Question:
Task: Vehicular Data Mining (activity analysis)
Background:
The mining of vehicular data has numerous practical applications such as in traffic control, planning of road networks, flow prediction, intelligent driver navigation, and many more. Vehicular Data Mining is one of the fundamental data mining problems which focuses on discovering patterns by analysing vehicular data (i.e. location data). One popularly used dataset in this context is the DRAWDAD roma taxi driver dataset. The dataset is available from: http://crawdad.org/roma/taxi/20140717/
Several studies have been conducted on this dataset. A good starting point to understanding Vehicular Data Mining can be found in:
Cristian Chilipirea, Andreea Petre, Ciprian Dobre, Florin Pop, Fatos Xhafa. Enabling Vehicular Data with Distributed Machine Learning. In Transactions on Computational Collective Intelligence XIX, Vol. 9380, pp. 89-102, 2015.
Definition of the task:
Use the CRAWDAD roma taxi driver dataset for activity analysis of taxi drivers in Rome. Find how the rate of activity varies over time. Questions to be answered are:
- Discover interesting cases of work ethics among the drivers. Describe how the activity pattern of those drivers deviate from the norm.
- Assume that you are to make recommendations to the CEO of the taxi company on how to improve the taxi services in Rome. Your recommendation should be based on the results of your data mining method and it should include additional information from external sources (i.e. roadmap of rome, population density information, information about other forms of public transports, location of relevant facilities and commercial entities, etc.)
Requirements:
- Present a general description of the dataset and present the general properties of the dataset.
- Propose and compare two different approaches to answering the mentioned questions. Discuss the strengths and weaknesses of the two approaches.
- Present, explain, and analyse your results. Offer a qualitative comparison between the two approaches. How did the result of the two approaches differ?
- Summarize: What new and interesting things did you discover?
Answer:
1 Introduction
Vehicular data consists of all measurements that are executed and generated on
the cars that participate in traffic. It can represents a vast amount of information
about the cities we live in. As a society, we have reached a point of very high
mobility; we all travel long distances to get to our jobs, to our schools, to our
leisure activities. All this movement represents the life of a city. We do have
the individual tools to measure what we are doing, what our destination is and
what is the fastest way to get to it, but when we look at large groups of people
participating in traffic we do not really understand what is happening, what are
the common flows or patterns that people follow, how we make use of our roads.
This is how we can use Vehicular Data, to understand the dynamics of our cities
and to try and improve them.
The types and the complexity of data are constantly growing. Until recently
only external data generators were used, like sensors on the road that would be
able to detect the amount of used capacity of the road at a given time. These are
complex systems that are difficult to build as well as expensive. Now, with the
growth in the number of wearable devices and vehicles with on board computers,
data can be generated from inside the traffic flow and at a higher velocity then
we were ever able to. A simple smartphone has numerous sensors such as GPS
or accelerometer, that can be used to gather data about a car: “Where is it?”,
“What speed does it have?” and even “Where is it going?”.
With the vast number of traffic participants and the large number of people
that own a smartphone, or a car capable of sending real time information about
itself, Vehicular Data is growing at a large pace. Public repositories with traffic
data already are available from projects such as T-Drive [22], Cabspotting [13],
and taxicabs [7]. The data being collected can lead to very large volumes, making
the process of extracting information out of it rather complex. Popular processing
methods rely on Machine Learning algorithms. Currently an important topic of
research, such algorithms provide an efficient way to analyze large amounts of
data. The complexity of processes that stand behind traffic flow is so large,
that only data mining algorithms – from the domains of structure mining, graph
mining, data streams, large-scale and temporal data mining – may bring efficient
solutions for these problems.
