then, we used the destination type as a static predictor in our model to distinguish different destinations. It is worth noting that using a hybrid neural network model with both dynamic and static predictors is considered “unorthodox” in that it tries to capture temporal dynamics and static settings simultaneously. However, the importance of static information is commonly brought up in many predictive studies using deep learning approaches so that hybrid models have been gradually used in fields like transportation and beyond ( 32 , 36 ). With that said, we made the presented LSTM model a hybrid one so as to contribute to this aspect.

Controlling Variables

To evaluate the model performance on different occasions, we used three binary identifiers to indicate the periodic attributes: weekend, non-working federal holiday ( 37 ), and university school session ( 38 ). Figure 3 shows the distribution of the travel demand in different periods for the six destinations. In each graph, the vertically long component segments (the “thicker” parts) indicate more observed trip amounts at this magnitude, and components that are horizontally long (the “longer” components) indicate a more dispersed distribution. The various distribution shapes in Figure 3 show that the joint effects of destination type and periodic characters are significant, which could hinder the performance of the prediction model when targeting the traffic to different destinations on different occasions.

Figure 3.

Number of trips distribution by time-of-year identifier by destination type: (a) weekend; (b) holiday; and (c) school session.

Multivariate LSTM Travel Demand Network

There were three steps in the complete modeling process of this study: data preparation, model building, and model training. Besides the data preparation introduced in the previous section, we also encoded the categorical destination predictor using a label encoder and transformed the categorical values into integers. After organizing the input data set, we then split it into two sets for training and testing. The training set included the daily observations from January 1 to August 31, 2019 (243 days), and the test set included the remaining observations from September 1 to December 31 (122 days). In total, there were 1458 samples in the training set and 732 in the validation set.

The second step was building an LSTM model to predict daily travel demand. Compared with traditional neural networks, the LSTM network has two significant improvements in its architecture design to reduce information loss for long-term learning. First, the predicted output of the previous timestep is the input of the next. This applies to all RNNs. Second, the memory unit in an LSTM model stores previous information in a gated neuron with an internal state, which can be retrieved and updated in the future learning process even if the two processes are distant. Figure 4 shows the structure of the LSTM network and the memory unit in our study. As is shown in Figure 4a, at each time t, values of the four predictors are parametrized into a 3D matrix X (samples, timesteps, and features) as the input layer ready to insert to the LSTM layer. In particular, the 2D matrix of one sample at time t consists of six observations with four features (each day has four features observed in six destinations). It is worth noting that in our LSTM model, the “timestep” is not the actual 1-day time change. Instead, it is the six steps needed to iterate over all destinations. Then, the LSTM layer makes the prediction for time t by passing the corresponding input matrix through an input gate (node I in Figure 4b) to the LSTM neurons and exporting the result through an output gate (node O in Figure 4b). After predicting at time t, the model passes on the prediction to the next prediction for time (t + 1) as part of the input matrix. Meanwhile, the learning information, such as the weights of predictors, is saved in the internal state cell. Since remembering long-term information can be computationally expensive, the memory unit has a forget gate (node F in Figure 4b) to control whether the information is relevant enough to keep or to need to be updated by more recent learning experiences.

Figure 4.

Long short-term memory (LSTM) recurrent neural network (RNN) model structure: (a) LSTM RNN architecture and (b) LSTM memory unit design.

Choosing appropriate values for the hyperparameters is an essential step in model building. In particular, in our LSTM model, we considered three parameters: number of neurons, number of epochs, and the batch size. Just as with neurons in our brains, the number of neurons in a neural network’s hidden layer determines how much information a computer program can receive, process, and transmit. Too few neurons will receive inadequate information in a complex data set and result in underfitting, while too many will cause overfitting problems and significantly increase the training time. The number of epochs defines the number of times that an algorithm runs through the entire training set. A batch can be regarded as a group of samples with one or more rows of data that an algorithm iterates and optimizes internal parameters for during each epoch. The size of epochs in a model and the batch size in an epoch jointly determine the program’s complexity, which further affects the accuracy and computational cost for making predictions. Based on the pre-analysis of the computational cost versus the accuracy, we decided to create 10 neurons in one hidden LSTM layer and run 100 epochs of training with a batch size of 6.

