Content uploaded by Bernd Ludwig
Author content
All content in this area was uploaded by Bernd Ludwig
Content may be subject to copyright.
Recommendation of Personalized Routes with
Public Transport Connections
Bernd Ludwig, Bjørn Zenker, Jan Schrader
Chair for Artificial Intelligence
Friedrich-Alexander-University Erlangen-Nuremberg
Haberstraße 2, D-91058 Erlangen
ludwig@cs.fau.de, bjoern.zenker@cs.fau.de, jan.schrader@cs.fau.de
Abstract. ROSE (ROuting SErvice) is an application for mobile phones,
which suggests events and locations to the user and guides him to them
via public transportation. Many different systems partly incorporating
recommending and navigation features exist. However, no system exists
which combines event recommendation and pedestrian navigation with
(live) public transport support. In this paper we describe ROSE a mobile
application which combines these features. Our motivation is to free the
passenger from many tedious tasks, e.g. finding an interesting event and
navigating to it. ROSE determines the best possible transport link and
then accompanies the passenger throughout his entire journey. It reacts
in real time to delays in the public transport system and calculates alter-
native routes when necessary. For route planning in this context, we will
propose a h
u-optimal algorithm for incorporating non-monotone multi
dimensional user preferences in an A*-like algorithm. We also present
an assignment of theoretical foundations to real world route planning
problems.
1 Introduction
The development of mobile hardware has lead to a new variety of navigation
systems. There are situations in which this is not the case, for example, tourists
in unfamiliar cities can only use the navigation system, after they have looked
up the places available to visit, e.g. a famous castle. Also city residents often do
not know where to go or what to do. If they want to enjoy a jazz concert, they
first have to look up when and where there is such a concert and then they have
to plan the trip to it. For this, they have to use at least two different information
sources: one to search for the concert and a navigation tool. In on-the-move
situations this is not practical.
To ease the whole process of trip planning, we are developing ROSE which
combines the recommendation of events and locations with navigation. To ac-
count for the weak hardware of mobile phones we use a client server architecture,
where the ROSE server processes all data from different service providers like
for example live public transport data, event and location directories or map
services.
1.1 Combining Pedestrian Navigation with Event Recommendation
and Live Public Transport Routing
We separated the system in three main parts:
Recommendation Part To get a recommendation, the user enters a query,
like ’eat pizza’, into his mobile phone. The recommender then generates a list of
suggestions based on the user input and the user preferences. In this example it
would likely be a list of restaurants which sell pizza.
For recommendation we you use currently stemmed string matching, results
are ranked using the Okapi BM25 measure [1].
The Okapi measure helps us normalizing documents of different length, be-
cause different event provider tend to send event descriptions of different sizes.
Normally (without Okapi) the longer texts tend to score higher and shorter
descriptions are getting unjustified penalized.
Route Generation Part After the user chose one of the presented options,
the system calculates a route from the current location to the selected goal.
To consider a diverse set of user preferences in route generation, we propose a
h
u-optimal algorithm in section 2.
To ease the traveling, public transportation is also considered. The system
calculates a route from the user’s current position to the nearest public transport
option, which means of transportation to take, where to change transportation
and how to walk from the last stop to the goal location.
Navigation Part Figure 3 shows, how the route is displayed on a map on the
mobile phone. If the route includes public transport, the next possible departure
time is shown and the user is informed, whether he has to hurry to catch a
bus. The system also informs the user if he has to hurry to reach his goal. As
map-data we are using OpenStreetMap.
All three parts can be loosely coupled: the results of the recommendation
are the input (goals) of the route generation. The result of the route generation
is the input (way) of the navigation service. Such a loose coupling lacks the
flexibility needed in many special situations, especially when errors occur, or the
user behaves in an unpredicted manner. For example, if the user misses a bus,
the system has to decide what to do: wait for the next bus, take another line, or
just walk. Another case occurs, when the user wants to get a recommendation
for multiple events, a route which goes through these events respectively. Often,
the events that match well with the user preferences are located too far away
from each other. This illustrates, loose coupling of recommendation and route
generation to be disadvantageous.
This is an example of the disadvantage of loose coupling of recommendation
and route generation.
To address these problems, e.g. considering distance when recommending
multiple events, we propose a closer coupling of recommendation, route genera-
tion and navigation. Close coupling results in a theoretical problem formulation,
as it can be seen in Section 3.
