Vol-3194/paper58
Jump to navigation
Jump to search
Paper
| Paper | |
|---|---|
 edit
| |
| description | |
| id | Vol-3194/paper58 |
| wikidataid | →Q117345115 |
| title | Personalized Context-Aware Recommender System for Travelers |
| pdfUrl | https://ceur-ws.org/Vol-3194/paper58.pdf |
| dblpUrl | https://dblp.org/rec/conf/sebd/ShekariSGRST22 |
| volume | Vol-3194→Vol-3194 |
| session | → |
Personalized Context-Aware Recommender System for Travelers
Personalized Context-Aware Recommender System
for Travelers1
(Discussion Paper)
Mahsa Shekari, Alireza Javadian Sabet, Chaofeng Guan, Matteo Rossi,
Fabio A. Schreiber and Letizia Tanca
Politecnico di Milano - Dipartimento di Elettronica, Informazione e Bioingegneria
Abstract
Nowadays, traveling has become more convenient thanks to many recent technological advancements.
However, the main problem is that, with non-customized offers, the risk for the travelers is to waste
their time looking for the most appropriate one. Consequently, travelers need a system capable of
understanding their contextual preferences for ranking travel offers accordingly. In this work, we
propose The Hybrid Offer Ranker (THOR) as a possible solution: on the one hand, we employ various
classification algorithms to learn the individuals’ contextual preferences; on the other hand, to help new
users the system has no information on (cold users), we employ unsupervised algorithms to identify
clusters of users with similar preferences and build group preference models accordingly.
Keywords
Context-Awareness, Data Mining, Recommender System, Mobility, User Modeling, Personalization
1. Introduction
Recent technological advancements have made traveling more convenient and more personalized
(think, e.g., of sharing mobility, or demand-responsive schemes). However, non-customized
travel offers can cause travelers to waste a lot of time in finding the most suitable ones [1].
Recommendation Systems (RS) can help in this regard, as they aim to understand the user’s
“preference” towards an item [2] and guide them in the offer selection process. Travel preferences
may be expressed by the context in which the traveler interacts with the system [3, 4], where
context is any information that can be used to characterize an entity’s situation (e.g., person,
place, physical environment) [5]. A system is context-aware if it uses context to provide the
user with relevant information and services, where relevance depends on the user’s task.
Unlike many works which focused on recommending only touristic destinations [6] to
travelers, in this work we rank a list of complete travel offers for the travelers according to
1
This work was supported by Shift2Rail and the EU Horizon 2020 research and innovation programme under
grant agreement No: 881825 (RIDE2RAIL)
SEBD 2022: The 30th Italian Symposium on Advanced Database Systems, June 19-22, 2022, Tirrenia (PI), Italy
$ mahsa.shekari@polimi.it (M. Shekari); alireza.javadian@mail.polimi.it (A. Javadian Sabet);
chaofeng.guan@mail.polimi.it (C. Guan); matteo.rossi@polimi.it (M. Rossi); fabio.schreiber@polimi.it
(F. A. Schreiber); letizia.tanca@polimi.it (L. Tanca)
� 0000-0002-5191-2961 (M. Shekari); 0000-0001-9459-2411 (A. Javadian Sabet); 0000-0002-9193-9560 (M. Rossi);
0000-0002-5237-0455 (F. A. Schreiber); 0000-0003-2607-3171 (L. Tanca)
© 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISSN 1613-0073
CEUR Workshop Proceedings (CEUR-WS.org)
�their context. This work has been carried out within the Ride2Rail (R2R)1 project, which is
part of the Innovation Programme 4 (IP4) of the Shift2Rail (S2R) initiative2 , which aims at
advancing the European transportation domain. The so-called Travel Companion (TC), which is
one of the contributions of the S2R IP4 ecosystem, is an application that can operate on various
devices to assist travelers before, during, and after each journey. To connect rural areas with
transportation hubs, the S2R ecosystem, through the TC, allows drivers to share their rides.