After building the structure, the final step was training and testing the model. We first normalized the input data to the range of 0–1 to avoid possible errors caused by mismatched scales of predictors. Then, we trained the model with the default sigmoid activation function for LSTM memory blocks using the training data set. After 100 epochs of model training, we made predictions on the test data set using the trained model. We defined the optimized state of an epoch in both stages by minimizing the mean squared error (MSE). The MSE takes values from 0 to 1 in which 0 means perfect fit. Finally, we used root mean square error (RMSE) and mean absolute error (MAE) in both the training and testing stages to evaluate the model accuracy. The LSTM model architecture was coded in Python using open-source codes in the Keras library ( 39 ). The modeling and evaluation process was conducted on a desktop computer with Intel^® Core^™ i5-9600K 3.70 GHz CPU, 16 GB RAM, and Nvidia GeForce RTX 2070 GPU.

Results

Travel Demand Forecast using LSTM Network

In total, the elapsed time of the program was approximately 120 s. Appendix A4 shows the loss values after each epoch for the training data and validation data. As we can see in the figure, both lines converged to a low level of loss values (<0.02) during the model training, indicating the model was precise. In addition, we can see that the model avoided overfitting the data in that the validation line is continuously decreasing (no degradation of model accuracy) and above the training line.

Figure 5 depicts the year-long predictions (dashed red lines) of the daily number of trips to different destinations vis-a-vis the actual values (solid blue lines). As we can see in the figure, in general, the model captured the fluctuation of travel demands to different locations well for the whole year. In particular, the prediction results during the training stage (January to August) seemed to be better than the testing stage (September to December), which makes sense for all unsupervised learning. Moreover, the result reveals the discrepancies of the prediction performance among different destinations in different stages. In the next section, we will evaluate the model performance in the two stages accordingly.

Figure 5.

Observed and predicted daily travel demand, Austin, Texas, 2019.

LSTM Model Result Evaluation

To answer the first research question, we compared the model accuracy in different stages for six destinations, as is shown in Table 2. To simplify the illustration, we summarize the results using the RMSE as the primary accuracy indicator. In general, the prediction accuracy for most destinations was low and stable between stages. In particular, the predicted trip volume to the inner-ring POI cluster was the most accurate (RMSE 131.56). It is probably because, as a center for leisure purposes near downtown, people were likely to go there more regularly. However, Figure 5 shows that the observed daily visits to the inner-ring POI cluster were fewer than the other two clusters on most days, indicating that people’s preference for traveling there for leisure purposes was the lowest. Thus, travel demand for the inner-ring POI cluster was more stable and predictable. On the contrary, the suburban and the urban-fringe POI cluster were more popular, but the traffic was more fluctuant over time. As a result, their prediction errors were slightly larger (RMSE 177.43 and 196.60, respectively). As another single-function destination, the airport acquired the second-best prediction result (RMSE 159.32) because of the stable daily travel demand.

Table 2.

Long Short-Term Memory Model Accuracy for Different Destinations, Sorted by Testing RMSE

Destination	Root mean square error (RMSE)			Mean absolute error (MAE)
Destination	Training	Testing	ΔRMSE	Training	Testing	ΔMAE
Inner POI	85.99	131.56	45.57	78.22	112.85	34.63
Airport	140.90	159.32	18.42	117.10	147.11	30.01
Suburban POI	159.07	177.43	18.37	126.20	167.34	41.14
Fringe POI	169.94	196.60	26.66	118.35	128.42	10.08
Downtown	245.85	282.16	36.31	163.79	241.14	77.35
University	338.58	707.37	368.78	270.16	650.68	380.52
Overall	206.76	339.81	133.04	145.64	241.26	95.62

Note: POI = point-of-interest.