1.2 Current State of the ROSE System
To address the limitations of mobile devices like limited computational power and
slow and expensive Internet access, we constructed a client server-architecture.
Expensive calculations are moved to the server and the transferred amount of
data is minimized. The ROSE system consists of the ROSE server, a J2EE
application which integrates different services from multiple service providers
and offers them as web services to the ROSE client, as you can see on the
left hand side in Figure 1. The connections to 3rd party data is shown on the
right hand side. In the middle you can see the ROSE databases which contain
preprocessed data from various providers.
Fig. 1. Overview of the ROSE server
At the moment, we are developing two prototypes of the client, running on
J2ME enabled mobile phones with GPS receiver and on Google Android.
The client sends the query over HTTP to the ROSE server. The server then
gathers, based on the user’s location, information from various providers, prepro-
cesses it, and sends the result to the client. Routes, timetables, and live public
transport data is obtained via VPN from a local public transport company.
On the client side, different localization services are integrated to cover a
large number of devices and to allow navigation in various locations. As the
start and end of a journey are often in buildings, and to support navigation
also in subways a suitable indoor localization technique is needed. Therefore,
we incorporate the Fraunhofer WiFi-localization module [2]. It is unique in that
way as it works autonomously on the mobile terminal and does not need access
to the WiFi access points. This protects user’s privacy.
2 ALGORITHMS FOR COMFORTABLE ROUTES
Criteria for evaluating the quality of a route are limited mostly by formal con-
straints dictated by the algorithm used to find optimal paths. Efficient greedy
graph search algorithms require the heuristic function to be monotone; A* even
requires the heuristics to be optimistic, i.e. to never overestimate real costs of
a path. In practice however, such constraints for heuristics are not adequate to
reason about user preferences. In a survey conducted at our computer science
institute among public transport users, the following criteria were marked as
important by the test candidates:
– No long waiting time until departure
– Short duration of the trip
– Short foot walks
– Few changes of transportation
– No long waiting time during changes
Optimistic estimates for these criteria are just the function f (x) = 0; this
amounts to omitting the criterion completely, which is an undesired consequence.
A second important finding of our study is that users do not evaluate the
utility of a route on a one-dimensional scale (where there is always an optimum
in a finite set or closed interval of utility values), but try to find a compromise
between multiple attributes that are not comparable among each other. This
finding is supported by investigations and studies in the field of psychology of
decision making reported in [3–5]: Users accept locally sub-optimal proposals if
they optimize the benefits of a proposal and minimize its risk globally (so-called
noncompensatory decisions).
For example, somebody traveling with a lot of luggage accepts using a bus
line that arrives some minutes later at the train station than the fastest one,
but is much less crowded. Obviously, in this context, the comfort of the trip is
valued higher than the duration. From an algorithmic point of view, this means
that the standard shortest path approach cannot be applied successfully in order
to satisfy the user needs as good as possible. However, searching according to a
heuristic function that forces the search procedure to visit (almost) the whole
search space is no attractive option for developing programs intended to run in
real-time.
2.1 An h
u-optimal Algorithm for Comfortable Routes
The key to an efficient solution that belongs to the same complexity class as A*
and retains it’s soundness and completeness is therefore to use a) a monotone
heuristic function h for computing correct solutions and b) to incorporate a non-
monotone, multi-attribute heuristic u for the user preferences into the search
procedure.
g
s
h
1
/u
1
h
2
/u
2
g
2
/t
2
g
1
/t
1
Fig. 2. h
u-optimal Expansion of the Search Space
In Fig. 2 you can see two paths from the start node s to a goal node g. g
i
denotes the actual costs from s to the node which is currently being expanded,
h
i
the estimated costs to g according to the heuristic function h. u
i
denotes the
actual valuation from s to the node which is currently being expanded, t
i
the
estimated valuation to g according to the multi-attribute heuristic function u
(user preferences).
If there is exactly one optimal successor state with |g
1
+ h
1
− g
2
− h
2
| > ,
the search procedure works as usual. In the case of |g
1
+ h
1
− g
2
− h
2
| ≤ , both
paths from s to g in Fig. 2 are undistinguishable in h.
In practical applications, if |g
1
+ h
1
− g
2
− h
2
| ≤ then often other criteria
become relevant for decisions in the sense of [5]. In order to explain the meaning
of , let us consider the two public transport connections from s to g in Fig. 2:
one takes 50 minutes, the other one 47 minutes. However, the first connection
requires fewer changes of transportation than the second one. If the user dislikes
changes, the first connnection is optimal according to the user preferences u
under the assumption of an -tolerance ≥ 3 minutes.
