2018 6th International Conference on Information and Communication Technology (ICoICT)
Route Recommendation using Community Detection Algorithm (Case: Kota Bandung) Yahya Peranginangin
Andrias Andi
Kristina Sisilia
Faculty of Electrical Engineering Telkom University Bandung, Indonesia yahyaperanginangin@telkomuniversity. ac.id
Faculty of Communication and Business Telkom University Bandung, Indonesia
[email protected] .id
Faculty of Communication and Business Telkom University Bandung, Indonesia
[email protected]
Abstract—Bandung public transport route, known as Angkutan Kota (Angkot), has 39 routes to serve intra city transportation. These routes connect more than 40 bus stations in star topology. The centrality analysis of this routes shows that many of them are overlapped into each other in certain location, thus may potentially cause congestion and increased competition for passengers among routes. To achieve operational break event, most fleet tend to conduct activities that reduce its service quality and disturb traffic flow. Having public transport with inconsistent level of service leave passengers with no choice but to find another mode of transportation, including owning private vehicle, which lead public transportation into deeper vicious circle. Some effort to break this circle is to subsidised public transport fare, car restraint (such as 3 in 1 policy), and priority lane for public transport (such as bus way). This paper suggests another way to increase usability of public transport by implementing graph theory to generate new routing strategy. Instead of having path formed from node sequence, we propose nodes that grouped geographically, creating modular routes that cover certain area and can be managed independently. The graph or network is formed by using bus stops/stations as nodes and routes to connect the nodes. We apply Louvain’s community detection algorithm to groups bus stops and stations into communities. By assuming that each community is a route coverage, we can define 14 Angkot routes. The proposed routes feature fewer routes and can significantly reduce overlapping between routes while still having similar performance with conventional route. Keywords—Graph Theory; Community Detection; Public Transport; Route.
I. INTRODUCTION Demand for public transportation service is derived from human activities. The quality of this service would define how people make use space around them. Public transport route is usually defined to be used for years with very minimum change. Modification or optimisation to these routes would require examination of all existing routes in a city [1]. Online sharing transport services seek to fill this gap. Companies such as Uber 1
Data is courtesy of data.go.id. Data is provided by Dinas Perhubungan Kota Bandung.
ISBN: 978-1-5386-4571-0 (c) 2018 IEEE
108
and Grab found opportunity in providing flexible and affordable transportations mode, combining advantages of affordability of public transport and flexibility of private vehicle. Online sharing transport service companies use algorithm and real-time data to predict most efficient route or to allocate their drivers in highdemand regions in order to maximize efficiency of transporting people or goods around a city. In the other side, public transport as mass transport mode still require stations or bus stops as part of its route. Optimising public transport routes would have to take existing bus stops and stations into consideration. The purpose of this paper is to generate alternative public transport routes that is easy to expand but still considering existing bus stops and stations. II. ANGKUTAN KOTA IN BANDUNG Angkutan Kota (Angkot for short) is the most common public transport mode in Bandung. It uses minibus with capacity 12-15 passengers. Kota Bandung has around 152 bus stops and stations. Public transport is dominate by Angkutan kota (Angkot) as it has significantly large number than other mode and it is considered the most affordable transport mode for most people. There are 39 Angkot routes and more than 5500 fleet to serve transportation around the city, excluding intercity routes1. Each route’s distance cover may range from 12 up to 46 km. Some routes are available 24 hours a day. Some routes are only available 16 to 18 hours a day. Usually they start the service from 5 in the morning up to 9 or 11 in the evening. Since early 21st century Angkutan kota has been facing many challenges. Rapid economic growth and new financing regulation made private transportation mode (such as motorcycle) become more affordable. Online transport service (such as Uber or Grab) introduced in early 2010s offered another alternative to conventional public transport. Flexibility and more predicted travel time make private vehicle and online transportation sharing service as good alternatives for busy people. On the other side it reduces demand for Angkot service. Lower demand creates internal competition between Angkot to
get passengers in a way to achieve operational break event point, even if it has to sacrifice service quality, such as predicted travel time, and causing congestion in certain bus stops. This situation creates downward demand spiral or vicious circle for Angkot. Lower service quality drives more passengers to find another transportation mode, even if they have to buy their own vehicle [2]. Public transport become less attractive and more expensive; private car or online sharing transport service become more attractive.
route based on combination of individual choice of trip and another variable such as price, travel time, or taste [5]. The results of these models are not set of routes, instead they help decision maker to decide most effective routes from predefined ones. The routes themselves usually generated by connecting between origin and destination of households or non-households (such as between residence area and commercial/office area). Most routes generated are connecting urban and sub-urban area, as we also found in Kota Bandung’s public transportation routes.