There is not much surprise that the topic of efficient processing of vehicular
data is today quite popular with scientists and practitioners alike. Competitions
like the IEEE ICDM Contest [18] are designed to ask researchers to devise the
Enabling Vehicular Data with Distributed Machine Learning 91
best possible algorithms that tackle problems of traffic flow prediction, for the
purpose of intelligent driver navigation and improved city planning. Projects
such as iDiary [14] develop advance algorithms to filter/derive/infer information
out of this Big vehicular Data. However, even if the vast majority of work
concentrate on designing the best solutions to deal with information extraction,
little work has been done towards optimizing the data mining process itself.
In this article, our first contribution consists in showing how using a basic,
simple Machine Learning algorithm, k-Nearest Neighbors, we are able to drastically
improve the processing of information about traffic inside a city (Sect. 3).
We next show how the ML algorithm can be distributed over multiple machines,
and how it scales with the number of processors it uses (Sect. 4). We present
in Sect. 5 a discussion about why we need to make this systems truly scalable
and why we should allocate researching resources into these problems. Section 6
presents our conclusions.
2 Related Work
The importance and variety in uses of Vehicular Data and the use of Machine
Learning to optimize traffic, is demonstrated by various authors [21,23,25]. Pau
and Tse [26] present a solution where vehicles are used to measure air parameters
as well as urban traffic. The data generated is then used to better understand
the pollution levels in cities such as Macao. Cars represent a big polluter and the
more information we have about cars the better we can understand their impact
on pollution. Fu et al. [15] present an assessment of vehicular pollution using data
about the estimated number of cars. This type of work can enormously benefit
from accurate vehicle usage data that could be extracted from large vehicular
data sets.
Safar [28] presents the use of vehicular data to answer kNN queries. The
author tries to find the k nearest objects to a location on a map. In addition,
vehicular data is used to improve the response time of these queries. We note
this work because of the use of both vehicular data and the k-Nearest Neighbors
(k-NN) algorithm (in fact, the work is among the first to make use of the k-NN
algorithm in conjunction with vehicular data). However, the way in which they
apply the later, and the purpose, are vastly different than our work.
Processing of vehicular data is commonly done with statistical techniques.
Williams and Hoel [31] present a solution where the data is provided by a road
control agency (the highway agency), and the processing is designed to extract
weekly or seasonal patterns in the usage of the analyzed highways.
Old systems (see Hsieh et al. [17] or Coifman et al. [11]) use video feeds
to do vehicle tracking and classification. Such systems can still be used as
Vehicular data generators, but results show they fail to scale well and/or incur
high/prohibitive costs for large-scale adoption (i.e., a lot of cameras need to be
placed around the city, and processing video feeds requires powerful specialized
equipment). Zhu et al. [33] present a system that uses RFID tags on vehicles,
and a city wide network is designed to gather vehicular data. Even though this
92 C. Chilipirea et al.
system is less expensive and more scalable, it still requires large investments in
the network and the willingness of the tracked vehicles to install an RFID tag
on their cars. Systems later evolved to transmit real time data from inside a
vehicle. Chadil et al. [9] present a system that uses a custom board with GPS
and GPRS to enable vehicle localization and tracking.
Even though vehicular data is very noisy, current research is trying to improve
its quality. Schubert et al. [29] present a comparison between the models used
to increase the accuracy of location data. With the high error rate of GPS
receivers, and the importance of geospatial localization for vehicular data, this
type of research is vital for the improvement of such data sets. Brakatsoulas
et al. [8] take a different approach and try to match GPS data to a street map,
doing all the necessary corrections in the process.
The use of machine learning techniques over Vehicular Data has been done
in works like [1], where this method is used to improve traffic signal control
systems by making these systems truly adaptive. However the dataset used is
considerably smaller, it covers only one intersection, while we are looking at city
wide data sets.
Similar to our work, Sun et al. [30] present the use of Bayesian networks to
achieve traffic flow forecasting. Unlike our work, their test set is provided by
the Traffic Management Bureau of Beijing and it is given in vehicles per hour,
a metric that is not always available. Because of this less noisy data set and its
size there was also no need to improve the execution time of their solution.
