CompRec-Trip: A composite recommendation system for travel planning.

Full text

Page 1

CompRec-Trip: a Composite Recommendation
System for Travel Planning
Min Xie†, Laks V.S. Lakshmanan†, Peter T. Wood‡
†Department of Computer Science, University of British Columbia
{minxie,laks}@cs.ubc.ca
‡Department of Computer Science and Information Systems, Birkbeck, University of London
ptw@dcs.bbk.ac.uk
Abstract—Classical recommender systems provide users with
a list of recommendations where each recommendation consists
of a single item, e.g., a book or a DVD. However, applications
such as travel planning can benefit from a system capable of
recommending packages of items, under a user-specified budget
and in the form of sets or sequences. In this context, there is a
need for a system that can recommend top-k packages for the
user to choose from. In this paper, we propose a novel system,
CompRec-Trip, which can automatically generate composite rec-
ommendations for travel planning. The system leverages rating
information from underlying recommender systems, allows flex-
ible package configuration and incorporates users’ cost budgets
on both time and money. Furthermore, the proposed CompRec-
Trip system has a rich graphical user interface which allows
users to customize the returned composite recommendations and
take into account external local information.
I. INTRODUCTION
Recommender systems (RS) are increasingly popular and
have become an essential driver of many applications including
web services [1]. Classical RS provide recommendations con-
sisting of single items, e.g., books or DVDs. However, there
are several applications that can benefit from a system capable
of recommending packages of items [2].
In travel planning, a user is interested in suggestions for
places to visit, or places of interest (POI). There may be a cost
to visiting each place (time, price, etc.) which the user may
want to constrain with a budget. Optionally, and independently
of the budget, there may be a notion of compatibility among
POIs in a package, modeled in the form of constraints: e.g.,
“the total distance covered in visiting all POIs in a package
should be ≤ 10 km.”, “no more than three museums in a
package”, “not more than two parks”, etc. The user may have
a limited budget and is interested in suggestions of compatible
packages of POIs such that the cost of each package is under
budget and its value (as judged from ratings) is as high as
possible. In these applications, there is a natural need for top-
k recommended packages for the user to choose from. Some
so-called “third generation” travel planning web sites, such as
NileGuide1and YourTour2, are starting to provide some of
these features, although in a limited form.
Motivated by such applications, we propose a novel sys-
tem called CompRec-Trip which can automatically generate
composite recommendations for travel planning, where each
1http://www.nileguide.com
2http://www.yourtour.com
recommendation comprises a set or sequence of POIs. Each
POI has both a value (rating or score) and a cost associated
with it, and the user specifies a maximum total cost (budget)
for any recommended package of POIs. Our CompRec-Trip
system consists of one or more recommender systems focusing
on different domains (e.g., hotels, museums, and events).
These component RS serve (i.e., recommend) top POIs in non-
increasing order of their value (explicit or predicted ratings).
In addition, the CompRec-Trip system also has access to in-
formation sources (which could be databases or web services)
which provide the cost associated with each POI.
In our setting, the problem of generating the top recom-
mendation (package) is NP-complete as it models either the
Knapsack problem [3] for packages as sets or the Orienteering
problem [4] for packages as sequences. Because of this and the
fact that we expect the component recommender systems to
provide ratings for large numbers of POIs and access to these
ratings can be relatively expensive3, we devise approximation
algorithms for generating top-k packages as recommendations.
Compared with previous work on considering complex or
composite recommendations4, our system has the following
distinguishing features: Package generation is based on rat-
ings from recommender systems; Allows flexible package
configuration; Captures users’ budgets on both time and
money; Allows flexible resulting package customization by
adding/removing POIs; Allows users to incorporate local in-
formation from websites such as Yahoo! Upcoming.
II. SYSTEM ARCHITECTURE
In a traditional RS, users rate items based on their personal
experience, and these ratings are used by the system to predict
ratings for items not rated by an active user. The predicted
ratings can be used to give the user a ranked recommendation
(item) list.
As shown in Figure 1, our composite recommendation
system is composed of one or more component RS and has
access to external sources that provide the cost of given items.
External sources can be local databases or web services. E.g.,
Yelp.com can be consulted for restaurants’ prices, and Google
Map can be consulted for distances between places. In terms
of computation, we abstract each RS as a system which serves
3Especially when the ratings need to be predicted.
4A detailed comparison can be found in [2].
978-1-4244-8958-9/11/$26.00 © 2011 IEEE ICDE Conference 20111352

