Understanding the Role of Historical Context in a Point of Interest
Recommendation System
Paulo Pombinho, Ana Paula Afonso and Maria Beatriz Carmo
Faculdade de Ciências, Universidade de Lisboa, Campo Grande, Lisboa, Portugal
Keywords: Context-Aware, Mobile Devices, Degree of Interest Function, Evaluation, Information Visualization.
Abstract: The increasingly large quantities of points of interest make choosing between all the available information a
painful task for the users. This limitation is aggravated by the reduced screen space of most mobile devices.
To minimize these issues, it is fundamental that the information shown to the user is relevant, helping them
in making good choices and decisions. We present a two phase evaluation of an adaptive degree of interest
function that uses location and temporal contexts combined with the historical context of the previous
searches to quantify the relevance of the points of interest shown to the user.
1 INTRODUCTION
Applications designed for the search of points of
interest in the vicinity of the users have become very
popular, allowing users to make good choices when
performing tasks like finding relevant locations in
their vicinity. However, despite their undeniable
usefulness, the increasingly huge amount of
information hinders the users to correctly perceive
all that is shown to them (Heimonen, 2002).
Furthermore, when coupled with the mobile device’s
small screen space available, a correct visualization
of the information is greatly hampered.
For this reason, it is essential that mobile
information visualization applications enforce that
what is shown on the screen is truly relevant for the
user (Holtzblatt, 2005), and it is fundamental to
include recommendations that guide the users in
choosing amongst the available information.
Recommender systems have been a popular
research topic, and are used in large online stores
(Jannach et al., 2012). These systems rely on
customers providing ratings and can be divided in
two different types: single rating systems, which
calculate an overall rating for each product for each
user, and multi-criteria systems that rate, not only,
an overall relevance, but also additional criteria /
attributes (Adomavicius and Kwon, 2007).
However, traditional recommender systems do
not take into account richer contexts, such as the
type of location or time of day, easily obtained using
current mobile devices (Adomavicius et al., 2011).
The adaptation to context dimensions is a key
feature to mitigate the limitations in the usability of
small screens. According to Reichenbacher (2008),
adaptive visualization concerns the adjustment of all
components of the visualization process, according
to a particular context. This principle is especially
important to increase the usability of searching
information in mobile devices and to reduce the
cognitive load inherent to mobile usage contexts.
In a previous work, we proposed a degree of
interest function (DOI) that uses information on the
user’s preferences and location to estimate the
relevance of each POI (Carmo et al., 2008). Despite
being considered useful, the user’s preferences had
to be explicitly stated by them and, consequently,
there was a need to specify several attributes and
weights, causing the DOI to be deemed confusing
(Pombinho et al., 2009).
Using the adaptive principle to solve the
limitations of the DOI function and reduce its
cognitive load, we proposed an adaptive degree of
interest function (ADOI) that avoids the necessity of
specifying a large number of attributes, while still
maintaining, if not improving, the calculated
relevance estimation (Pombinho et al., 2012).
In this paper we evaluate the ADOI function, to
understand if it increases the usability of the system,
while also improving the relevance calculation.
The next section presents a brief overview of our
ADOI function, its distance functions and the
adaptive historical context used.
537
Pombinho P., Afonso A. and Carmo M..
Understanding the Role of Historical Context in a Point of Interest Recommendation System.
DOI: 10.5220/0004276305370541
In Proceedings of the International Conference on Computer Graphics Theory and Applications and International Conference on Information
Visualization Theory and Applications (IVAPP-2013), pages 537-541
ISBN: 978-989-8565-46-4
Copyright
c
 2013 SCITEPRESS (Science and Technology Publications, Lda.)

2 ADAPTIVE DOI
One reason why the DOI was considered confusing
was the need for the users to specify a large set of
attributes for each query and understand and specify
the weights for each attribute. To overcome these
limitations we proposed an ADOI function that
included enhancements to the previous DOI. These
modifications aim to avoid the specification of a
large number of attributes, while making part of this
process automatic and transparent to the user.
We will briefly describe these enhancements,
however, they are described in more detail in
(Pombinho et al., 2012).
As important as understanding what points of
interest exist in the vicinity of the user is to identify
which are open by the time the user gets there. As an
example, if the user is searching for a gas station, it
is not useful to display results that might not be open
when the user finally arrives there. For this reason,
we have added a new temporal distance function and
time attribute to the adaptive DOI function.
To avoid the need, for the user, to specify all the
different attributes and weights of the DOI, and
reduce the inherent cognitive load, the ADOI uses an
historical context that will enhance the queries, by
automatically specifying attributes using the
information from the user’s previous queries.
For each pair (attribute type, attribute value) we
store a count of how many times it was queried.
Whenever the user makes a query, the attributes
specified and their values are updated in the internal
database. This historical log allows a summary of
the interest of the user to be assembled over time.
