Route Recommendation Algorithm for Railway Transit Travelers

based on Classification of Personal Characteristics

Yan Hong and Du Xiaoping

Software College of Beihang University, Beijing 100191, China

Keywords: Urban Railway Traffic, Route Selection, Traveler Classification.

Abstract: With the rapid development of urban rail transit network, traveler’s route decision become more difficult to

make and travelers’ route preferences vary with their characteristics. This study proposed a route recom-

mendation algorithm with the least generalized travel cost based on the classification of traveler’s personal

characteristic. The generalized travel cost model was established with the consideration of LOS variables

(e.g. in-vehicle time, transfer time, number of transfers, in-vehicle traveler density, etc) and then a traveler

classifier was constructed based on the K- nearest neighbor algorithm by machine learning how travelers’

characteristics affect their route choice intentions, thus the optimal route with the least generalized cost for

each type of travelers being generated. Finally, the model and algorithm were verified to be valid with the

data from Beijing subway network.

1 INTRODUCTION

As the rail transit network has formed in more and

more cities and the seamless transfer operation mode

is adopted, travelers will have multiple route choices

between a pair of OD (origin to destination). The

traditional route selection algorithm couldn’t meet

different route preferences of different travelers with

different characteristics. In recent years many schol-

ars have studied on the problem of traveler’s route

selection problem in urban rail transit network, such

as Zhang designed the route planning algorithm

based on the MNL (Multinomial Logit) model

(Zhang Y S, Yao Y, 2013), Zhao Nan

studied the

multi route selection problem of Shenzhen rail transit

based on the normal distribution model (ZHAO Nan,

LI Chao, 2012) and Liu

constructed a personalized

route planning algorithm for rail transit travelers

combined with travelers’ attributes based on the

MNL model (Liu Sha-sha, Yao En-jian, Zhang

Yong-sheng, 2014). However none of these studies

focused on how travelers’ attributes affect their route

choice intention. So this paper extended the method

of existing route planning algorithm by combining

with the construction of a traveler classifier based on

the K nearest neighbor algorithm, which at the same

time reconstructed the generalized travel cost model

taking into consideration the factors of pass-ups,

transfer time and in-vehicle traveler density.

2 GENERALIZED TRAVEL COST

MODEL FOR SUBWAY TRAV-

ELERS

Under the condition of seamless transfer, the route

selection problem in urban rail transit network is a

decision making problem from behavioral science. In

order to simulate the traveler’s selection behavior,

we can define a generalized travel cost for each route

(Si Bing-feng, Mao Bao-hua, Liu Zhi-li, 2007),

which take into consideration all the factors

concluded when a traveler select a route. The

Modeling process of the generalized travel cost is as

follows.

Suppose that Fare is the generalized travel cost of

a route between the OD pair, n stands for the transfer

station, N represents the transfer times and i

represents the section between two sites on the route.

Fare can be made up of two parts, the basic time T

and the extra cost E.

Fare = T + E

(1)

The basic time T includes the in-vehicle time

in veh

−

and the transfer time

trans

. Transfer time

120

Xiaoping D. and Hong Y.

Route Recommendation Algorithm for Railway Transit Travelers based on Classiﬁcation of Personal Characteristics.

DOI: 10.5220/0006020201200125

In Proceedings of the Information Science and Management Engineering III (ISME 2015), pages 120-125

ISBN: 978-989-758-163-2

consists of transfer-walking time

and platform-

waiting time

, as follows.

in veh trans

T t t

−

= +

(2)

(

)

trans n n

t e w

= +

∑

(3)

In formula (3), e

stands for the transfer-walking

time in transfer station n and w

represents the

platform-waiting time, which is related to the

average departure interval Tn and the average

number of pass-ups j

(

)

1 2

n n n

w j T

= + ⋅

(4)

The extra cost is generated by transfer and

congestion. Suppose that e

trans

and e

comf

represent

the cost from transfer and congestion, as follow.

trans conmf

E e e

= +

(5)

The transfer will generate extra cost for it takes

physical energy so that travelers have the fear to do

it. According to the result of the questionnaire of

traveler’s trip characteristics in urban rail transit,

which was part of the 2014 rail transit passenger

flow investigation project, different travelers have

different expect to the reduced travel time of

increasing a transfer. The one who pursues faster

expects less, quite proportion of them would choose

the option of “5 minutes”, meaning that they’d rather

sacrifice the comfort to save time. At the meanwhile

the one who pursues a minimum of transfers or most

comfort in vehicle would choose the longer time

such as “10 minutes”. So the transfer cost can be

expressed as follow:

trans

e N

= ⋅

(6)

The parameter

indicates the expected reduced

time when a traveler increases a transfer.