Page 2

items in non-increasing order of their value (rating or score)
upon request. In addition, the system includes a compatibility
checker module, which checks whether a package satisfies
compatibility constraints, if any. The compatibility checker
consults necessary information sources in order to verify
compatibility.
The user interacts with the system by specifying cost
budgets, an integer k, and optionally compatibility constraints
on packages. The system finds the top-k packages of items
with the highest total value such that each package has a total
cost under budget and is compatible.
Fig. 1. System Architecture
III. PROBLEM STATEMENT
In this paper, for simplicity, we focus on the composite
recommendation problem with one component RS and with-
out compatibility constraints; extensions will be discussed in
Section IV-D.
Given a set N of POIs, a set U of users, an active user
u ∈ U, and a POI t ∈ N, we denote by vu(t) the value of
POI t for user u. We denote the value as v(t) when the active
user is understood. A RS predicts v(t) when it is not available,
by using the active user’s past behaviour and possibly that of
other similar users. For t ∈ N, we denote by tc(t) the time cost
and by mc(t) the monetary cost, of POI t. Given a set of POIs
R ⊆ N, we define v(R) = Σt∈Rv(t), tc(R) = Σt∈Rtc(t) and
mc(R) = Σt∈Rmc(t).
Definition 1: Top-k Composite Set Recommendations:
Given an instance I of a composite recommendation system
consisting of one component RS and an external information
source, a cost budget Bton time, a cost budget Bmon money
and an integer k, find the top-k packages P1,...,Pksuch that
each Pi has mc(Pi) ≤ Bm, tc(Pi) ≤ Bt, and among all
feasible packages5, P1,...,Pkhave the k highest total values,
i.e., v(P) ≤ v(Pi) for all feasible packages P ?∈ {P1,...,Pk}.
When k = 1, the top-k composite set recommendation
problem (CompRec-Set) can be viewed as a variation of the
classical 0/1 2-dimensional knapsack problem [3] with the
restriction that POIs can be accessed only in non-increasing
order of their value. Without loss of generality, we assume
all POIs have (time and money) costs smaller than the cost
budgets Btand Bm.
Note that ratings of POIs from the component RS are
retrieved using sorted access, while the cost of a given POI
is obtained via random access. Let cs and cr be the costs
associated with these accesses. Then the total access cost of
5We say a package if feasible if it satisfies all the budgetary constraints.
processing n POIs is n×(cs+cr). Notice that the number of
POIs in the system is often huge, and csand crcan be large
compared to the cost of in-memory operations, as often for
both accesses, information may need to be transmitted through
the Internet.
So, well-known algorithms for the knapsack problem which
need to access all POIs [3] may not be realistic. Thus, an
efficient algorithm for top-k CompRec-Set should minimize
the total access cost, i.e., it should minimize the number of
POIs accessed.
In our system, we also assume some background cost
information about each POI is available. The background cost
information can be a histogram H collected from the cost
database or just the minimum POI monetary (time) cost mcmin
(tcmin). This information can be materialized in our system
and be refreshed regularly by rescanning the cost database.
We denote the background cost information as BG.
In addition to the cost associated with each POI, we may
also need to consider the cost spent on traveling to correspond-
ing POIs. For each POI pair (t1,t2), t1,t2∈ N, we denote by
d(t1,t2) the shortest distance between t1and t2. And given a
set of POIs R ⊆ N, we define w(R) as the minimum distance
walk which covers all POIs in R, and let tcw(R) and mcw(R)
be the corresponding time and monetary cost for taking w(R)
by assuming an average speed/cost per unit of distance.
Definition 2: Top-k Composite Sequence Recommenda-
tions: Given an instance I of a composite recommendation
system consisting of one component RS and an external
information source, a cost budget Bton time, a cost budget
Bm on money and an integer k, find the top-k packages
P1,...,Pk such that each Pihas mc(Pi) + mcw(Pi) ≤ Bm,
tc(Pi) + tcw(Pi) ≤ Bt, and among all feasible packages,
P1,...,Pk have the k highest total values, i.e., v(P) ≤
v(Pi),1 ≤ i ≤ k for all feasible packages P ?∈ {P1,...,Pk}.
When k = 1, the top-k composite sequence recommenda-
tion problem (CompRec-Seq) can be viewed as a variation of
the orienteering problem [4] with the restriction that POIs can
be accessed only in non-increasing order of their value. Similar
to the composite set recommendation problem, we assume we
have some simple background cost information such as the
minimum shortest distance dminbetween POIs.
IV. IMPLEMENTATION
A. Composite Set Recommendation
For the composite set recommendation problem, we can
adopt the algorithm template studied in our previous work
[2], which is shown in Algorithm 1.
In this algorithm template, one POI is retrieved from an
underlying RS at each iteration (lines 3–4). After accessing
this new POI, we can use a pseudo-polynomial time algorithm
[3] to find an optimal solution Roover the accessed POI-
set S (line 5). We can also derive a tight upper bound6
V∗of the true optimal solution from the calculation of Ro
6An upper bound is tight iff there exists a possible unseen POI configuration
for which the upper bound is achievable by using a subset of accessed POIs
and valid unseen POIs. Here, an unseen POI t is valid iff v(t) ≤ vmin,
tc(t) ≥ tcminand mc(t) ≥ mcmin.
1353