In contrast, since people visited downtown for various reasons, the traffic demand tended to be more fluctuant. There was a possibility that the model accounted for some commute downtown trips as a fraction of the demand on non-working days, making the model less accurate for downtown (RMSE 282.16) than for single-function destinations. Lastly, the university had the worst prediction accuracy (RMSE 707.37), although it was a single-function destination. The result is not surprising in that the model significantly underestimated the traffic surge caused by the opening of the fall semester in September. From Figure 5, we can observe the peaks in September were much higher than those during the spring semester. As a result, the model failed to gain enough knowledge from the past to predict the unusual surges.

To answer the second research question, we further compared the prediction error of travel demands for different locations on different occasions in the testing stage (Figure 6). The result shows that prediction errors under some circumstances were tightly clustered near zero, while others were more dispersed or even deviated from the zero reference line. The randomness of prediction errors increases the complexity of the model interpretation. To simplify the interpretation process, we define a case of demand underestimation as the box being entirely on the reference line’s left side. Correspondingly, a case of demand overestimation is the box being entirely on the right side.

Figure 6.

Prediction error of travel demand for different destinations: (a) weekend; (b) holiday; and (c) school session.

In particular, the LSTM model tended to overestimate the travel demand for downtown on holidays and non-school days compared with their counterparts. As for the university, although the prediction errors are the largest and the most dispersed, we can see from Figure 6b and 6c that the model underestimated the demand on most school days by some way while overestimating it on non-school days. Finally, since the airport’s travel demand was relatively stable, the airport has predicted travel demands that were generally accurate with a slight degree of under- or overestimation.

Interestingly, when comparing the prediction errors of travel demand for POI clusters in different locations at different times, we saw different results. First, although the weekend travel demand for the inner-ring POI cluster was underestimated, and although the weekday demand was overestimated, the errors in both cases were insignificant. Similarly, the prediction result for the suburban POI cluster was relatively accurate except on holidays, when people traveled there significantly less than the model predicted. As for the POI cluster on the city’s northern periphery, the model tended to underestimate people’s travel intention to go there on weekends yet overestimate it on weekdays and holidays. In the next section, we summarize the main findings from the analyses and provide explanations to justify them.

Discussion

In this research, we used the LSTM neural network in deep learning to predict the travel demands for different urban districts in the city of Austin. In addition, we evaluated the impacts of destination types on prediction accuracy at different times. This study showed that as a built environment factor, the endogenous attributes of a destination, namely the urban functionality and the regional location, have inconsistent predictive power over space and time when forecasting people’s travel behavior.

First, this study found that the predictive power of a destination is affected by its urban functionality. The most significant evidence was found in the comparison between the university case and the downtown case. The model underestimated the travel demand to the university by a large amount because it did not anticipate the traffic surge caused by fall semester opening. In contrast, the model overestimated the travel demand to downtown on non-working days because it counted commuting trips and leisure trips together. The conclusion in this case is that destinations with different urban functions have inconsistent predictive power in different time periods because people conduct different activities in different places at different times.

Second, this study found that even if the destination type is the same, the regional location can affect the predictive power. In particular, the inner-ring POI cluster was more predictable than the suburban cluster, and the urban-fringe cluster was the hardest to predict. This predictability discrepancy is related to the locational difference because destinations farther away from downtown have longer system travel distances. As a result, the likelihood of people going there is lower if negative externalities make traveling difficult. Moreover, by comparing the prediction errors in different periods, we also found cases of under- or overestimation for the three clusters. The conclusion in this case is that destinations with the same urban function also have inconsistent predictive power at different times because of their various regional locations.