The need for scalability in the processing of vehicular data is also identified
by Biem et al. [6], where they use a small compute cluster to process GPS data
from taxis and trucks. This data is processed as streams and it is used to create
real-time speed maps over the city.
Zhang et al. [32] present a way to distribute kNN Joins over MapReduce.
Unlike our solution, their work concentrates on Joins on extremely large data
sets, the main algorithm is however similar and they do obtain a speedup of 4
with 20 reducers. A different solution to the same problem of kNN Joins is presented
by Lu et al. [20]. The authors also chose to distribute the computation
using the Hadoop MapReduce framework. Another article discussing machine
learning algorithms over Hadoop is [16], where multiple machine learning algorithms
are tested over the Hadoop framework. Similarly, authors in [19] present
a similar solution, Distributed GraphLab, aimed for Cloud applications.
Reducing the execution time of k-nearest neighbors has also been done by
parallelizing it using GPU cores [5]. Their solution can be used in combination
with the solution we present in this paper to achieve even greater processing
speed boosts.
3 Application of ML Algorithms on Traffic Data
The traces we use were obtained from the CRAWDAD public repository [12],
a community resource for archiving wireless data. The website contains a large
Enabling Vehicular Data with Distributed Machine Learning 93
number of traces that can be used in Mobile Ad-Hoc Networking and Vehicular
Mobile Ad-Hoc Networking simulations. We selected two popular vehicular
traces to run our experiments on:
– Roma trace [2];
– San Francisco epfl trac [27].
Table 1 includes the main characteristics of these two traces. San Francisco
is the largest one, with double the number of data items and more than half the
number of taxis being recorded. The traces themselves have also been taken six
years apart, with the more recent one being the San Francisco trace.
Table 1. Trace characteristics.
Roma San Francisco
of cars 316 536
of data items 11.219.955 21.673.309
Start Date 17-May-2008 01-February-2014
End Date 10-June-2008 02-March-2014
Both trace contains location coordinates of taxis (in a format similar to
taxi id, timestamp and location (latitude and longitude)). We chose these two
data sets because they span over a moderately large time period, and they both
include a large number of entries (an aspect particularly important to evaluate
scalability aspects).
We believe that in any type of long time measurement regarding human
activity one should be able to distinguish a day-night pattern. This pattern has
a few main characteristics. First, it is an excellent way to validate the data, and
a lack of such a pattern might indicate that something is wrong with the data
generation/recording process or that the data measures something that is not
directly affected by human behavior.
The data was formatted to take the following header: time of day (broken in
30 min intervals); day of week; speed. The id field was not relevant for our experiments.
We also removed all the data points where speed was lower than 0.5 km/h
for a longer time period (to exclude parked vehicles from our measurements).
The newly formatted data was then processed using the k-Nearest Neighbors
(k-NN) algorithm. The entire data set was used as a training set, and for evaluation
we created a secondary set with 48 time of day intervals and null speed
values. The predicted results, indicative of the k-NN model, are visible in Figs. 1
and 2. In the figure we represented the mean speed computed for the various
times of day (this is much lower than the maximum permitted speed, as there
are always cars driving slowly, as expected in congested cities). Furthermore,
the traces contain data about monitored taxis, which have an unusual driving
behavior compared to normal cars. For example, they slow down when they want
to search for a customer or for a building from which they received a request.
94 C. Chilipirea et al.
In the same traffic conditions we expect that normal cars would introduce a
higher mean speed. But there are still a lot of traffic events that lower the mean
speed: cars have to stop at cross-roads, cars stop to pick up other passengers, and
cars slow down when the driver is searching for a location, emergency vehicles
force drivers to slow down or stop to give priority. All these lower the mean
speed of cars within a city.
For all processing and all graphs we used a k value of 500, and the k-Nearest
Neighbor algorithm was set to calculate the mean speed over all the neighbor
values.
Fig. 1. k-Nearest Neighbors model for speed, Roma.