Page 3

using MaxValBound [2] (line 6). The algorithm terminates if
v(Ro) ≥
(lines 7–8)7.
Algorithm 1: CR-Set-Top1(N, Bt, Bm, BG)
1 S ← An empty buffer
2 while TRUE do
3
t ← N.getNext()
4
S.Insert(t)
5
Ro← Optimal2DKP(S, Bt, Bm)
6
V∗= MaxValBound(S,BG)
7
if v(Ro) ≥
8
return Ro
1
α× V∗; if not, it continues to access the next POI
1
α× V∗
Similar to [2], it can be shown that the proposed CR-Set-
Top1 is both correct and instance optimal [5]. However, CR-
Set-Top1 relies on an exact algorithm for the knapsack prob-
lem which may lead to high computational cost. To remedy
this, we can develop more efficient heuristic algorithms but
may have to give up instance optimality. The idea is that
instead of using an exact algorithm to get the best package
for the currently accessed set of POIs S, we can use a simple
greedy heuristic to form a high quality package RGfrom S
and then test whether RGis globally a high quality package.
We have shown in previous work [2] that algorithms using
the greedy heuristic often have performance very close to the
instance optimal algorithm.
B. Composite Sequence Recommendation
Composite sequence recommendation is closely related to
the orienteering problem [4], which seeks a maximum value
walk on a graph subject to a budget constraint on the cost. To
simplify the presentation, we will ignore discussing cost on
each POI (i.e., the so-called node cost) as this can be handled
by a reduction to the original orienteering problem [6].
Similar to the composite set recommendation problem, to
minimize the number of POIs retrieved while having the result
quality guaranteed, we need to adapt the algorithm template as
described by Algorithm 1 and iteratively calculate the optimal
solution for the accessed POI-set S along with a tight upper
bound on the possible true optimal solution to the underlying
composite sequence recommendation problem instance.
Given an (exponential-time) exact orienteering solver, we
can calculate the optimal solution for the subgraph G of
the original POI graph, induced by the accessed POI-set S.
However, it is more challenging to get a tight upper bound
on the value of the true optimal solution for the composite
sequence recommendation problem.
Let dmin be the background cost information about the
minimum distance between POIs, vmin = mint∈Sv(t) and
t∗be an “imaginary” unseen POI which has value vmin. A
tight upper bound on the possible true optimal solution can
be computed by the procedure MaxOriValBound shown in
Algorithm 2.
It can be proven that, using procedure MaxOriValBound, we
can give a correct instance-optimal α-approximation algorithm
7Decreasing the approximation factor α can lead to higher quality packages;
however, it will also result in higher computational cost. In our system, we
set α = 2.
CR-Seq-Top1 for the composite sequence recommendation