In this work, we continue the implementation of the recommender system introduced by
Javadian et al. [4]. The system employs the Context Dimension Tree (CDT) methodology [7, 8]
to design a contextual preference model presented in [3]. More precisely, this work aims at
answering the following questions:
1: How can we create a personalized preference model for ranking travel offers according to
the contextual preferences of the traveler using their historical records?
2: How can we create an initial preference model for new travelers using other travelers’
data with similar characteristics?
To implement a personalized Context-Aware RS for Travelers, we propose The Hybrid Offer
Ranker (THOR). THOR models the problem as a binary classification problem which predicts if
the traveler will buy any of the available travel offers or not. Using each traveler data, a unique
contextual preference model is built for each traveler. A ranking mechanism is designed in case of
multiple travel offers, which assigns the probabilities to rank the offers according to the user’s
personal preferences. In addition, the proposed system utilizes clustering algorithms to build a
preference model for each group by recognizing users with similar desires. Finally, by generating
an extensive dataset considering some precise rules, we test the system’s performance (in terms
of both computation time and accuracy). The developed RS showed promising results in terms
of the accuracy of its predictions.
The rest of the work is organised as follows. Sec. 2 presents the related work. Sec. 3 details
the proposed system and validation procedure. Finally, Sec. 4 concludes and outlines some
future work.
2. Background and Related Work
RSs typically fall into one of the following categories [9]: (i) Content-based, which provide
recommendations based on the user’s past purchases; (ii) Collaborative, which recommend
based on other users with similar preferences; and (iii) Hybrid, which combine (i) and (ii).
The framework developed by Sebasti et al. [6] collects users’ information and general pref-
erences by asking them to enter their data and introduce their specific preferences for the
current visit; then, it generates a list of activities that are likely of interest to the user. A
hybrid RS first classifies users into different classes; then, according to the user’s records, a
content-based approach recommends places, and a filter selects the fitting places considering
the current request. Javadian et al. [10] propose a data-mining-based RS to rank the offers
by considering a score according to the characteristics of the user’s mobility request. The
proposed method by Fang et al. [11] automatically generates temporal feature vectors from a
1
https://ride2rail.eu
2
http://www.shift2rail.org/
�collection of documents based on Wikipedia and Twitter. Additionally, Coelho et al. [12] employ
travel-related tweets to personalize recommendations regarding POIs for the user. In this work,
historical buildings are considered for identifying POIs and classification models. Consonni
et al. [13] collect travel-centered mobility data and characterize the time spent traveling in
terms of the activities performed, i.e., fitness, enjoyment, and productivity; understanding the
travelers and their choices according to the nature of their travel time can help the system
learn the preferences more accurately. In [14] the sorting and selection actions are provided as
perspectives related to trip search options. This allows the system to rapidly learn the users’
preferences while they select different offers.
A content-based RS needs to access a sufficient number of users’ records in order to select
the user’s preferences. As a result, a new user might not receive accurate recommendations
if they have very few records. The “cold start problem” [15] occurs when there is not enough
historical data. Some techniques mentioned in [16] employ strategies based on each item’s
popularity and entropy on the user profile to determine the best recommendation items for
new users. Another approach is recommending to a new user the top popular offers. This
can increase the user’s purchase likelihood, but also decreases the level of personalization.
Finally, collaborative methods can be helpful from the personalization point of view, but the
recommendations’ precision might remain quite low.
In this work, we combine collaborative and content-based methods to develop a novel hybrid
approach. We consider a content-based method to find appropriate travel offers for new users.
Therefore, we use a collaborative approach to build recommender models for various groups of
users with similar characteristics and use them to provide recommendations to new users.
3. The Hybrid Offer Ranker (THOR)
We call the proposed RS The Hybrid Offer Ranker (THOR). In this section we provide the reader
with some details of THOR’s implementation and functions.