Although the maximum number of fleet allowed by government is more than 5500, the actual number of Angkot that still in service today is already decreased to 30%2, or about 1600 fleet. This is mainly caused by the increasing of private vehicle ownership in Bandung3. By 2016 there are more than 1.7 million of private vehicle in Kota Bandung [3]. Compare to its population, it means there are 683 private vehicles for every 1000 population and 0,64 Angkot per 1000 population. Public transport availability in Kota Bandung is still below developing countries standard for bus transportation planning (1.5 per 1000 population) and Europe standard (1 bus per 1000 population) [4]. As for private vehicle ownership, Kota Bandung can be placed among top ten countries for number of vehicles per 1000 population. To calculate its waiting time, we take several assumptions based on available traffic data. By assuming that each fleet is operated 2 shift, and average speed is 25km/h (average speed target in Kota Bandung when high traffic time) the waiting time between Angkot in a bus stop is 0,53 to 8,5 minutes with average waiting time is 2 minutes. Compare to Singapore MRT’s waiting time (2-3 minutes during peak hours and 5-7 minutes during offpeak times), Angkot’s waiting time is still in reasonable range4.
Fig. 2. Distribution of bus stops and stations in Bandung (top) and how routes connect bus stations creates hub and spoke structure (bottom). Fig. 1. Common attempt to break circle of vicious in public transportation (Hensher and Button, 2008)
III. METHOD There are several modelling approaches to design public transportation route, such as Elasticity Model, Assignment Model, Activity-based approach, Tour-based model, and Tripbased model (Four-Step Model). Those models try to optimise 2
Topology of public transport routes in Bandung follow hub and spoke structure. Several main bus stations (hub) located at urban area connect to smaller stations that surrounded the city. Most hubs are located in commercial or office area. Expansion of city area to east created additional bus stations that also connect to hubs located in urban area. Between stations, Kota Bandung have more than 120 bus stops. Bus stops distribution in Bandung is denser in the city centre and getting less dense in residential area. Fig. 2 (top) shows that there are many residence 4
http://jabar.metrotvnews.com/read/2017/10/19/775746 3 http://www.pikiran-rakyat.com/bandungraya/2014/09/02/295217/tren-penggunaan-angkot-menurun33-trayek-mati
109
https://www.lta.gov.sg/content/ltaweb/en/publictransport/mrt-and-lrt-trains/riding-a-train.html
areas still not covered by public transport service. Additional routes are expected in the future to serve residence area. Adding more routes using existing method is a complex thing to do. Thorough study on existing as well as on new routes should be conducted. New routes most likely will affect existing one and would add load to urban traffic. This paper proposes another topology for public transportation route. We use graph theory to analyse existing public transport route and create more flexible route with less overlapping between routes and can be adjusted modularly for future expansion. On this paper we only consider the network structure of public transport route instead of the structure of actual road network. To generate the network or graph we use bus stop/station and Angkutan Kota’s routes as data source. As this paper focus on exploring new method in generating route, testing it with real or artificial traffic data will be covered in the next paper.
modularity [6]. The metrics above can be interpreted as shown in Table I. TABLE I.
INTERPRETATION OF GRAPH METRICS IN BANDUNG’S PUBLIC TRANSPORTATION
Metric
Degree Centrality
Path Length
IV. ANGKOT BANDUNG IN GRAPH Social Network Analysis (SNA) uses graph theory to quantify the quality relationship between people. We use the same theory to analyse public transport route in Bandung. To build a network, we define bus stops and stations as node and Angkot routes as its edge. We use the route to simplify the network instead of using physical road to connect the nodes. In this paper we use Angkot routes as the network edges for it is the most popular transport mode in Bandung and its large number of fleet compare to Damri buses (5521 Angkots compare to 1464 buses). In the rest of this paper we will use the word “node”, “bus stop” or “bus station” interchangeably.
Diameter
We defined this network as directed and weighted network. The direction of this network is determined by the traffic direction and how an Angkot approach each node sequentially along its route. On the other hand, a bus stop/station as a node can be traversed by more than one route. Every overlapping route that traverses a node will be added to its weight. A node that traverse by more than one route will have higher weight than a node that traverse by only one route.