Congestion in vehicle will also generate extra

cost. According to the research, the comfort level in

vehicle can be determined by the density of standing

travelers

, which can be divided as follows:

Comfort standard: 0-3.5 people/m

. Passengers

can move freely feeling comfortable and satisfied.

Moreover there is a great chance to have a seat

during the trip.

Congestion standard: 3.5-7.5 people/m

. The

congestion generates some cost.

Excess capacity standard: 7.5 people/m

and

above. Passengers will feel obviously crowded.

Standing passengers will breakthrough into the

seating area so that seating passengers will also feel

inconvenience. Congestion now generates larger

cost.

Now suppose

represents the density of

standing travelers and the unit is people/m

. Mi

stands for the extra cost caused by congestion during

interval i. Congestion cost

comf

can be expressed

as follows.

comf i

e M

∑

(7)

0; 3.5

;3.5 6.5

; 6.5

β ρ

ω ρ





= ≤ <





≥



(8)

In the formula above

and

respectively

represents the cost in congestion standard and excess

capacity standard during interval i.

3 CLASSIFICATION OF TRAV-

ELERS’ ROUTE CHOICE IN-

TENTION

In this section, we first classified the traveler’s route

choice intention into 3 types, based on which the

parameters of generalized travel cost model were

defined. Then by using the method of machine

learning, we studied how to classify the travelers'

route choice intention according to the travelers'

attributes. Finally the steps of classify algorithm of

the travelers were given based on the K nearest

neighbor algorithm.

This paper deeply analyzed the questionnaire of

traveler’s trip characteristics in urban rail transit.

The questionnaire contents include travelers’

attributes, trip characteristics and route choice

intention. About the route choice intention 3 options

were set: “shortest time”, “transfer least” and “most

comfortable in-vehicle”, respondents were asked to

sort the weight of the 3 factors when making a route

decision. In the real situation travelers often don’t

take only a single factor as a comprehensive

consideration, so we made statistics by two priority

factors considered by travelers: 34% gave priority to

the factors of time and transfer, among those many

had a medium or a short trip distance or in purpose

of commuting; 15% gave priority to the factors of

transfer and comfort, most of them were not in

purpose of commuting or they are elder people; 12%

gave priority to the factors of time and comfort, most

Route Recommendation Algorithm for Railway Transit Travelers based on Classification of Personal Characteristics

121

Route Recommendation Algorithm for Railway Transit Travelers based on Classiﬁcation of Personal Characteristics

121

of them were in purpose of commuting and had a

longer trip distance. So the three categories were got

and for each type of travelers we defined the

parameter values of the generalized cost model

based on the questionnaire, results are in table 1.

Table 1: Generalized travel cost model parameter values of

3 categories of travellers.

Category

Priorities

A time; transfer 5 0 0

B transfer; comfort

10 0.5 1

C time; comfort 0 1 2

Two types of traveler characters affect their route

choice intention: travelers’ own attributes and trip

characteristics. Travelers’ own attributes include

age, gender, and personality and so on. Personality

factor are more random so it is excluded from the

study. Travelers’ trip characteristics include travel

distance and purpose. Then we need to construct a

classifier to input the traveler’s characteristics and

output the type of traveler’s route choice intention.

The input characteristics of this paper are: age,

gender, travel purpose and travel distance.

Considering different travelers with different

gender have different feelings about the distance and

comfort and the sample under different travel

purpose are obviously different, so we divided the

sample into four parts: male non-commuters, male

commuters, female non-commuters and female

commuters. Sample distributions are as figure 1.

From the figure above something can be seen,

such as the elderly female tend to choose more

comfortable route, while young man would pursue

faster route, far distance would make travelers

choose more comfortable route and commuters

would choose a route in a shorter time.

Considering age and distance are continuous

variables, KNN algorithm was used respectively for

the four sample sets to construct the classifier. KNN

algorithm works as follows: There is a training

sample set, and the relationship between each record

Figure 1: Travellers’ priorities in different sample sets.

ISME 2015 - Information Science and Management Engineering III

122

ISME 2015 - International Conference on Information System and Management Engineering

122

and its category label is known. Input new data with

no labels, compare the characteristics of each new

record with the data in the sample set, then the

algorithm extracts the new data’s category labels

according to the most similar data (nearest neighbor)

in the sample set. Generally speaking we only select

first k records most similar to the new record from

the sample set and k is usually an integer less than

20. Finally choose the most common category label

of the k similar records as the new record’s category

label.