In order to select a path that is both optimal in the sense of h given and
in the sense of u, we compute the ranking p
1
= t
1
+ u
1
and p
2
= t
2
+ u
2
of
both options. The vectors p
1
and p
2
represent all user preferences in u. Each
dimension represents a single preference (cf. the list of criteria in section 2). As
the preference space to which p
1
and p
2
belong is partially ordered, in general
there is no unique minimal element. Therefore, we compute a pareto-optimal set
which includes all paths which cannot be distinguished neither by h nor by u.
The final selection of a path which is h
-optimal and u-optimal is performed by
applying a decision procedure (as described in section 2.3). We call the result
h
u-optimal.
2.2 Approximation of the h
u-optimal Algorithm
Instead of implementing the h
u-optimal algorithm directly, we simulated it’s
behavior by having the path search procedure compute the n-best list of solutions
for a given destination. Each of these n solutions is evaluated according to the
user preferences. The final set of m < n solutions is the set of m routes which
are h
u-optimal.
In out first prototype, we implemented the n-best approach in order to obtain
an evaluation platform as fast as possible. The algorithm works in two steps:
1. Compute n best results
Just computing the n best routes using always the same heuristics often leads
to proposals that only differ minimally among each other — in particular,
if just one line serves as public transport to the destination. Therefore, it is
better to compute n routes using n different (optimistic) heuristics and to
compare the n resulting best routes. This observation has also been made
by [6].
2. Rank the n best lists
In order to get a global score for each user preference and each route, we
sum up the contributions of each segment of the route to each criterion. The
sum is called the rating of the route corresponding to the criterion under
investigation. Finally, a total score is computed: The easiest approach to take
multiple criteria into account is to compute a weighted sum of all criteria by
multiplying each rating with the weight for the preference as entered by the
user in the configuration dialog for the ROSE system (see Figure 1 left).
A more elaborate approach is to compute a pareto-optimum for the rating
of the n routes. Beyond that, it is possible in our system to apply other
multi-attribute decision rules, as the Take The Best decision rule proposed
by [7].
We have implemented the described approach in JavaME for a standard
Nokia N95. Figure 1 right shows a screenshot for a calculated route in N
¨
urnberg
(Germany) town centre. On the average, the algorithm takes just a few seconds
to compute at least five solutions for the user’s request. The underlying public
transportation network consists of about one hundred bus stops. There are ca. 20
Fig. 3. left: Configuration Dialog, right: Calculated route on mobile phone (map from
OpenStreetMap under CC-BY-SA)
bus lines, each of them departing from 5 to 10 times per hour. As a consequence,
the memory consumption of the search procedure is high: the frequency of the
bus lines leads to large open and close lists. To reduce the search space, we
implemented a number of domain-dependent pruning strategies.
2.3 Multi-Attribute Decisions
A common and often used decision rule for multi-attribute decisions is the
weighted sum. As known from decision theory, weighted sums are problematic as
incommensurable data is mapped onto a unique scale. For example, the duration
of a trip and the amount of space in a bus available for a passenger’s luggage do
not have the same units. Therefore, just multiplying such values with a weight
factor and adding it to some other weighted value makes it hard for the user
to understand the system’s decision. Anyway, it is difficult for the designer of
such a system to improve the selection process by adjusting the weights, as it
is impossible to state that the user always prefers one criteria over others, e.g.
short waiting time is preferred over trip duration.
In order to better model the way users take their decisions among several
proposed routes, we implemented a module for multi-attribute decisions. Its
basis is formed by a decision-oriented programming language. Figure 2 shows a
code fragment for a decision in this language called MADL (Multi Attributive
Decision Language). The decision named ’doWhat’, using the pareto decision
rule, consists of three alternatives, namely taking bus 293, waiting and taking
bus 288 or walking. The goal is to minimize the three given attributes ’waiting
Time’, ’duration’ and ’trip duration’.