Conclusion

Forecasting travel behavior by querying physical built environments is critical in many transportation fields such as traffic management and active transportation planning ( 19 , 40 ). The advancement of technology boosts people’s mobility as well as the ways of understanding the advancing travel behaviors using big-data technology. Transportation researchers have conducted a handful of novel cross-sectional studies using machine learning approaches to discover the nonlinear relationship between transportation phenomena and the physical environment and verify that the latter’s impact is threshold-based ( 41 – 43 ). Building on this effort, this paper further sheds light on the possibility that the nonlinear relationship may also vary over time. Results show that the endogenous nature of a place significantly influences people’s travel demand at different times. Thus, the influence of destination attributes in a forecast model is also temporally inconsistent.

The paper has several limitations. First, the journeys in this study were identified by GPS signals from smart devices. The geographic accuracy of GPS is an unknown but a critical precondition to achieve satisfying model accuracy. However, we believe that extensive data covering a long time range could mitigate expected errors in individual observations. Nevertheless, it is more desirable to have more accurate GPS records. Second, by sampling from smart devices, travel demands from certain demographic groups that are less likely to possess smartphones may be overlooked. This limits the representativeness of most studies using big urban data ( 12 , 44 , 45 ). For example, according to a digital inclusion survey in 2018, the smartphone penetration rate in Austin was 78%, and the rate among some vulnerable demographic groups was lower than others ( 46 ). The implication of this is that we could leave out some of the population during modeling of the majority. We should not expect that the result is reliable across different demographic populations. Third, this study did not discuss the influence of origins on travel demand for different destinations. The underlying connections between origins and destinations may be at work. For example, people living near a desirable cluster of restaurants and shopping stores would like to go there more frequently than to other clusters. Future research should incorporate more predictors concerning origin attributes, such as population in adjacent neighborhoods and accessibility from different origins.

Notwithstanding, this study makes substantial contributions to the literature of travel behavioral research. First, using a novel dataset, this study reaffirms that deep learning approaches like RNN have the ability to implicitly model the complicated relationship between travel behavior and built environment in larger geography. Besides, the findings of this study suggest that the heterogenous nature of different urban districts contributes more to the variations of the traffic demand of these places. Second, in the history of travel behavioral research, many scholars have attempted to generalize the relationship with numerous case studies ( 47 , 48 ). They have pushed the research boundary from linear relationships to nonlinear ones. However, few have added temporal attributes as a third element in the equation and discussed how the relationship might differ when predicting the future. This study fills the gap in this regard. The LSTM model in this study shows the heterogeneous impacts of destination attributes on predicting people’s travel demand over space and time.

Future research on travel demand forecasts using big data could focus on the following aspects. First, the presented study results reveal that forecasting unusual traffic surges from limited training samples is difficult, even using successive learning algorithms. To better capture the unusualness, forecasters should fully grapple with the profound opportunities in the big-data era and utilize more years of data from various sources ( 49 ). Second, even though the predictability of deep learning models is better than classic statistical models, whether the improvement is statistically significant and cost-effective in practice is still up for debate ( 50 ). For instance, it should be noted that the model in this study relied on the availability of total trips, as did the models in previous studies. When the information package is not complete, traditional trip-based econometric models might be more efficient in practice ( 51 ). In light of this, this study calls for transportation scholars to contribute more empirical studies discussing the advantages and disadvantages of using emerging big-data applications in travel demand forecast in practice.

Supplemental Material

sj-docx-1-trr-10.1177_0361198121994582 – Supplemental material for Using Deep Learning to Understand Travel Demands in Different Urban Districts

Supplemental material, sj-docx-1-trr-10.1177_0361198121994582 for Using Deep Learning to Understand Travel Demands in Different Urban Districts by Shunhua Bai and Junfeng Jiao in Transportation Research Record

Footnotes

Acknowledgements

The authors would like to thank SafeGraph for sharing their data with the academic community.

Author Contributions

The authors confirm contribution to the paper as follows: study conception and design: SB; data collection: SB; analysis and interpretation of results: SB, JJ; draft manuscript preparation: SB, JJ. All authors reviewed the results and approved the final version of the manuscript.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was supported by the U.S. DOT Cooperative Mobility for Competitive Megaregions University Transportation Center and Good Systems, a research grand challenge at the University of Texas at Austin.