For instance, if the user always goes to Italian
restaurants, it is possible to use this information to
automatically specify the “type of restaurant”
attribute without further action from the user.
The user preferences and the type of searches
made by them are, however, not always the same
according to the location and temporal contexts. For
this reason, we allow users to define geographical
areas that are relevant for them (for example, a work
or a home area). When the user performs a query,
the logs are recorded in the appropriate geographical
area / time of day section. Whenever a new query is
made, the attributes are automatically adapted to the
user’s current location and temporal contexts.
3 USER EVALUATION
Since the proposed ADOI relies on an historical
context that represents each user’s implicit
preferences, obtained from previous interactions
with the system, we need this information to
correctly perform an evaluation of the functions.
The best approach would be to have a significant
number of users trying the application prototype
during several months, slowly building the historical
context to match their interests. However, due to
consequent need of a very large number of available
trial mobile devices and also due to time constraints,
this approach was not deemed practical.
Another approach would be to deploy an
application, and have a large number of users
performing the evaluation on their own mobile
devices. However, for it to be feasible, we would
need POI data for an extensive area with information
about the POI attributes which, again, was not
possible.
For this reasons, to obtain, the user’s general
interests, we conducted a preliminary evaluation.
3.1 Phase One
In our study we had 13 participants, six male and
seven female, with ages ranging between 21 and 62.
Since the goal of this phase was to obtain the
user’s interests, it consisted of a questionnaire that
evaluated three different scenarios and also the
general preferences. To avoid biasing the evaluation,
the users were not briefed about the true purpose of
the study; instead they were only informed that it
was a survey about restaurants.
Despite having, as a goal, obtaining information
for the second phase of the evaluation, some relevant
data was obtained.
Regarding the general preferences of the users, it
was interesting to find that “Type” and “Price” were
by far considered the most important. On the
contrary, the existence of a “Seafront” and the
“Classification” were considered less important.
Figure 1: Average classification and standard deviation for
each attribute / scenario.
Enforcing our idea that different contexts create
different needs, we had 62% of the users state that
they had different preferences in the three scenarios
0%
20%
40%
60%
80%
100%
General
Lunch/
Work
Dinner/
Home
Vacation
IVAPP2013-InternationalConferenceonInformationVisualizationTheoryandApplications
538

and, for most the closer scenario was the search near
home at dinner. Finally, when we consider all the
answers for the different scenarios, we can conclude
that there are some attributes that have significant
differences between them (Figure 1).
Although, as stated, our main goal in this phase
was to obtain data to allow the second phase of the
evaluation, the results obtained, despite covering a
reduced number of users, do reveal a strong
tendency for different interests in different contexts.
3.2 Phase Two - Hypothesis
The main objective of this evaluation phase was to
understand if the proposed concepts were easily
comprehended by the users and if they achieved a
relevance calculation that, when compared to the
actual interest of the users in specific POI, had
closer results than the standard DOI.
From our understanding, and the previous
analysis of the first phase of the evaluation, we
considered five hypotheses:
H1 – The Exploratory DOI will be considered more
useful in scenarios where the user wants to choose a
different POI from those already known.
H2 – Both types of users (with and without
preference differences) will prefer the Adaptive
DOI, finding it easier to use.
H3 –The temporal distance will be more used than
the geographical distance.
H4 – The Standard DOI will have similar results to
the Adaptive DOI for users with no different
preferences for different scenarios.
H5 – The Adaptive DOI will obtain closer results
(both in the order and value) to the true preferences
of the users who have different preferences for
different scenarios.
3.3 Phase Two - Procedure and Tasks
Since the participants in the second phase were the
same from the first, and to minimize biasing from
one phase to the other, we had a time lapse of over a
month. Furthermore, using information obtained
from the first phase, we instantiated, for each user,
the Historical database with values resembling the
users interest for each scenario.
This evaluation phase consisted of five tasks:
First Task – In the initial task, users were presented,
for each three different scenarios (used in the first
phase), with a list of POI and their attributes. They
were then asked to order them by giving them a
classification from 0 to 100%. This task has the
objective of validating H4 and H5, allowing the
comparison of both the order and values given by
the users, to the values obtained by the DOI and the
ADOI, for similar scenarios.
Second and Third Tasks – The second task was an
introductory task, where each concept was briefly
explained to the users and enough time was given
for them to freely experience with the application.
Similarly, task three allowed the user to test the
three DOI modes: Standard DOI, Exploratory DOI
and Adaptive DOI.
Fourth Task –The fourth task placed the users in a
scenario were implicitly they were told to find
“something new” and also that they were short on
time. This task had the objective of finding out, if
most users choose to use the Exploratory DOI, thus
validating H1. Furthermore, the use of the temporal
distances was also analysed (for validation of H3).
Fifth Task – Finally, in the last task the user had to
search for a restaurant near their workplace, at noon.
The objective of this task was to understand how
many users choose each DOI mode. From the
information obtained we intended to understand if
H2 and H3 are valid.