Modularity
Density
Interpretation How many alternative routes a bus stop has. The higher the degree, the more alternative routes available. The highest weighted degree in this network is 26 (13 in-degree and 13 out-degree centrality). Node with higher degree centrality can be considered as interchange node. How long (in hop or transit) is the average shortest distance between two bus stops in Bandung. The average path length for Angkot Bandung network is 6.3. It means that to reach her destination across the city, a passenger needs to get through 6-7 bus stops in average. How long is the maximum hop/transit taken to reach a destination in Bandung. Diameter of public transport network in Bandung is 18. To reach the longest destination using the shortest path, a passenger requires to make 18 hop/transit. How many connections are there between node compare to all possible connection. Angkot network as in Fig. 3 has density 0.015. It means there are 15 connections in every 1000 possible connections. Low density network tends to result larger diameter; thus, it would take longer distance to reach destination. Tendency to split into communities. Modularity score ranged from 0 to 1. Modularity that closer 0 means the network is easier to split into communities. The non-weighted network modularity is 0,562. The weighted one is 0,634. Considering overlapping routes in public transport network will tend to give bigger size and smaller number of communities.
To measure how many routes are overlapping in a node we use weighted degree centrality, as it is directly related to actual route. We do not use betweenness centrality as it measures the potential of a node to become a connector between nodes. TABLE II. No
Fig. 3. Bus stops network connected by Angkot routes.
We analyse public transport network using SNA metrics: centrality, path distance, density, clustering coefficient, and
110
TOP 10 WEIGHTED DEGREE CENTRALITY In-Degree
Out-Degree
1
Halte GKI Kebon Jati
Node (Bus Stop/Station) Name
13
13
2
Terminal Abdul Muis
13
13
3
Terminal St.Hall
13
13
4
Halte Dewi Sartika
12
12
5
Halte Hotel California
12
12
6
Terminal Cicaheum
11
11
7
Halte Padjajaran BPPK
10
10
8
Halte Padjajaran SMKN12
10
10
9
Halte Pasar Bunga
10
10
10
Halte Jakarta STIMIK
9
9
Based on the total directed degree centrality (in and out degree), number of relation occur on nodes are following power law distribution, similar with the one that usually found at social relationship. If we assume that most of bus stop have similar size and capacity, some nodes have to handle a lot of passengers, but a lot of nodes will not be utilised properly as the passenger rate would be very low. Routes with low rate of passengers tend to use ‘hail and ride’ instead of waiting at designated bus stop [7].
represent road of Kota Bandung. We only took network structure into consideration to generate the new route in this paper.
The problem of route accumulation does not exclusively belong to bus station. Table II above shows that not every top highest degree node is a bus station. Only 3 bus stations are in the top ten list of highest weighted degree. Most important nodes above located in the centre of the city. All those nodes have similar score between in-degree and out-degree. At the densest area, there can be up to 13 route that compete to get passengers. Those nodes canbe used by passenger to transit or change route (interchange nodes). These may add more congestion to the traffic in around the nodes.
∆𝑄 = $
There are about 37 bus station in Bandung. Almost 70% of existing routes start or end in one of only five bus station, i.e. Abdul Muis, St. Hall, Cicaheum, Ciroyom and Dago. If there will be new routes in the future, they would likely use these stations as its destination. Thus, every initiation to expand Angkot route will directly impact urban area and existing route.
We use graph theory to directly generate routes for Angkot. By using existing nodes and routes into consideration, we develop a public transport network. To separate nodes into communities, we use Louvain method [9]. ∑&' ()&,&' +,
−.
∑/0/ ()& + +,
∑
3
+
3
+
/0/ /0/ &' 1 2 − $ +, − . +, 1 − . +, 1 2
(1)
The modularity gain (DQ) of an isolated node in a community can be calculated using (1) where ki is the sum of weights of node i, ki,in is the sum of weights of node I to nodes inside its community, and m is the sum of weights of all nodes in the network. A node will be iteratively attached and detached into and from communities to find maximum gain of its modularity. When a node inserted into its neighbour community and the maximum gain achieved the node will be signed into its neighbour community. To adjust performance of the routes generated, we can change the resolution of the modularity [10].