Figure 1 present the average traveling speed per time of day, obtained for the
Roma trace. We can discern a day-night pattern in the obtained speed model.
During the day, the speed lowers considerably compared to nighttime. This happens
mainly because of high traffic during the day, with a lot of cars stuck in
traffic – the mean speed is lowered for all the cars that participate in said traffic.
This model is also a good predictor of what mean speeds to expect at a certain
hour, or when it is best to make a trip so that you can achieve maximum speed.
The figure is averaged for the entire city, but using the same principle, we can
and did obtain data for different parts of the city, or even with fine granularity
we can reach the street level.
The San Francisco model (Fig. 2), shows a similar day-night pattern. This
model also gave us some previously unexpected information. We are able to
clearly identify the “rush hours”, moments during the day where there are so
many cars in traffic that the mean speed is lowered by a large factor. In the mode
we can observe both the morning and the afternoon “rush hour” moments, where
the mean speed is at its minimum. After a closer inspection of the Roma model,
we identified the same drops in mean speed at similar time intervals.
What we did not expect was that when we overlap the two models, we would
find a relatively close correlation between them. In Fig. 3 we plotted the overlapped
data. Here, most of the raises and drops in the mean speed almost align.
Enabling Vehicular Data with Distributed Machine Learning 95
Fig. 2. k-Nearest Neighbors model for speed, San Francisco.
Considering the day-night cycle in mean speed is a sort of a “heart-beat” of the
city, we find it extremely interesting that two very different cities in different
parts of the world with different cultures, and data gathered years apart, have
such a similar “heart-beat”.
We note that the mean speed over the entire day in San Francisco is higher
than the mean speed of the Roma trace. We believe this to caused because of
geography; the city of Roma has hills, unlike San Francisco, and because it is an
older city, dating back to the start of the Roman empire, the streets are not as
wide, permitting less traffic, at lower speeds.
Fig. 3. k-NN San Francisco/Roma model comparison.
96 C. Chilipirea et al.
Fig. 4. The city of Roma as obtained from Open Street Maps [24].
For the Roma trace we made another analysis. We wanted to extract density
maps and see how these change as effect of the day-night difference.
We divided the map into 10.000, or 100 by 100, sub-regions. Then, for each
sub-region, we counted the number of cars between the hours 01:00 and 07:00
for the night, and between the hours 14:00 and 20:00 for the day, for each day
individually. This data was then processed using the k-NN algorithm to extract
a mean value for each sub region. The results can be seen in Figs. 5 and 6. In
these figures, a black color means a high number of cars, while a white color
means a smaller number of cars, even 0.
It is very clear that during the day the car density is higher in the center
of the city. By comparing these images with Fig. 4, the map of the city, we can
see how the cars follow the main city roads. We note that using this analysis we
were able to map most of the roads in Roma using only GPS data from taxis.
4 Distributing k-Nearest Neighbors
Running k-Nearest Neighbors over multiple machines can be done in multiple
ways. We chose to use the Massage Passing Interface (MPI) framework. MPI is
a distributed framework (and a message passing library interface speci¨ıˇnAcation
for parallel programming) mostly used in high performance computing clusters.
MPI is appropriate for data-dependent iterative algorithms based on message
passing. People have long opposed MPI to MapReduce (the next widest used
model/framework for parallel/distributed ML algorithms – Chen et al. [10] make
Enabling Vehicular Data with Distributed Machine Learning 97
Fig. 5. Car density during the night.
Fig. 6. Car density during the day.
98 C. Chilipirea et al.
an advanced comparison of the two), which is most suitable for non-iterative
algorithms without lots of data exchanges. We chose MPI as an alternative
to what we found in the literature about distributing k-NN using the Hadoop
MapReduce framework (see Zhang et al. [32], Lu et al. [20], or Gillick et al. [16]).
To the best of our knowledge, today there is no MPI implementation of k-NN
available.