The main function offered by THOR allows for the ranking of travel offers presented to TC
users according to their contextual preferences. Figure 1 provides an overview of the THOR
system. The main procedure is as follows. As soon as a TC submits a desired source and
destination, as well as potential contextual preferences (mobility request), the Travel Service
Provider (TSP) returns a list of non-ranked travel offers satisfying the user’s request. In the
next step, for each of the travel offers, the Offer Categorizer module (OC) computes category
scores (e.g., environmentally friendly and comfortable). The resulting enriched travel offer will
be used by THOR for learning and ranking. Then, the enriched offers are combined with the
user’s most recent profile information which will be used as the Ranker’s input data.
In order to provide Context-Aware recommendations to the users, THOR employs the
aspects defining the users’ choice criteria. To identify these criteria, THOR employs the CDT
methodology which models the context as a rooted tree (See Fig. 2). The idea behind this
modeling approach is that it builds the model with the most general context as the top-level
nodes and gradually breaks the context into atomic characteristics while traversing the tree. In
our case, the main context nodes are as follows.
• User Profile contains the static information of the user. For example, the personal data,
�Figure 1: High-level representation of THOR.
profile type, disability information, etc.
• Context contains the momentary information related to each of the mobility requests.
For example, the time (e.g., day vs. night), if the user was accompanying anyone, etc.
• Search Options represent specific explicit requests by the user while submitting a mo-
bility request. For example, if they prefer specific means of transportation (e.g., airplane).
• Available Travel Offers encompasses the list of travel offers characterized by their offer
category scores as well as their leg-level information. This node helps us to understand
for a given mobility request, which travel offers were presented to the user and what was
the user choice among them.
Combining the mentioned context information enables THOR to better learn the contextual
preferences of each user separately. To exemplify, please consider a female user who mostly
travels for leisure with her partner in economy class flights. After giving birth to her child,
she changes her preference to train because of the comfort for her child. THOR learns such
contextual preferences and when she sends a mobility request accompanying only her partner,
probably the economy flights will be ranked higher, while if she accompanies her child, offers
containing train could be ranked higher. It should be noted that, in this example, for brevity
we simplified the contextual preferences. In reality, the user choice is affected by a non-linear
combination of the mentioned features.
Learner module is in charge of regularly updating each user’s personal preference model; it
utilizes various classification algorithms (KNN [17], SVC [18], DT [19], RF [20], and LR [21]) to
build classifiers which take as input the historical data of a single user and outputs the user’s
� Traveler Context Dimension Tree (TCDT)
BehavioralStatus (1:2) Health (1:1)
PersonalData (1:1)
Inactive Surfing Traveling HealthRelated (1:1)
Profile Membership
Issue
Constraint
WalkingSpead
Request (1:1) Interface (1:1) Gesture (0:1) (0:n) (0:n) (1:1)
MaritalStatus
DrivingExperience
Education
Citizenship
Age
Chattiness
TopicOfInterest
SpeakingLanguage
BirthGender
NGO
LoyaltyCard
PaymentCard
(0:n) (1:n) (1:1) (1:1) (1:1) (1:1) (1:1) (1:1) (1:n) (0:n) (0:n) (0:n)
Topic
InterfaceType
HIVal
WalkingSpeadType
HealthConstraint
Language
Igonred
EducationLevel
Interested
NGOType
TOIVal
AgeCategory
Level
Gender
DELevel
CitizenshipType
(0:n) (0:n)
LoyaltyCardType
OfferCategory
OfferCategory
Mobility
ConnectionAvailability
AllowedInterchanges
Refundability
Length
LiveStatusNotification
ProfileMode
Incentive
