=Paper=
{{Paper
|id=Vol-1743/ks3
|storemode=property
|title=Personalised Recommendations for Context Aware Suggestions
|pdfUrl=https://ceur-ws.org/Vol-1743/ks3.pdf
|volume=Vol-1743
|authors=Fabio Crestani
|dblpUrl=https://dblp.org/rec/conf/simbig/Crestani16
}}
==Personalised Recommendations for Context Aware Suggestions==
https://ceur-ws.org/Vol-1743/ks3.pdf
Personalised Recommendations for Context Aware Suggestions
Fabio Crestani
Faculty of Informatics
Universitá della Svizzera Italiana (USI)
Lugano, Switzerland
fabio.crestani@usi.ch
Abstract that, appropriately modified, can help provide so-
lutions to these tasks.
Modern Information Retrieval has moved Recent years have witnessed an increasing use
from standard text retrieval to novel ap- of location-based social networks (LBSNs) such
plications of the same technology. Con- as Yelp, TripAdvisor, and Foursquare. These so-
textual suggestion is an example of this cial networks collect valuable information about
type of applications. The TREC Contex- users’ mobility records, which often consist of
tual Suggestion track addresses the prob- their check-in data and may also include users’
lem of suggesting contextually relevant at- ratings and reviews. A service that could be of
tractions to a user visiting a new city based interest to users of such networks could be related
on his/her recorded preferences from past to providing them recommendation of location to
visits to other cities. In this invited talk visits. In fact, being able to recommend person-
I will reframe the problem of represent- alised venues to users plays a key role in satisfying
ing and using context and briefly report the user needs on such social networks.
our two past approaches to capturing the Recent research on recommending systems has
user profile to enable a system to provide focused on using collaborative-filtering technique,
more accurate and relevant recommenda- where the system recommends venues based on
tions. The results of our participation in users’ data whose preferences are similar to those
the 2013 and 2015 TREC tracks, report- of the target user. Collaborative-filtering ap-
ing how we can use such contextual infor- proaches are very effective, but they suffer from
mation as geographical location, time, and the cold-start (i.e., they need to collect enough in-
friends’ interests, show that our system not formation about a user for making recommenda-
only significantly outperforms the base- tions) and the data-sparseness problems. Further-
lines method, but also performs better than more, these approaches rely mostly on check-in
most other participants to that track, man- data to learn the preferences of users and such in-
aging to achieve the best results in nearly formation is often insufficient to get a complete
all test contexts. picture of what the user likes or dislikes of a spe-
cific venue (e.g., the food, the view, the music).
1 Introduction In order to overcome this limitation, recent ap-
proaches try to model the users by applying a
The research ara of Information Retrieval (IR), deeper analysis on users’ past ratings as well as
historically concerned with retrieving information their reviews. In addition, following the principle
from large archives in response to a user query, of collaborative filtering, they exploit the reviews
has been evolving rapidly in recent years. This of different users with similar preferences.
evolution has brought IR researchers to deal with
problems that are very different from standard IR, 2 Contextual Suggestion
like for example Topic Detection and Tracking,
Blog and Tweet retrieval, Knowledge Base Accel- The TREC Contextual Suggestion Track started in
eration, Temporal Summarisation, Novelty Detec- 2012 and continued to 2016, the current year. It
tion, etc. IR provides a large number of techniques investigates search techniques for complex infor-
19
ences, the physical environment, and the social sit-
uation affecting the user. Capturing it, enables to
differentiate from moment to moment in the life of
a user, providing better suggestions.
Our approach to the TREC Contextual Sugges-
tion task involved using external information to
enrich the available information about the user and
the user’s context. In 2013 we enrich the user ge-
ographical context (i.e. his location in time and
space), while in 2015 we enrich the contextual in-
formation about the different venues and the opin-
Figure 1: The TREC Contextual Suggestion Sce- ion of different users (i.e. his social context), to
nario. make it possible to provide more valuable sugges-
tions.