Our solution reads both the training set and the test set in the first task
(rank == 0), and this data is spread to all other tasks running on the distributed
machines. Each machine then processes its part of the test set.
In k-Nearest Neighbors each element in the test set needs to be compared with
all elements in the training set to determine: which are the k-nearest neighbors.
After all the tasks are finished processing their part of the test set, the predicted
value is calculated for each point in the data set and the data is gathered to the
first task.
During our experiments we used different values for k, and we split the Roma
test set into 80% for the training set and 20% for the test set.
The tests were run on a cluster of IBM Xeon E5630 rated at 2.53 GHz with
32GB or RAM. During the experiments we used between 1 and 10 different such
machines. We note here that each experiment was run 3 times and the results
were very similar, in the graph we show the mean execution time between the 3
different runs.
In Fig. 7 we can observe the execution time for the k-NN algorithm on the
Roma data set. We used k values of 100, 300, 500, 700 and 1000.
Fig. 7. Execution Time for distributed k-NN.
Enabling Vehicular Data with Distributed Machine Learning 99
Fig. 8. Speedup for distributed k-NN.
The speedup for our implementation of the algorithm is almost linear with
the number of compute nodes. This is shown in Fig. 8, which shows the speedup
for all the values of k we used in our experiment. For smaller values of k, the
speedup is insignificant (this is visible for k=100, and degrades for lower values).
5 Discussion
Each experiment in this paper took between 30 min to 1 h to execute on one
single core machine, excluding data parsing or formatting. The traces contained
data for just 1 month for over 500 cars. When we are looking at scaling these
solution to city size we consider that every car is a potential data generator. This
means millions of cars per city each generating one line of data every few seconds.
To make things even worse, some machine learning algorithms, such as neural
networks or convolution neural networks, have even higher execution times than
the one we presented, but they could be used to extract even more interesting
information from the data sets and to make more accurate predictions.
With the extremely large number of data generators and the high execution
time for machine learning algorithms building a complex real time solution can
prove difficult. It is important to find the best ways to use all the resources provided
by a modern computers, executing code on multiple cores, using both CPU
and GPU, as well as enable the algorithms to make use of multiple machines.
Another factor that can raise the complexity of future Intelligent Transportation
System (ITS) applications is the complexity of the generated data
itself. Soon only timestamps and GPS coordinates will not be enough, and more
100 C. Chilipirea et al.
complex sensors will provide a large variety of data. With the fast integration
of smartphones that have powerful photo and video cameras, it is no longer far
fetch to consider that soon one system will have to process video data generated
by all the cars inside a city.
6 Conclusion
In this article we presented our analysis over two distinct vehicular traces, one for
the city or Roma and another for San Francisco, using the k-Nearest Neighbor
algorithm. We chose to use the k-Nearest Neighbors algorithm because it is one
of the simplest (and yet powerful) machine learning tool.
With the analysis we showed that with the use of machine learning algorithms
important information can be extracted from even basic vehicular data sets.
The used data sets only contain time stamp and GPS coordinates (latitude and
longitude) but other data sets may contain data from other sensors such as
microphones, to measure noise levels or chemical sensors that could measure the
level of pollution. With the minimal available data we manage to identify a day
night cycle pattern, rush hours and we manage to build a model that could be
used to predict high traffic intervals and regions with high vehicular density.
Then we showed how machine learning algorithms such as k-Nearest Neighbors
can exploit the power of compute clusters or even clouds by reducing execution
time through distributed processing. We achieve this using the popular
MPI framework.
The need for highly scalable machine learning algorithms for use with vehicular
data is further explored in our discussion section.
As future work we believe more machine learning algorithms should be tested
in combination with vehicular data. We believe that with the right combination
of data and algorithm surprising and interesting information could be extracted.
More machine learning algorithms should be parallelized or distributed so
that they can be efficiently executed over multiple machines. A lot of work in
this direction has been done in projects such as Mahout [3] or Spark MLLib [4],
but not all machine learning algorithms have been integrated into these package.