The advantages of using MADL are: all decision specific information is stored
in one designated place and easily understandable and adaptable by humans. It
ParetoDecision doWhat
{
ALT [bus293]
ALT [wait, bus288]
ALT [walk]
GOAL MIN!(waitingTime)
GOAL MIN!(changes)
GOAL MIN!(tripDuration)
}
Fig. 4. MADL decision
allows configuring the parameters that influence a decision and the strategy (de-
cision rule) used to make a decision in a given situation. Furthermore, MADL
allows combination and inheritance of decisions. It fosters the usage and evalua-
tion of different decision rules, as they are already implemented and can be used
out of the box.
2.4 Comparison
Our algorithmic idea is similar to the approaches of various recommendation and
navigation systems. COMPASS [8] and Magitti [9] employ different prediction
strategies in recommendation and rate the yielded results based on user prefer-
ences. P-TOUR [10] uses a genetic algorithm to find different near-best solutions
and presents them to the user in a k-means clustered overview, from which he
can select his preferred route. RouteCheckr [11] employs a multi-criteria Dijk-
stra based algorithm, which remains limited to the usage of the weighted sum.
As it is not using an estimation function, it does not have to cope with non-
monotone and optimal criteria, but is presumably slower as an algorithm using
such a heuristic. Hochmair [12] studies, which decision rules bicyclists utilize in
route planning and discusses different decision rules. He concludes that a com-
pensatory decision rule should be used, but he does not implement this concept
in a new algorithm.
In contrast, PECITAS [6] is a mobile personalizable navigation system, which
advices routes using means of public transportation. However, it does not include
recommendation of events or locations and routes are restricted to one starting
point and one destination point only. For user adaptation, PECITAS generates
multiple routes by using different heuristics (e.g. fastest route, or not taking any
bus) and ranks them according to user preferences (walking preferences, number
of bus changes, arrival at destination, sightseeing).
Compared to these systems, the unique features of ROSE are:
– routing to multiple destinations,
– integration of recommendation and route generation with live public trans-
port support and navigation
– usage of various compensatory decision rules.
3 ASSIGNMENT TO THEORETICAL PROBLEMS
The presentation of the ROSE system and its capabilities highlights the fact
that applications such as ROSE should offer two modes of usage:
Single destination mode: the user wants to reach a single location or com-
plete just one tasks and needs assistance in finding the appropriate location
and a transportation link to it.
Multiple destination mode: the user wants to reach more then one location
or complete several tasks. The locations or the respective tasks are partially
ordered leading to some time constraints that have to be respected by ROSE
when it recommends locations and transportation links.
In order to give an overview about the algorithmic complexity for the func-
tionality that has to be provided by ROSE, in this section we sketch the graph
theoretical problem for each (level of) functionality. The most fundamental prob-
lem finding the shortest path from one location to another is known as the short-
est path problem and investigated in depth. It solves the problem of finding a
transportation link without respecting a time table. An extension to this prob-
lem includes public transport. A theoretical definition of this problem is known
as time-dependent shortest path problem (TDSP) and according algorithms can
be found in [13] and [14].
A different extension is to find a best path between multiple, differently val-
ued locations. This is known as the orienteering problem (OP) [15], which is NP-
hard. Furtheron, considering opening hours of locations leads to the orienteering
problem with time windows (OPTW). [16] gives an overview. This problem is
extended for the tourist route recommendation context by [17] which takes mul-
tiple constrained attributes into account, e.g. a maximum price of the trip. [16]
presents an iterated local search based algorithm which yields good results (’[...]
below 2 seconds for problems with up to 100 locations [...]’). For our application
scenario we need another extension to support public transport in route finding.
As [14] shows, normal shortest path problems have little in common with public
transport best path problems. [18] introduces the time-dependent orienteering
problem (TDO) which adds public transport support to the OP.
By taking time windows into account, TDO is extended to the Multi Path
Orienteering Problem with Time Windows (MPOPTW). [19] recently proposed
an algorithm for this problem.
Table 1 shows an overview about some route planning problems, their corre-
sponding theoretical problems and adequate algorithms.