V. RE-CLUSTERING PUBLIC TRANSPORT NETWORK Initiative to save public transport usually come from government. To increase public preference to public transport, city government may initiate Some of them are in the form subsidy to public transport operator to reduce impact of increasing operating cost and lower demand for bus, some are in the form of creating priority lane for bus and car restraint to increase preference for public transport over private car or online sharing transportation service. A major way to optimise public transport is to introduce new routing system, such as switching existing routing system into grid system for better redundancy or into connective network for simpler routing. Routes generated using community detection algorithm will be in the form of coverage area instead of path formed from the node sequence. Therefore, before separating the network into communities, there are two things that we should take into consideration: the network will be considered as undirected network, and the weight of the nodes need to be disregarded. The network in this paper consists of Angkot routes in the form of closed path that layered together into on graph. Each path is a return path between two nodes. Based on Transformation Maintaining Directionality approach, we can replace the direction of an edge with undirected edge with additional weight [8]. Since the connections between nodes are restricted by traffic flow, thus we do not find any reciprocal connection between two nodes. Therefore, we can use undirected community detection method, such as Louvain method, to analyse this network. As the node’s weight (the weight that given to a node when two or more routes are overlapped) is representing existing routes, disregarding it will give network connection that
111
Fig. 4. Community detection with resolution 1.0 for non-weighted network.
Community detection method divides nodes in a network into groups. Node with more connection among them would likely to be in the same group. Adding weight into the calculation would change the importance of a node in its community. The higher the weight, the more important the node is, and it will attract more of its neighbour into its community. The community formed will be bigger in size and smaller in number. Based on non-weighted modularity, nodes are divided into 8 communities/routes. Each route covers from 10 to 31 bus stops/stations with average 19 bus stops/stations. Compare to existing routes that cover from 2 to up to 33 nodes with average 11 stops, the new routes still have higher average coverage node per route. Reducing resolution of community detection algorithm to 60% resulting in an increase in the number of routes up to 14 routes and average node coverage per route become 10.85, close to existing average node coverage per route. While existing route use 47 routes (8 routes are intercity route), the
proposed routes only use less than 30% routes to cover all the city area.
route only covers certain and localised area. To give better service, government can increase the density of route modularly or by adding more route in as a separate modular.
Fig. 5. Distribution of node for each route. A route covers from 6 up to 15 nodes.
In addition to fewer routes, these new routes also reduce route overlapping significantly. Instead overlapping routes only exist at interchange node, i.e. among neighbours’ communities. This will lead to lower competition between routes and will reduce waiting time in order to fill passenger seats. Although these new routes still have interchange nodes, but such nodes are spread evenly across the city. Thus, its impact to local traffic can be minimised. TABLE III.
COMPARING WEIGHTED AND NON-WEIGHTED DEGREE CENTRALITY TO ILLUSTRATE HOW EFFECTIVE THE PROPOSED ROUTE IN REDUCING BUS STOPS AND STATIONS LOAD. (% REDUCE OF DEGREE CENTRALITY IS CALCULATED BY COMPARING DDEGREE CENTRALITY (WEIGHTED – NON WEIGHTED) WITH ITS ORIGIN DEGREE (WEIGHTED DEGREE CENTRALITY) No
Node (Bus Stop/Station)
Weighted
Non-Weigh.
Reduce
1
Halte GKI Kebon Jati
26
15
35%
2
Terminal Abdul Muis
26
10
62%
3
Terminal St.Hall
26
14
46%
4
Halte Dewi Sartika
24
8
67%
5
Halte Hotel California
24
13
46%
6
Terminal Cicaheum
22
14
36%
7
Halte Padjajaran BPPK
20
4
80%
8
Halte Padjajaran SMKN12
20
5
75%
9
Halte Pasar Bunga
20
10
50%
10
Halte Jakarta STIMIK
18
6
67%
Fig. 6. An example of a route that is not a close route after dividing network using community detection technique.
An optimised route may be in the form of non-cycled network, as seen in fig. 6. But as the network is not based on real road, new edges or nodes may be introduced to create a closed loop. But we also can treat each route differently. While higher load route can have fixed route between nodes, low load route can become more flexible without a fixed path, just like online sharing transportation service allocate its driver in an area [7]. These additions will affect the structure of this community but would only give minimal impact to overall network. Thus, managing route would become much easier. VI. CONCLUSIONS Routing recommendation using community detection technique in this research is considered heuristic method. Community detection method can be used to generate routes based on the structure of city streets. There are many algorithms to separate networks into communities. This paper only demonstrates the usage of Louvain method. Adjusting resolution in modularity, the proposed routes still can achieve similar performance with existing public transport routing in Bandung, in terms of coverage node per route. By using clustered routes, overlapping can be reduced and route can be managed locally with minimum effect to another routes. The needs of new routes caused by more resident and business area in the future will not put extra load into other routes, especially the ones in urban areas. More reliable and more predicted public transport will lead to more preference into public transport usage.