Supplemental Material

Supplemental material for this article is available online.

References

Marshall

Grady

Travel Demand Modeling for Regional Visioning and Scenario Analysis. Transportation Research Record: Journal of the Transportation Research Board, 2005. 1921: 44–52.

Oppenheim

Urban Travel Demand Modeling: From Individual Choices to General Equilibrium. Wiley, New York City, NY, 1995, P. 480. https://www.cabdirect.org/cabdirect/abstract/19951809327.

McNally

M. G.

Kulkarni

Assessment of Influence of Land Use–Transportation System on Travel Behavior. Transportation Research Record: Journal of the Transportation Research Board, 1997. 1607: 105–115.

Ashok

Ben-Akiva

M. E.

Estimation and Prediction of Time-Dependent Origin-Destination Flows with a Stochastic Mapping to Path Flows and Link Flows. Transportation Science, Vol. 36, No. 2, 2002, pp. 184–198.

LeSage

J. P.

Pace

R. K.

Spatial Econometric Modeling of Origin-Destination Flows. Journal of Regional Science, Vol. 48, No. 5, 2008, pp. 941–967.

Jiao

Bischak

Hyden

The Impact of Shared Mobility on Trip Generation Behavior in the US: Findings from the 2017 National Household Travel Survey. Travel Behaviour and Society, Vol. 19, 2020, pp. 1–7.

Bai

Jiao

Dockless E-Scooter Usage Patterns and Urban Built Environments: A Comparison Study of Austin, TX, and Minneapolis, MN. Travel Behaviour and Society, Vol. 20, 2020, pp. 264–272.

Farber

Morang

M. Z.

Widener

M. J.

Temporal Variability in Transit-Based Accessibility to Supermarkets. Applied Geography, Vol. 53, 2014, pp. 149–159.

Golob

T. F.

Meurs

A Structural Model of Temporal Change in Multi-Modal Travel Demand. Transportation Research Part A: General, Vol. 21, No. 6, 1987, pp. 391–400.

10.

Jiao

Bai

Understanding the Shared E-Scooter Travels in Austin, TX. International Journal of Geo-Information, Vol. 9, No. 2, 2020, pp. 135–147.

11.

McNally

M. G.

The Four Step Model. Institute of Transportation Studies, University of California, Irvine, CA, 2000.

12.

Hashem

I. A. T.

Chang

Anuar

N. B.

Adewole

Yaqoob

Gani

Ahmed

Chiroma

The Role of Big Data in Smart City. International Journal of Information Management, Vol. 36, No. 5, 2016, pp. 748–758.

13.

Lim

Kim

K.-J.

Maglio

P. P.

Smart Cities with Big Data: Reference Models, Challenges, and Considerations. Cities, Vol. 82, 2018, pp. 86–99.

14.

Chen

Susilo

Liu

Wang

The Promises of Big Data and Small Data for Travel Behavior (Aka Human Mobility) Analysis. Transportation Research Part C: Emerging Technologies, Vol. 68, 2016, pp. 285–299.

15.

Nguyen

Kieu

L.-M.

Wen

Cai

Deep Learning Methods in Transportation Domain: A Review. IET Intelligent Transport Systems, Vol. 12, No. 9, 2018, pp. 998–1004.

16.

Karlaftis

M. G.

Vlahogianni

E. I.

Statistical Methods versus Neural Networks in Transportation Research: Differences, Similarities and Some Insights. Transportation Research Part C: Emerging Technologies, Vol. 19, No. 3, 2011, pp. 387–399.

17.

Pendyala

Goulias

Time Use and Activity Perspectives in Travel Behavior Research. Transportation, Vol. 29, 2002, pp. 1–4.

18.