TripInfo
Recurring
Segment
Flexibility (0:1) Accompanying (0:n) (2:n) (1:1) (1:1) (1:1) (1:1) (1:1) (1:1) (1:1) (0:n) (0:n)
Date Loc
Item
Pet
AcPerson Person
(0:n) (0:n) (0:n)
Flexibility Flexibility
PS
IncenVal
AIntVal
LengthType
RefVal
LiSNotVal
RecurringType
SegmentType
ProfileModeType
TripInfoValue
AcPet
AcItem
Destination
Departure
Arrival
(0:n) (0:n)
Source
(0:n)
(1:n) RequestedServices
AcPersonVal
DistFlexibilityType
DateRange
DistFlexibilityType
AcItemVal
DateRange
AcPetVal
ReqTravel ReqMeal
RMealType
RMode
RVehicleCondition
RMealExclusive
ROfferCategory
IsMealMandatory
(1:n) (0:n) (0:n) (0:n) (1:1)
RCarrierr
(0:n) (n:n)
RTSProduct
(n:n)
RMealEVal
RVecConVal
Bool
RTPVal
RMealTVal
RModeVal
ROfCVal
RCarVal
ReqModeService
RCarService ROtherMode
ModeService Service
Seat
RPassenger (1:1) RCar (1:1) RDriver (1:1)
Model
RMaritalStatus
RLanguage
Safety
RNGO
RGender
RChattiness
JustVerified
REducationLevel
RTopicOfInterest
RDrivingExperience
Age
Type
Fuel
RAgeCategory
ModeSeat
Number
Gender
Seat
Service
(1:1) (0:n) (0:n) (0:1) (1:n) (0:n) (1:1) (0:n) (0:n) (1:1) (1:1) (1:1) (1:1) (n:n) (n:n) (n:n) (1:1) (n:n) (n:n)
OMSeatVal
OMSerVal
RDEVal
RMSVal
RTOIVal
RCarSafVal
RACVal
RCarSeatVal
RCarModVal
RChVal
RCarServVal
RELVal
RLVal
GenderType
RCarTypeVal
RGVal
RNGOVal
AgeCategory
Bool
RCarFTVal
Int
Figure 2: Proposed Traveler Context Dimension Tree. Black circles, white circles, and white squares
represent, respectively, dimensions, concepts and parameters.
contextual preference model. Concerning the newly registered users for whom the system
has no historical data, the Learner module invokes its clustering block. A clustering block is
designed using 𝐾-means [22] and Density-based spatial clustering of applications with noise
(DBSCAN) [23] algorithms. The idea is that, at first, using the user profiles of all the users,
THOR identifies clusters of users with similar profile information. In the next step, for each
cluster, it collects the historical data of all the users belonging to that group. Then, using
the classification block mentioned earlier, it learns a group-level contextual preference model.
Having a clustering model and group-level contextual preference models for groups in hand, as
soon as a new user registers to the system, the Learner module identifies the group which the
�user belongs to. Finally, the group-level contextual preference model will be used as the user’s
personal preference model. The model will be in use until the system collects enough historical
data from the user for building a pure personal contextual preference model.
The Ranker module is the main interface for computing the rankings of enriched travel
offers. At first the Ranker retrieves the corresponding recommender model of the user as well
as the enriched travel offers. In the next step, the Ranker predicts if the user will buy or not any
of the offers and saves the results. Finally, the system computes the ranking of the list of travel
offers through the confidence score of the predictions and presents them to the user.
Since the TC is going to be used by the users in the near future, we do not have access to
any real data. As a result, to test the performance of the system, both from the computation
time and accuracy points of view, we generated a dataset according to pre-defined distributions
and rules. Having a rule-based generated dataset enabled us to compare the performance of
the various classifiers and the system as a whole. On a fully controlled setting in which user’s
historical data were carefully inserted using human supervision, THOR reaches 95% accuracy.
4. Conclusion and Future Work
In this work, we tackled the issue of personalization of the user experience in the transportation
domain by designing and implementing THOR, a system that: (i) is able to learn individuals
contextual preferences using their historical records (RQ1); and (ii) employs the wisdom of
crowds to tackle the so-called cold start problem (RQ2). THOR ranks the travel offers for each
individual using the learned models.