mation needs that are highly dependent on con- 3.1 TREC 2013
text and user interests. The task was to take the
representation of these user interests (profiles) and In (Rikitianskiy et al., 2014) we described our ap-
contexts and to produce a list of ranked sugges- proach for TREC 2013, aimed at making context-
tions for each profile-context pair. The scenario sensitive recommendations to tourists visiting a
used consistently by the track was that of a user new city. We presented a new approach to recom-
visiting a new city and receiving suggestions of mending places to users incorporating geograph-
places (e.g. bars, restaurants, museums, etc.) to ical information as context and exploiting data
visit based on what the new city made available from multiple sources. Our method is based on
and his preferences as extracted from the user pro- quite a simple strategy of using the descriptions
file (see figure 1). A full description of the task can of previously rated places in closed geographical
be found (Dean-Hall et al., 2013) . The similarities proximity to build user profiles. We also intro-
with collaborative filtering are obvious, the only duced a number of novel additions which have
difference is that we know too little about each in- clearly lead to improved performance. In fact,
dividual user to be possible to use any good col- the analysis of the results from the TREC evalua-
laborative filtering algorithm. Obviously the track tions performed by a large group of users, demon-
evolved over the years, slightly changing the geo- strated the high level of performance delivered by
graphical context and providing richer users’ pro- our method, showing that it is able to significantly
files, but still making it impossible to use well es- outperform the two track baselines and all other
tablished collaborative filtering algorithms. track entrants in the majority of cases. In fact,
when compared to the 34 other competing systems
In the following we report on the approaches we
in the track, it delivered results which were well
followed for our 2013 and 2015 participations to
above the median. In nearly half of all contexts,
this track1 and on the use of external information
our approach was able to deliver the best set of
to enlarge the user profile to make it possible to
results, confirming that the choices made during
provide more effective contextual suggestions.
the development of the system were sensible and
3 On the Use of External Information for beneficial. More details can be found in the above
cited paper.
Contextual Suggestion
3.2 TREC 2015
Context has a very loose definition in the area of
IR. It is related to all aspects that influence the user The TREC 2015 Contextual Suggestion Track
perception of an information need or of the rel- changed little compared to previous years, but
evance of a document to such information need. we experimented a quite different approach.
This includes the time, the location, the prefer- In (Aliannejadi et al., 2016) we presented a novel
method for suggesting venues to users, where the
1
In 2014 we did not take part as the author was on sab- users are modelled based on venues’ content as
batical. We are also taking part in the 2016 track, but the
results have yet to be released, so we will not comment on well as other users’ reviews of the same venues.
the approach taken. For the former we use the categories of the venues
20
enriched by keywords extracted from users’ online Acknowledgments
reviews, which provide a more detailed descrip-
The work reported in this paper was done in col-
tion of the venue itself. Although the venue in-
laboration with my PhD students at the Faculty of
formation is valuable for inferring “what type” of
Informatics of the Universitá della Svizzera Ital-
places a user may like or dislike, it does not give
iana and, in particular, with Mohammad Nejadi.
any clue on the reasons “why” a user rated as pos-
The work is financially supported by the Swiss Na-
itive or negative a particular venue. We needed to
tional Science Foundation (SNSF) under the grant
exploit the user’s opinions in order to understand
”Relevance Criteria Combination for Mobile In-
what the user may have appreciated of a place and
formation Retrieval (RelMobIR)”.
to get better recommendations for future venues.
One way to obtain these opinions is mining the
users’ reviews and see how much they liked the References
venue and, more importantly, for which reasons:
M. Aliannejadi, I. Mele, and F. Crestani. 2016. User
was it for the quality of food, for the good ser- model enrichment for venue recommendation. In
vice, for the cozy environment, or for the loca- Proceedings of the Asian Information Retrieval Sym-
tion? In cases where we lacked reviews from some posium (AIRS), Beijing, China, December.
of the users (e.g., they rated a venue but omitted
A. Dean-Hall, C.L.A. Clarke, N. Simone, J. Kamps,
to review it) and therefore could extract opinions, P. Thomas, and E.M. Voorhees. 2013. Overview of
we applied the collaborative-filtering principle and the TREC 2013 contextual suggestion track. In Pro-
sed reviews from other users with similar interests ceedings of The Twenty-Second Text REtrieval Con-
and tastes. Our intuition was that a user’s opinion ference, TREC 2013, Gaithersburg, Maryland, USA,
November.
regarding an attraction could be learned based on
the opinions of others who expressed the same or A. Rikitianskiy, M. Harvey, and F. Crestani. 2014.
similar ratings for the same venue. To do this we A personalised recommendation system for context-
aware suggestions. In Proceedings of the Euro-
exploited information from multiple sources (e.g. pean Conference in Information Retrieval Research
Yelp and Foursquare) and combine them to gain (ECIR), pages 63–74, Amsterdam, The Netherlands,
better performance. In the cited paper we showed March.
how our model outperforms all the other runs by a
significant margin and was placed as the first run
in the track. See the paper for details on the tech-
nique used.
4 Conclusions and Future Work
The importance of context in IR has long been
recognised and context has been used in many dif-
ferent applications of IR and related fields. Con-
textual suggestion is a difficult problem because
of the many and different factors that make up the
context and that have an influence on the effective-
ness of the suggestion. Considering all available
factors and, of course, finding an effective combi-
nation of them is the best approach, but it needs
to be personalised and efficiently computed to be
effective. This is the current direction of research
of my research group in the context of a couple
of project we are involved in. This invited talk
reported on the successful results of our participa-
tion in TREC 2013 and 2015 and on how we used
context as an effective mean to provide better sug-
gestions.
21