Sana

Castiglione

Cooper

Tischler

Using Google’s Passive Data and Machine Learning for Origin-Destination Demand Estimation. Transportation Research Record: Journal of the Transportation Research Board, 2018. 2672: 73–82.

19.

Milne

Watling

Big Data and Understanding Change in the Context of Planning Transport Systems. Journal of Transport Geography, Vol. 76, 2019, pp. 235–244.

20.

Tao

Cao

Fan

Das

Ramaswami

Exploring the Nonlinear Relationship between the Built Environment and Active Travel in the Twin Cities. Journal of Planning Education and Research, 2020, p. 0739456X20915765. https://doi.org/10.1177/0739456X20915765.

21.

Chen

Gao

Yang

Chen

How to Make Dockless Bikeshare Good for Cities: Curbing Oversupplied Bikes. Transportation Research Record: Journal of the Transportation Research Board, 2019. 2673: 618–627.

22.

Clifton

K. J.

Singleton

P. A.

Muhs

C. D.

Schneider

R. J.

Development of Destination Choice Models for Pedestrian Travel. Transportation Research Part A: Policy and Practice, Vol. 94, 2016, pp. 255–265.

23.

Neural Networks Based Real-Time Transit Passenger Volume Prediction. Proc., 2009 2nd International Conference on Power Electronics and Intelligent Transportation System (PEITS), Shenzhen, China, IEEE, New York, 2009, pp. 303–306. http://ieeexplore.ieee.org/document/5406782/.

24.

Çetiner

Sari

Borat

A Neural Network Based Traffic-Flow Prediction Model. Mathematical and Computational Applications, Vol. 15, No. 2, 2010, pp. 269–278.

25.

Rumelhart

D. E.

Hintont

G. E.

Williams

R. J.

Learning Representations by Back-Propagating Errors. Nature, Vol. 323, No. 6088, 1986, pp. 533–536.

26.

Huang

Song

Hong

Xie

Deep Architecture for Traffic Flow Prediction: Deep Belief Networks with Multitask Learning. IEEE Transactions on Intelligent Transportation Systems, Vol. 15, No. 5, 2014, pp. 2191–2201.

27.

Gers

Schmidhuber

Cummins

Learning to Forget: Continual Prediction with LSTM. Neural Computation, Vol. 12, No. 10, 2000, pp. 2451–2471.

28.

Hochreiter

Schmidhuber

Long Short-Term Memory. Neural Computation, Vol. 9, No. 8, 1997, pp. 1735–1780.

29.

Khan

S. M.

Dey

Chowdhury

Development and Evaluation of Recurrent Neural Network-Based Models for Hourly Traffic Volume and Annual Average Daily Traffic Prediction. Transportation Research Record: Journal of the Transportation Research Board, 2019. 2673: 489–503.

30.

Tian

Pan

Predicting Short-Term Traffic Flow by Long Short-Term Memory Recurrent Neural Network. Proc., 2015 IEEE International Conference on Smart City, Chengdu, China, IEEE, New York, 2015, pp. 153–158.

31.

Zhao

Chen

P. C. Y.

Liu

LSTM Network: A Deep Learning Approach for Short-Term Traffic Forecast. IET Intelligent Transport Systems, Vol. 11, No. 2, 2017, pp. 68–75.

32.

Orsini

Gastaldi

Mantecchini

Rossi

Neural Networks Trained with WiFi Traces to Predict Airport Passenger Behavior. Proc., 2019 6th International Conference on Models and Technologies for Intelligent Transportation Systems (MT-ITS), Cracow, Poland, IEEE, New York, 2019, pp. 1–7.

33.

Song

Daily Traffic Flow Forecasting through a Contextual Convolutional Recurrent Neural Network Modeling Inter-and Intra-Day Traffic Patterns. IEEE Transactions on Intelligent Transportation Systems, 2020, pp. 1–10.

34.

Yao

Tang

Wei

Zheng

Revisiting Spatial-Temporal Similarity: A Deep Learning Framework for Traffic Prediction. Proc., 33rd AAAI Conference on Artificial Intelligence, Palo Alto, CA, 2019, pp. 5668–5675.