4 CONCLUSION AND OUTLOOK
We introduced ROSE, a mobile system for recommending events and planning
routes. For considering arbitrary user preferences when using A* like search,
Route Planning
Problem
Graph-Theoretical
Problem
Most Efficient
Algorithm
Single destination
Shortest Path Dijkstra, A*
with public transport Time-dependent shortest
path (TDSP)
PFS [14] , Ding [13]
Multiple destination recommendation
Orienteering Problem (OP) Fischetti, Salazar, Toth
[15]
with time windows Orienteering Problem with
time windows(OPTW)
For an overview see [16]
with time windows and
constrained attributes
Multi-constrained team
orienteering problem with
time windows
(MCTOPTW)
Garcia et al. [16]
with public transport Time-dependent
orienteering problem
(TDO)
Fomin, Lingas [18]
with public transport and
time windows
Multi Path Orienteering
Problem with Time
Windows (MPOPTW)
Garcia et al. [19]
Table 1. Assignment of route planning problems to theoretical problems
we developed a h
u-optimal algorithm. Combined with MADL it allows easy
integration of multiple criteria and different decision rules into its heuristic.
We studied mappings of different real world problems to theoretical problems
and found out, that at the moment, no theoretical problem formulation for our
scenario exists.
Our next steps are developing a suitable heuristic algorithm for solving the
MPOPTW problem in our context and to conduct user studies to find criteria
for the heuristic and more information about decision rules in the combination
of recommendation and route finding.
Additionally, the impact of close and loose coupling of the recommendation
and route finding services on the user satisfaction shall be researched. As we
presume that the algorithmic complexity of close coupling is too big and that
loose coupling is not satisfactory, we want to research, how an intermediate
coupling between these extremes could be realized.
References
1. Singhal, A.: Modern information retrieval: A brief overview. Bulletin of the IEEE
Computer Society Technical Committee on Data Engineering 24(4) (2001) 35–42
2. Meyer, S., V.T.H.S.: Wi-fi coverage and propagation for localization purposes in
permanently changing urban areas. International Journal of Information Technol-
ogy and Web Engineering, Amsterdam, The Netherlands (2008)
3. Tversky, A.: Elimination by aspects: A theory of choice. Psychological Review 79
(1972) 281–299
4. Tversky, A.: Choice by elimination. Journal of Mathematical Psychology 9 (1972)
341–367
5. Kahneman, D., Slovic, P., Tversky, A.: Judgement under uncertainty - Heuristics
and biases. Cambridge University Press, Cambridge (1981)
6. Tumas G., Ricci, F.: Personalized mobile city transport advisory system. ENTER
Conference 2009 (2009)
7. Gigerenzer, G., G.D.G.: Reasoning the fast and frugal way: Models of bounded
rationality. Psychological Review 103 (1996) 650–669
8. van Setten, M., P.S.K.J.: Context-aware recommendations in the mobile tourist
application compass. Adaptive Hypermedia and Adaptive Web-Based Systems
(1111) 235–244
9. Bellotti, V.e.a.: Activity-based serendipitous recommendations with the magitti
mobile leisure guide. CHI 2008 Proceedings (2008)
10. Maruyama, A., S.N.M.Y.Y.K.: P-tour: A personal navigation system for tourism.
Proc. of 11th World Congress on ITS (2004)
11. Voelkel, T., W.G.: Routecheckr: personalized multicriteria routing for mobility
impaired pedestrians. Proceedings of the 10th international ACM SIGACCESS
conference on Computers and accessibility (2008) 185–192
12. Hochmair, H.: Decision support for bicycle route planning in urban environments.
Proceedings of the 7th AGILE Conference on Geographic Information Science,
Crete University Press, Heraklion, Greece (2004)
13. Ding, B., X.Y.J.Q.L.: Finding time-dependent shortest paths over large graphs.
EDBT Proceedings (2008)
14. Huang, R.: A schedule-based pathfinding algorithm for transit networks using
pattern first search. Geoinformatica, Greece 11 (2007) 269–285
15. Fischetti, M., S.J.T.P.: Solving the orienteering problem through branch-and-cut.
INFORMS J. Comput. 10 (1998) 133–148
16. Garcia, A., V.P.S.W.L.M.T.A.O.: Iterated local search applied to the multi-
constrained team orienteering problem with time windows. submitted to Com-
puters & Industrial Engineering (to appear)
17. Vansteenwegen, P., S.W.G.A.: Personalised tourist guide: Multi-constraint team
orienteering problem with time windows. Proceedings of ORBEL2009, Leuven,
Belgium (2009)
18. Fomin, F. V., L.A.: Approximation algorithms for time-dependent orienteering,
information processing letters. Information Processing Letters 83 (2002) 57–62
19. Garcia, A., A.O.O.O.V.P., Linaza, M.: Public transportation algorithm for an
intelligent routing system. 16th ITS world congress, September 21-25, Stockholm,
Sweden: to appear (2009)