In future works, we plan to use appropriate feature selection methods [24] to reduce the
complexity of the models. We also plan to build the social media (SM) core proposed in [4]
as a tool to characterize urban mobility patterns [25]. Having an appropriate SM core can
help decision-makers understand the user preferences during events [26, 27] that bring many
travelers to specific European cities.
References
[1] X. Chen, Q. Liu, X. Qiao, Approaching another tourism recommender, in: 2020 IEEE
20th International Conference on Software Quality, Reliability and Security Companion
(QRS-C), 2020, pp. 556–562. doi:10.1109/QRS-C51114.2020.00097.
[2] F. Ricci, L. Rokach, B. Shapira, Introduction to Recommender Systems Handbook, Springer
US, Boston, MA, 2011, pp. 1–35. doi:10.1007/978-0-387-85820-3\_1.
[3] A. J. Sabet, M. Rossi, F. A. Schreiber, L. Tanca, Context awareness in the travel companion
of the shift2rail initiative, in: Proceedings of the 28th Italian Symposium on Advanced
Database Systems, volume 2646 of CEUR Workshop Proceedings, CEUR-WS.org, 2020, pp.
202–209. URL: http://ceur-ws.org/Vol-2646/15-paper.pdf.
[4] A. Javadian Sabet, M. Rossi, F. A. Schreiber, L. Tanca, Towards learning travelers’
preferences in a context-aware fashion, in: Ambient Intelligence – Software and Ap-
plications, Springer International Publishing, Cham, 2021, pp. 203–212. doi:10.1007/
978-3-030-58356-9\_20.
� [5] A. Dey, Understanding and using context, Personal and Ubiquitous Computing 5 (2001)
4–7. doi:10.1007/s007790170019.
[6] L. Sebastia, I. Garcia, E. Onaindia, C. Guzman, e-tourism: A tourist recommendation and
planning application, in: 2008 20th IEEE International Conference on Tools with Artificial
Intelligence, volume 2, 2008, pp. 89–96. doi:10.1109/ICTAI.2008.18.
[7] C. Bolchini, E. Quintarelli, L. Tanca, Carve: Context-aware automatic view definition over
relational databases, Inf. Syst. 38 (2013) 45–67.
[8] P. Agnello, S. M. Ansaldi, E. Lenzi, A. Mongelluzzo, D. Piantella, M. Roveri, F. A. Schreiber,
A. Scutti, M. Shekari, L. Tanca, Reckon: a real-world, context-aware knowledge-based
lab, in: 29th Italian Symposium on Advanced Database Systems, SEBD 2021, volume 2994,
CEUR-WS, 2021, pp. 466–473.
[9] Y. Cai, H. F. Leung, Q. Li, H. Min, J. Tang, J. Li, Typicality-based collaborative filtering
recommendation, IEEE Transactions on Knowledge & Data Engineering 26 (2014) 766–779.
[10] A. J. Sabet, S. Gopalakrishnan, M. Rossi, F. A. Schreiber, L. Tanca, Preference mining in
the travel domain, in: 2021 IEEE International Conference on Artificial Intelligence and
Computer Applications (ICAICA), 2021, pp. 358–365. doi:10.1109/ICAICA52286.2021.
9498231.
[11] G.-S. Fang, S. Kamei, S. Fujita, Automatic generation of temporal feature vectors with
application to tourism recommender systems, in: 2016 Fourth International Symposium
on Computing and Networking, 2016, pp. 676–680. doi:10.1109/CANDAR.2016.0121.
[12] J. Coelho, P. Nitu, P. Madiraju, A personalized travel recommendation system using social
media analysis, in: 2018 IEEE International Congress on Big Data (BigData Congress),
2018, pp. 260–263. doi:10.1109/BigDataCongress.2018.00046.