In this paper, 90% of the existing data was used

as the training sample and the remaining 10% was

used as test the accuracy of the classifier. K was

valued 10. There will be some random error in the

classifier for the travelers’ route choice intention is

also influenced by the personalities and some

random factors. The final test results are shown in

Table 2, the error rate is within 20% in the four

sample sets so the classifier is considered valid.

Table 2: Test results of all kinds of samples.

No. Sample type Sample size Error rate

1 Male non-commuters 197 0.179

2 Male commuters 443 0.192

3 Female non-commuters

179 0.188

4 Female commuters 378 0.199

Thus traveler classification algorithm is as

follows:

Step 1 Input the age, gender, travel purpose and

distance of a traveler;

Step 2 Judge the training set type according to

travelers’ travel purpose and gender;

Step 3 Calculate the distance between the new

data point and the training data points in dimensions

of age and travel distance.

Step 4 Sort the distance by ascending order and

select the first 10;

Step 5 Confirm the categories of the 10 travelers’

route choice intention and return the final category

with the highest frequency.

4 PERSONALIZED ROUTE REC-

OMMENDATION ALGORITHM

Usually when travelers travel by rail transit they

don’t consider all the paths between the OD, instead

they only consider a part of them, which we call

effective path set. So the difficulty is to find the

effective path set. This paper searched the effective

path set based on the depth first traversal algorithm

and the basic idea is as follows: look for a path

connected from the OD which meets the constraints

based on the traversal algorithm; record the path if it

meets the conditions or go back to the father nodes

to traverse again if it doesn’t; repeat the trial of

selection and return until you find all the effective

paths.

In the establishment of effective path set, based

on the principle of least generalized travel cost, the

personalized route recommendation algorithm

proceeds as follows:

Step 1 Input the information of an OD and a

traveler’s age, gender, travel purpose and departure

time;

Step 2 Calculate the shortest distance and search

the effective path set between the OD.

Step 3 Input to the traveler classifier with

traveler’s attributes and trip characteristics and the

traveler’s category label will be output;

Step 4 According to the traveler’s category,

combined with the real-time traffic congestion data;

calculate the generalized travel cost of each effective

path.

Step 5 Select the one with the least cost in the

effective path set as the optimal route for the

traveler.

5 THE EXAMPLE ANALYSIS

In order to test the validity of the model and

algorithm, this paper selected the “Zhichun Road

Station to Songjiazhuang Station” as the OD pair

from the rail transit network in Beijing. Taking into

account that the degree of crowdedness varies with

the times of the day, this study chose the off peak

periods and the evening peak periods to analyze.

According to the algorithm there are 5 effective

paths:

Zhichun Road Station - No. 10 Line

(clockwise) - Songjiazhuang Station;

Zhichun Road Station - No. 10 Line

(counterclockwise) - Songjiazhuang Station;

Zhichun Road Station- No. 10 Line -

Huixinxijie Nankou Station - No. 5 Line -

Songjiazhuang Station;

Zhichun Road Station - No. 13 Line -

Xizhimen Station - No. Line 4 - Jiaomen West

Station - No. 10 Line - Songjiazhuang Station;

Route Recommendation Algorithm for Railway Transit Travelers based on Classification of Personal Characteristics

123

Route Recommendation Algorithm for Railway Transit Travelers based on Classiﬁcation of Personal Characteristics

123

Zhichun Road Station - No. 13 Line -

Xizhimen Station - No. 2 Line - Chongwenmen

Station - No. 5 Line - Songjiazhuang Station.

The network is shown in Figure 2.

Figure 2: Part of the Beijing rail transit network.

For easily express, the three types of generalized

travel cost model are numbered with letters: A,

giving priority to time and transfer factor; B, giving

priority to transfer and comfort factor; C, giving

priority to time and comfort factor. Table 3 shows

the basic data and Table 4 shows the results

calculated of the three type’s model in the two

periods and five routes.

It can be seen that in the evening peak period the

optimal route of all the three traveler categories is

route , this is because in the evening peak all

metro lines are crowded and line 5 and line 4 are

particularly serious, so no transfer and relatively

light congestion makes path the most

reasonable. In the off peak hours, the best choice for

B type travelers (giving priority to transfer and

comfort factor) is route , for in this moment

vehicles in this direction of line 10 have lower

passenger density than other lines and passengers

have a great chance to have seats, so it is a good

choice for people who pursue a comfortable travel.