problem. Similar to composite set recommendation, to get
better practical performance, we can utilize approximation
algorithms for the orienteering problem such as [4] instead
of exact algorithms. However, the resulting algorithm will not
be instance optimal.
Algorithm 2: MaxOriValBound(G,Bt,Bm,BG)
1 foreach Edge (v1,v2) ∈ G do
2
n∗= ?d(v1,v2)
3
Add to G a path P between v1and v2which is composed of n∗
“imaginary” POIs, each a copy of t∗, and d(P) = d(v1,v2)
4 S = OptimalOrienteer(G,Bt,Bm)
5 Augment S with “imaginary” POIs t∗if possible
6 return v(S)
dmin
?
C. Top-k Composite Recommendation
In addition to the best composite recommendation, it is
often useful to provide the user with the top-k composite
recommendations, where k is a small constant. Similar to the
top-1 case, due to the hardness of the underlying problem, we
seek an efficient approximation algorithm which can give us
high quality recommendations.
Given an instance I of the top-k composite recommendation
problem, assume RIis the set of all feasible composite recom-
mendations, i.e., RI= {R |R ⊆ N ∧mc(R) ≤ Bm,tc(R) ≤
Bt} for composite set recommendation and RI= {R | R ⊆
N ∧ mc(R) + mcw(R) ≤ Bm,tc(R) + tcw(R) ≤ Bt}
for composite sequence recommendation. Following Fagin et
al. [5], we define an α-approximation of the top-k composite
recommendations to be any set Rkof min(k, |RI|) composite
recommendations, such that, for all R ∈ Rkand R?∈ RI\Rk,
v(R) ≥1
Lawler’s procedure [7] is a general technique for enumerat-
ing optimal top-k answers to an optimization problem, which
relies on an efficient algorithm to find the optimal solution
to the problem. We have proven in our previous work [2]
that for the top-k composite recommendation problem, if the
algorithm for the optimal solution in Lawler’s procedure is
replaced by an α-approximation algorithm, instead of optimal
top-k answers, we get an α-approximation to the optimal set
of top-k composite recommendations. So in our system we
leverage Lawler’s procedure and CR-Set-Top1/CR-Seq-Top1
to produce top-k approximate composite recommendations.
D. Discussion
To capture further constraints such as “not more than two
parks” in our algorithms, we can define a Boolean compatibil-
ity function C over the packages under consideration. Given a
package P, C(P) = true iff all constraints on POIs in P are
satisfied. We can add a call to C in the previously proposed
algorithms after each candidate package has been found. If the
package fails the compatibility check, we just discard it and
search for the next candidate package.
It is worth noting that the Boolean compatibility function
defined here allows for great generality. However, depending
on the application needs, for scenarios where only one specific
type of constraint is considered, e.g., having one POI from
α× v(R?).
1354