[13] C. Consonni, S. Basile, M. Manca, L. Boratto, A. Freitas, T. Kovacikova, G. Pourhashem,
Y. Cornet, What’s your value of travel time? collecting traveler-centered mobility data via
crowdsourcing, 2021. arXiv:2104.05809.
[14] L. Boratto, M. Manca, G. Lugano, M. Gogola, Characterizing user behavior in journey
planning, Computing 102 (2020). doi:10.1007/s00607-019-00775-8.
[15] A. I. Schein, A. Popescul, L. H. Ungar, D. M. Pennock, Methods and metrics for cold-start
recommendations, in: Proc. of the 25th ann. int. ACM SIGIR conf. on Research and dev. in
IR, 2002, pp. 253–260.
[16] A. Sang, S. K. Vishwakarma, A ranking based recommender system for cold start data
sparsity problem, in: 10th IC3, 2017, pp. 1–3. doi:10.1109/IC3.2017.8284347.
[17] H. b. Jaafar, N. b. Mukahar, D. A. Binti Ramli, A methodology of nearest neighbor: Design
and comparison of biometric image database, in: 2016 IEEE Student Conference on Research
and Development (SCOReD), 2016, pp. 1–6. doi:10.1109/SCORED.2016.7810073.
[18] L.-S. Lan, M-svc (mixed-norm svc) - a novel form of support vector classifier, in: IEEE
ISCAS, 2006, pp. 4 pp.–3264. doi:10.1109/ISCAS.2006.1693321.
[19] Y. Yu, F. Zhong-liang, Z. Xiang-hui, C. Wen-fang, Combining classifier based on decision
tree, in: WASE ICIE, volume 2, 2009, pp. 37–40. doi:10.1109/ICIE.2009.12.
[20] Z. Bingzhen, Q. Xiaoming, Y. Hemeng, Z. Zhubo, A random forest classification model
for transmission line image processing, in: 15th ICCSE, 2020, pp. 613–617. doi:10.1109/
ICCSE49874.2020.9201900.
[21] X. Zou, Y. Hu, Z. Tian, K. Shen, Logistic regression model optimization and case analysis,
� in: IEEE 7th ICCSNT, 2019, pp. 135–139. doi:10.1109/ICCSNT47585.2019.8962457.
[22] S. Lu, H. Yu, X. Wang, Q. Zhang, F. Li, Z. Liu, F. Ning, Clustering method of raw meal
composition based on pca and kmeans, in: 2018 37th Chinese Control Conference (CCC),
2018, pp. 9007–9010. doi:10.23919/ChiCC.2018.8482823.
[23] A. Smiti, Z. Elouedi, Dbscan-gm: An improved clustering method based on gaussian
means and dbscan techniques, in: 2012 IEEE 16th International Conference on Intelligent
Engineering Systems (INES), 2012, pp. 573–578. doi:10.1109/INES.2012.6249802.
[24] A. Brankovic, M. Hosseini, L. Piroddi, A distributed feature selection algorithm based
on distance correlation with an application to microarrays, IEEE/ACM TCBB 16 (2019)
1802–1815. doi:10.1109/TCBB.2018.2833482.
[25] M. Manca, L. Boratto, V. Morell Roman, O. Martori i Gallissà, A. Kaltenbrunner, Using social
media to characterize urban mobility patterns: State-of-the-art survey and case-study,
OSNEM 1 (2017) 56–69. doi:https://doi.org/10.1016/j.osnem.2017.04.002.
[26] A. Javadian Sabet, M. Brambilla, M. Hosseini, A multi-perspective approach for analyzing
long-running live events on social media. a case study on the “big four” international
fashion weeks, Online Social Networks and Media 24 (2021) 100140. doi:https://doi.
org/10.1016/j.osnem.2021.100140.
[27] M. Brambilla, A. Javadian Sabet, M. Hosseini, The role of social media in long-running
live events: The case of the big four fashion weeks dataset, Data in Brief 35 (2021) 106840.
doi:https://doi.org/10.1016/j.dib.2021.106840.