For instance neither contain a k-Nearest Neighbors implementation.
More vehicular traces need to be generated and analyzed. The vehicular
traces we used only contained data from taxis, which made them slightly biased.
A vehicular trace that contains all kinds of vehicles would be extremely interesting
for the research community. It would also be fascinating to have public
traces that contain other sensor data such as pollution or noise levels.
Acknowledgment. This work was supported by the Romanian national projectMobi-
Way, Project PN-II-PT-PCCA-2013-4-0321. The authors would like to thank reviewers
for their constructive comments and valuable insights.
Enabling Vehicular Data with Distributed Machine Learning 101
References
1. Abdulhai, B., Pringle, R., Karakoulas, G.J.: Reinforcement learning for true adaptive
traffic signal control. J. Transp. Eng. 129(3), 278–285 (2003)
2. Amici, R., Bonola, M., Bracciale, L., Rabuffi, A., Loreti, P., Bianchi, G.: Performance
assessment of an epidemic protocol in vanet using real traces. Procedia
Comput. Sci. 40, 92–99 (2014)
3. Apache: Mahout (2015). https://mahout.apache.org/
4. Apache: Spark mllib (2015). https://spark.apache.org/mllib/
5. Barrientos, R., G´omez, J., Tenllado, C., Prieto, M.: Heap based k-nearest neighbor
search on gpus. In: Congreso Espanol de Inform´atica (CEDI), pp. 559–566 (2010)
6. Biem, A., Bouillet, E., Feng, H., Ranganathan, A., Riabov, A., Verscheure, O.,
Koutsopoulos, H., Moran, C.: Ibm infosphere streams for scalable, real-time, intelligent
transportation services. In: Proceedings of the 2010 ACM SIGMOD International
Conference on Management of Data, pp. 1093–1104. ACM (2010)
7. Bracciale, L., Bonola, M., Loreti, P., Bianchi, G., Amici, R., Rabuffi, A.: CRAWDAD
data set roma/taxi (v. 2014–07-17), Jul 2014. Downloaded from http://
crawdad.org/roma/taxi/
8. Brakatsoulas, S., Pfoser, D., Salas, R.,Wenk, C.: On map-matching vehicle tracking
data. In: Proceedings of the 31st International Conference on Very Large Data
Bases, pp. 853–864. VLDB Endowment (2005)
9. Chadil, N., Russameesawang, A., Keeratiwintakorn, P.: Real-time tracking management
system using gps, gprs and google earth. In: 5th International Conference
on Electrical Engineering/Electronics, Computer, Telecommunications and Information
Technology, ECTI-CON 2008, vol. 1, pp. 393–396. IEEE (2008)
10. Chen, W.Y., Song, Y., Bai, H., Lin, C.J., Chang, E.Y.: Parallel spectral clustering
in distributed systems. IEEE Trans. Pattern Anal. Mach. Intell. 33(3), 568–586
(2011)
11. Coifman, B., Beymer, D., McLauchlan, P., Malik, J.: A real-time computer vision
system for vehicle tracking and traffic surveillance. Transp. Res. Part C: Emerg.
Technol. 6(4), 271–288 (1998)
12. Dartmouth: Crowdad (2015). http://crawdad.org/
13. exploratorium: Cabspotting (2015). http://cabspotting.org/index.html
14. Feldman, D., Sugaya, A., Sung, C., Rus, D.: idiary: From gps signals to a textsearchable
diary. In: Proceedings of the 11th ACM Conference on Embedded Networked
Sensor Systems, p. 6. ACM (2013)