Page 4

each of three predefined categories, more efficient algorithms
like Rank Join [8] can be leveraged.
Furthermore, although in the previous algorithms we assume
there is only one component RS, it is straightforward to
combine recommendation lists from several component RS
into a single list, based on ratings.
E. Datasets
For the implementation of our system, ratings, monetary
cost and location information are crawled from Yelp.com. For
time cost information, we assume that the time spent on each
POI should be proportional to its area/size (this information
is crawled from Wikipedia), and multiplied by a factor which
depends on the type of the POI. For example, parks are often
larger in area than museums; however, visitors tend to spend
more time in museums per unit area, so we may apply a
larger factor to museums to take this into consideration. For
POIs such as restaurants, we assume a default time cost. An
alternative way to get time cost information is to analyze
online user-generated content as described by [6].
We can obtain all the distance information using the ge-
ographic coordinates of POIs crawled from Yelp.com. We
can also obtain better quality distance information from the
Google map service. To estimate the time/money to be spent
on the trip, depending on which transportation method the user
chooses, we assume an average speed/cost for the travel.
V. GRAPHICAL USER INTERFACE
Compared with existing trip recommendation systems, one
of the highlights of our CompRec-Trip system is that users
have great power in interacting with the recommended pack-
ages and in further tuning of the results. An illustration of the
system interface is shown in Figure 2. As can be seen, there
are three major components of our system: 1. Destination and
customization interface; 2. Recommendation list; 3. Package
contents view.
1. Destination and Customization Interface: Users of
CompRec-Trip can get composite set or sequence recommen-
dations through the top tab bar by specifying the travel destina-
tion, cost budgets, number k of top packages to be generated
and also various other features including: preferences about
POIs such as number of different types of POI to be included,
and the method of transport. Furthermore, from the tab bar,
users can choose to include the latest updates from external
websites such as Yahoo! Upcoming.
2. Recommendation List: Once a user has (fully or partially)
specified the travel targets and preferences about the trip,
the CompRec-Trip system automatically generates the top-k
composite recommendations. The recommended packages will
be listed on the right-hand panel, and the user can click on each
to get a live view of the package on the left hand map interface.
If the user is not satisfied with the recommended packages,
he/she can click the “Add POI (Remove POI)” button on the
current active package. A local search interface will be shown
and the user can use this to filter POIs he/she is (is not)
interested in. Any modification to the package takes effect
immediately and the change is reflected on either the current
Fig. 2.Graphical User Interface
active recommendation or all recommendations, depending on
whether the user wants to make this fix global.
3. Package Contents View: A large part of the user interface
is occupied by the visualization of the current active recom-
mended package. A map control is shown according to the
selected package, the contents of the package are visualized on
the map and nearby interesting POIs or updates from external
information sources can be shown as an option. The user can
click on each of the POIs contained in the package to get a
brief overview of this POI and obtain links for getting further
information. Also the user can interact with this package by
removing some POIs from the package or clicking on the map
to search for some other local interesting POIs nearby, which
she can add to the current package.
VI. DEMONSTRATION
Our demo will show how top-k composite set/sequence
recommendations will be constructed by specifying various
preferences such as cost budgets, transportation method and
other constraints. We will visualize each package on a map
interface and show how users can interact with the resulting
packages by adding and removing POIs. Furthermore, we will
also show how external local information can be seamlessly
incorporated into the recommendation results and be used to
help users explore local interesting activities. The importance
of efficient top-k composite recommendation algorithms along
with incremental computing will be highlighted.
REFERENCES
[1] G. Adomavicius and A. Tuzhilin, “Toward the next generation of recom-
mender systems: A survey of the state-of-the-art and possible extensions,”
IEEE TKDE, vol. 17, no. 6, pp. 734–749, 2005.
[2] M. Xie, L. V. S. Lakshmanan, and P. T. Wood, “Breaking out of the box
of recommendations: From items to packages,” in ACM RecSys, 2010,
pp. 151–158.
[3] H. Kellerer, U. Pferschy, and D. Pisinger, Knapsack Problems. Springer,
2004.
[4] C. Chekuri and M. P´ al, “A recursive greedy algorithm for walks in
directed graphs,” in FOCS, 2005, pp. 245–253.
[5] R. Fagin, A. Lotem, and M. Naor, “Optimal aggregation algorithms for
middleware,” J. Comput. Syst. Sci., vol. 66, no. 4, pp. 614–656, 2003.
[6] M. De Choudhury, M. Feldman, S. Amer-Yahia, N. Golbandi, R. Lempel,
and C. Yu, “Automatic construction of travel itineraries using social
breadcrumbs,” in ACM Hypertext, 2010, pp. 35–44.
[7] E. L. Lawler, “A procedure for computing the k best solutions to discrete
optimization problems and its application to the shortest path problem,”
Man. Sci., vol. 18, no. 7, pp. 401–405, 1972.
[8] J. Finger and N. Polyzotis, “Robust and efficient algorithms for rank join
evaluation,” in ACM SIGMOD, 2009, pp. 415–428.
1355