Table III shows comparison of degree centrality between existing routes (weighted degree centrality) and new routes (non-weighted degree centrality). We can measure degree centrality of new routes as non-weighted as we can remove overlapping routes in most area. The comparison can be interpreted as the reduce of traffic load casude by reducing overlapping route alone. By reducing route overlapping alone can reduce traffic load caused by Angkot up to 80%. Adding new routes will give very minimal impact to other routes, as each
112
This new method of generating route can also give a new approach in route experiment. While a high load route can have fixed path, a new route can be introduced in a particular area of Kota Bandung. This new route can very flexible. Using app to assign its route, based on passenger’s request, while collecting data to figure out the most effective route to be implemented in that area. This method is meant to complement existing method used to design public transport route, especially when there is need to analyse large and networked data. This paper only makes use of
simple data from existing Angkot routes and bus stops/stations location. Using more data, such as citizen daily movement or daily traffic, would increase the accuracy of modelling public transport route. REFERENCES [1] Auckland Transport, "New public transport network," Auckland Transport, 2017. [Online]. Available: https://at.govt.nz/projectsroadworks/new-public-transport-network. [Accessed 24 February 2018].
[9] V. D. Blondel, J.-L. Guillaume, R. Lambiotte and E. Lefebvre, "Fast unfolding of communities in large networks," Journal of Statistical Mechanics: Theory and Experiment volume 2008, vol. 2008, October 2008. [10] R. Lambiotte, J.-C. Delvenne and M. J. Barahona, "Random Walks, Markov Processes and the Multiscale Modular Organization of Complex Networks," IEEE Transactions on Network Science and Engineering, vol. 1, no. 2, pp. 76-90, 2014. [11] N. Dugué and A. Perez, "Directed Louvain: maximizing modularity in directed networks," Université d’Orléans, Orléans, 2015. [12] A. Ceder and N. H. M. Wilson, “Bus Network Design”, Transportation Research Part B: Methodological, vol. 20B, no. 4, pp. 331-344, 1986
[2] J. d. D. Ortúzar and L. G. Wilumsen, Modelling Transport 4th Edition, John Wiley & Sons, Ltd., 2011.
[13] Y. Peranginangin and A. Alamsyah, “Social engagement analysis in online conversation of Indonesia higher education: Case Study Telkom Univesity,” in 3rd International Conference on Information and Communication Technology (ICoICT ), Nusa Dua, Bali, 2015.
[3] Badan Pusat Statistik, "bandungkota.bps.go.id," 11 8 2017. [Online]. Available: https://bandungkota.bps.go.id/publication/2017/08/11/7cf46753e6cb99 92a7e401b6/kota-bandung-dalam-angka-2017.html. [Accessed 20 2 2018].
[14] R. Fernandez, “Can Santiago de Chile's transport policy break the vicious circle?” Traffic Engineering and Control, vol. 6, no. 49, pp 332-339, 2005.
[4] R. H. Samet, "Technological Evolution of Infrastructure System," in Long-Range Futures Research: An Application of Complexity Science, North Charleston, BookSurge Publishing, 2008, pp. 245-251.
[15] A. Alamsyah and Y. Peranginangin, “Network market analysis using large scale social network conversation of Indonesia's fast food industry,” in 3rd International Conference on Information and Communication Technology (ICoICT ), Nusa Dua, Bali, 2015.
[5] D. A. Hensher and K. J. Button, Handbook of Transport Modelling, 2nd Edition, Bingley: Emerald Group Publishing Ltd., 2008.
[16] R. Liu, S. Feng, R. Shi, and W. Guo, “Weighted graph clustering for community detection of large social networks,” in 2nd International Conference on Information Technology and Quantitative Management, Moscow, Russia, 2014.
[6] M. Newman, Networks: An Introduction, New York: Oxford Univesity Press, 2010. [7] S. M. Nourbakhsh and Y. Ouyang, "A structured flexible transit system for low demand areas," Transportation Research Part B, vol. 46, pp. 204-216, 2012.
[17] A. L. Barabási, Linked: The New Science of Networks, Cambridge, Massachussets: Perseus Publishing, 2002.
[8] F. D. Malliaros and M. Vazirgiannis, "Clustering and Community Detection in Directed Networks: A Survey," Physics Reports, vol. 533, no. 4, pp. 95-142, 2013.
113