35.

National Weather Service Forecast Office. National Oceanic and Atmospheric Administration Online Weather Data. Austin/San Antonio, Texas, 2020. https://w2.weather.gov/climate/xmacis.php?wfo=ewx.

36.

Esteban

Staeck

Yang

Tresp

Predicting Clinical Events by Combining Static and Dynamic Information using Recurrent Neural Networks. arXiv:160202685 [cs], 2016. http://arxiv.org/abs/1602.02685.

37.

U.S. Office of Personnel Management. Federal Holidays. US Office of Personnel Management, 2019. https://www.opm.gov/policy-data-oversight/pay-leave/federal-holidays/.

38.

Office of the Registrar. Academic Calendar Archive. University of Texas, Austin, 2020. https://registrar.utexas.edu/calendars/archive.

39.

Chollet

F. Keras

. GitHub. Keras, 2015. https://github.com/fchollet/keras.

40.

Buliung

R. N.

Kanaroglou

P. S.

Activity–Travel Behaviour Research: Conceptual Issues, State of the Art, and Emerging Perspectives on Behavioural Analysis and Simulation Modelling. Transport Reviews, Vol. 27, No. 2, 2007, pp. 151–187.

41.

Ding

Cao

X. J.

Næss

Applying Gradient Boosting Decision Trees to Examine Non-Linear Effects of the Built Environment on Driving Distance in Oslo. Transportation Research Part A: Policy and Practice, Vol. 110, 2018, pp. 107–117.

42.

Ding

Chen

Jiao

Non-Linear Effects of the Built Environment on Automobile-Involved Pedestrian Crash Frequency: A Machine Learning Approach. Accident Analysis & Prevention, Vol. 112, 2018, pp. 116–126.

43.

van Wee

Handy

Key Research Themes on Urban Space, Scale, and Sustainable Urban Mobility. International Journal of Sustainable Transportation, Vol. 10, No. 1, 2016, pp. 18–24.

44.

Hard

E. N.

Chigoy

B. T.

Songchitruksa

Farnsworth

S. P.

Borchardt

D. W.

Green

L. L.

Comparison of Cell, GPS, and Bluetooth Derived External O-D Data: Results from the 2014 Tyler, Texas Study. Proceedings of the 96th Annual Meeting of the Transportation Research Board, Washington, D.C., 2017. https://trid.trb.org/view/1439253.

45.

Townsend

Zambrano-Barragán

Big Urban Data: A Strategic Guide for Cities. Inter-American Development Bank, 2019. https://publications.iadb.org/en/big-urban-data-strategic-guide-cities.

46.

Austin. 2018Austin Digital Assessment. 2018. https://data.austintexas.gov/stories/s/Devices/wjkx-wkvk/.

47.

Ewing

Cervero

Travel and the Built Environment: A Meta-Analysis. Journal of the American Planning Association, Vol. 76, No. 3, 2010, pp. 265–294.

48.

Transportation Research Board, and Institute of Medicine. Special Report 282: Does the Built Environment Influence Physical Activity? Examining the Evidence. Transportation Research Board of the National Academies, Washington, D.C., 2005.

49.

Z.-C.

Huang

H.-J.

Yang

Fifty Years of the Bottleneck Model: A Bibliometric Review and Future Research Directions. Transportation Research Part B: Methodological, Vol. 139, 2020, pp. 311–342.

50.

Wang

Zhang

Liu

Dai

Lee

L. H.

Enhancing Transportation Systems via Deep Learning: A Survey. Transportation Research Part C: Emerging Technologies, Vol. 99, 2019, pp. 144–163.

51.

Miller

H. J.

Chapter 5: Theories and Models. In The Geography of Urban Transportation ( Giuliano

Hanson

, eds.), 4th ed. New York City, NY, Guilford Press, 2017, pp. 113–138.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

1.00 MB