15. Fu, L., Hao, J., He, D., He, K., Li, P.: Assessment of vehicular pollution in china.
J. Air Waste Manage. Assoc. 51(5), 658–668 (2001)
16. Gillick, D., Faria, A., DeNero, J.: Mapreduce: Distributed computing for machine
learning. Berkley, 18 Dec 2006
17. Hsieh, J.W., Yu, S.H., Chen, Y.S., Hu, W.F.: Automatic traffic surveillance system
for vehicle tracking and classification. IEEE Trans. Intell. Transp. Syst. 7(2), 175–
187 (2006)
18. IEEE, TomTom: Ieee icdm contest: Tomtom traffic prediction for intelligent gps
navigation (2010). http://tunedit.org/challenge/IEEE-ICDM-2010
19. Low, Y., Bickson, D., Gonzalez, J., Guestrin, C., Kyrola, A., Hellerstein, J.M.:
Distributed graphlab: a framework for machine learning and data mining in the
cloud. Proc. VLDB Endowment 5(8), 716–727 (2012)
20. Lu, W., Shen, Y., Chen, S., Ooi, B.C.: Efficient processing of k nearest neighbor
joins using mapreduce. Proc. VLDB Endowment 5(10), 1016–1027 (2012)
102 C. Chilipirea et al.
21. Mavromoustakis, C.X., Kormentzas, G., Mastorakis, G., Bourdena, A., Pallis, E.,
Rodrigues, J.: Context-oriented opportunistic cloud offload processing for energy
conservation in wireless devices. In: Globecom Workshops (GC Wkshps), pp. 24–
30. IEEE (2014)
22. Microsoft, R.: T-drive: Driving directions based on taxi traces (2015). http://
research.microsoft.com/en-us/projects/tdrive/
23. Mousicou, P., Mavromoustakis, C.X., Bourdena, A., Mastorakis, G., Pallis, E.:
Performance evaluation of dynamic cloud resource migration based on temporal
and capacity-aware policy for efficient resource sharing. In: Proceedings of the 2nd
ACM Workshop on High Performance Mobile Opportunistic Systems, pp. 59–66.
ACM (2013)
24. OpenStreetMap: Openstreetmap (2015). https://www.openstreetmap.org
25. Papadakis, S.E., Stykas, V., Mastorakis, G., Mavromoustakis, C.X., et al.: A hyperbox
approach using relational databases for large scale machine learning. In: 2014
International Conference on Telecommunications and Multimedia (TEMU), pp.
69–73. IEEE (2014)
26. Pau, G., Tse, R.: Challenges and opportunities in immersive vehicular sensing:
lessons from urban deployments. Sig. Process. Image Commun. 27(8), 900–908
(2012)
27. Pi´orkowski, M., Sarafijanovic-Djukic, N., Grossglauser, M.: A parsimonious model
of mobile partitioned networks with clustering. In: First International Communication
Systems and Networks and Workshops, COMSNETS 2009, pp. 1–10. IEEE
(2009)
28. Safar, M.: K nearest neighbor search in navigation systems. Mob. Inf. Syst. 1(3),
207–224 (2005)
29. Schubert, R., Richter, E., Wanielik, G.: Comparison and evaluation of advanced
motion models for vehicle tracking. In: 2008 11th International Conference on
Information Fusion, pp. 1–6. IEEE (2008)
30. Sun, S., Zhang, C., Yu, G.: A bayesian network approach to traffic flow forecasting.
IEEE Trans. Intell. Transp. Syst. 7(1), 124–132 (2006)
31. Williams, B.M., Hoel, L.A.: Modeling and forecasting vehicular traffic flow as a
seasonal arima process: Theoretical basis and empirical results. J. Transp. Eng.
129(6), 664–672 (2003)
32. Zhang, C., Li, F., Jestes, J.: Efficient parallel knn joins for large data in mapreduce.
In: Proceedings of the 15th International Conference on Extending Database
Technology, pp. 38–49. ACM (2012)
33. Zhu, H., Zhu, Y., Li, M., Ni, L.M.: Hero: Online real-time vehicle tracking in
shanghai. In: IEEE The 27th Conference on Computer Communications, INFOCOM
2008. IEEE (2008)