In the off peak period, the optimal route of the C

type traveler (giving priority to time and comfort) is

path , on which the sites are fewest and the

passenger density is low, so it is a good choice for

people who is in a hurry and doesn’t like feeling

crowded.

The OD pair Songjiazhuang Station to Zhichun

Road Station and a group of random information of

travelers were put into the algorithm and the results

are shown in Table 5. The distance of this OD is

quite far, so the algorithm is humanized for travelers

to consider the factors of transfer and comfort. It can

be seen that when travelers are not commuters the

algorithm classify them as type B (giving priority to

transfer and comfort), path which is none of

transfer and relatively faster and less of congestion is

recommended to them in the evening peak periods

while in the off peak hours path which is none of

transfer and congestion is recommended. When

travelers in the off peak time and in purpose of

commuting, the path which is the shortest with a

high probability to have seats is recommended. The

result varies with individuals and times, which shows

the humanization and rationality of the algorithm.

Table 3: Basic data of 5 effective paths in different times.

Period

Route

No.

Distance[km]

Transfer

times

Sites on

route

In-vehicle

time[min]

Transfer

time[min]

Average density in-

vehicle[people/m

]

Off peak

hours

1 25389 0 22 50.6 0 3.74

2 28174 0 23 52.9 0 2.8

3 21562 1 20 45 1.75 4.25

4 21108 2 17 41 7.25 4.17

5 19925 2 15 38 9 3.21

Evening

peak hours

1 25389 0 22 50.6 0 4.35

2 28174 0 23 52.9 0 4.25

3 21562 1 20 45 15 5.83

4 21108 2 17 41 11.5 6.02

5 19925 2 15 38 15 3.75

ISME 2015 - Information Science and Management Engineering III

124

ISME 2015 - International Conference on Information System and Management Engineering

124

Table 4: Calculated costs of the three traveler types based on Table 3.

Period

Route

No.

trans

conmf

Fare

A B C A B C A B C

Off

peak

hours

1 50.6 0 0 0 0 11 22 50.6 61.6 72.6

2 52.9 0 0 0 0 2 4 52.9 54.9 56.9

3 46.75 5 10 0 0 10 20 51.75 66.75 66.75

4 48.25 10 20 0 0 8.5 17 58.25 76.75 65.25

5 47 10 20 0 0 4 8 57 71 55

Evening

peak

hours

1 50.6 0 0 0 0 11 22 50.6 61.6 72.6

2 52.9 0 0 0 0 11.5 23 52.9 64.4 75.9

3 60 5 10 0 0 17 34 65 87 94

4 52.5 10 20 0 0 14 28 62.5 86.5 80.5

5 53 10 20 0 0 10 20 63 83 73

Table 5: Calculated results based on randomized traveler information.

Period No. Gender Age Commuters or not Traveler category Optimal route

Off peak hours

female 26 no B 2

male 45 no B 2

male 28 yes C 5

Evening peak hours

male 23 yes C 1

female 35 no B 1

6 SUMMARY

This paper firstly defined the generalized travel cost

model considering the factors of crowded degree at

different times, transfer and pass-ups and so on.

Then Based on the research on the influence of the

travelers' characteristics on their route choice

intentions, the traveler classifier was constructed.

The classifier tested effective according to the

questionnaire data. On the basis of above, a route

recommendation algorithm for different types of

travelers was put forward. Through case analysis, the

method was proved to be reasonable.

REFERENCES

Zhang Y S, Yao Y. A route planning algorithm based on

subway traveler routes choice behavior analysis[C].

Proceedings of the Second International Conference

on Transportation Information and Safety. Wuhan,

China: American Society of Civil Engineers, 2013:

1972-1978.

ZHAO Nan, LI Chao. Research on Multi-route Selection

of Urban Rail Transit[J]. China Academy of Railway

Sciences, 2012(10): 72-76.

Liu Sha-sha, Yao En-jian, Zhang Yong-sheng. Personal-

ized Route Planning Algorithm for Urban Rail Transit

Travelers[J]. Journal of Transportation System Engi-

neering and Information Technology, 2014(5): 100-

104+132.

Si Bing-feng, Mao Bao-hua, Liu Zhi-li. Traveler Flow As-

signment Model and Algorithm for Urban Railway

Traffic Network under the Condition of Seamless

Transfer[J]. Journal of The China Railway Society,

2007(6): 12-18.

Route Recommendation Algorithm for Railway Transit Travelers based on Classification of Personal Characteristics

125

Route Recommendation Algorithm for Railway Transit Travelers based on Classiﬁcation of Personal Characteristics

125