3.4 Phase 2 - Results and Discussion
To perform the comparison between the relevancies
given, by the users, to each POI, and the values that
were obtained by the Standard DOI and the
Adaptive DOI, we subtracted the values stated by
the users from both the DOI and the ADOI values,
and compared the results.
We found that, for both, the relevancies
calculated by the algorithms are, on average, lower
than those given by the users. This fact is probably
due to the penalizing nature of both DOI functions,
when compared to the way the users classify the
POIs. One user stated that by simply being a
restaurant he would classify it with a 50% value.
Figure 2: DOI vs. ADOI: average value differences.
Concerning the absolute relevance value
differences (Figure 2), both functions had deviations
from the values given by the users (around 20%).
22,6
24,5
19,7
18,1
19,1
16,6
0
10
20
30
Total WithPref. WithoutPref
ValueDifferences(AbsoluteAverage)
Standard
DOI
Adaptive
DOI
UnderstandingtheRoleofHistoricalContextinaPointofInterestRecommendationSystem
539

However, when we consider only the differences in
the order (Figure 3) we have, on average a
difference under 1, which indicates that it is, in
general, in accordance with the user’s preferences.
In all the scenarios considered, the Adaptive DOI
had better results than the Standard DOI, obtaining
both values and order that better matched the values
indicated by the users. This confirms H5; however,
surprisingly H4 is disproved since even for users
with no difference in preference, the Adaptive DOI
had better results. In fact, the results obtained with
the different groups of users do not shown a
significant difference between them.
Figure 3: DOI vs. ADOI: average order differences.
To validate H1, we examined how many users
chose to perform their query of task 4 with the
Exploratory Mode. Despite the slightly worse
usefulness classification, only two participants used
other DOI modes, thus validating H1.
Concerning H2, the hypothesis is only partially
validated. While in task 5, two thirds of the
participants preferred to use the Adaptive DOI,
when asked which one they would prefer, we
obtained mixed responses, with an equal number of
users preferring each mode. Instead, more than half
the users prefer to have both functions available.
This is, in part, contrary to what we would suppose,
since the Adaptive mode consistently obtains results
that better match the user’s classifications.
Finally, regarding H3, our results partially
contradict our hypothesis. Despite being, in general
more used than the geographical distance, when we
asked the participants which one they would prefer,
we had twice as many participants choosing the
geographical distance. It should, however, be
stressed that, more than half the participants would
prefer to have both distances available.
4 CONCLUSIONS AND FUTURE
WORK
Our work provides evidence that user preferences
change, sometimes significantly, depending on the
context in which they are (both temporal and
geographical).
We can also clearly witness an improvement in
both the values and the order of the POIs when using
the adaptive DOI. This improvement suggests that
the use of richer contextual information can
significantly improve the way applications model
and identify the user interests, enabling a better
selection of the information presented to the user
and its relevance. By having a better judgment on
the choice of presented information, and displaying
it in a way that more closely resembles the frame of
mind of the user, we can considerably reduce the
cognitive load associated with these systems and
increase their usability.
We also witnessed some classifications by the
users that raised interesting questions. For example,
one of the users classified a restaurant with 0%,
because the Type was vegetarian, and the user really
disliked that type of food. This hints that, possibly,
there should not only exist positive preferences, but
also, negative ones.
Regarding future work, we intend to explore a
number of different contexts that could also be used
to further filter and partition the Historical Context
database. The use of the current climate conditions
in the area of the user, for example, can alter the
preference for restaurants with or without a seafront,
depending on whether it is raining or not. Similarly,
when the users were considering the vacation
scenario in a different country, many expressed the
desire to choose the restaurant type as “typical”.
This indicates that the notion of being abroad (easily
found by analyzing the user coordinates) can also
significantly alter the user preferences.
REFERENCES
Adomavicius, G., Kwon, Y., 2007. New Recommendation
Techniques for Multi-Criteria Rating Systems. In
Intelligent Systems, IEEE, 22 (3), 48-55.
Adomavicius, G., Mobasher, B., Ricci, F., Tuzhilin, A.,
2011. Context-Aware Recommender Systems. In AI
Magazine, 32 (3), 67-80.
Carmo, M. B., Afonso, A. P., Pombinho, P., Vaz, A.,
2008. Visualization of Geographic Query Results for
Small Screen Devices. In Proc. of the Visual 2008,
LNCS 5188, 167-178.
Heimonen, T., 2002. Information Visualization on Small
Display Devices. Master Thesis. Department of
Computer Sciences, University of Tampere.
Holtzblatt, K., 2005. Designing for the Mobile Device:
Experiences, Challenges and Methods. In
Communications of the ACM, 48 (7), 33-35.
0,83
0,85
0,80
0,57
0,55
0,61
0,0
0,5
1,0
Total WithPref. WithoutPref
OrderDifferences(AbsoluteAverage)
Standard
DOI
Adaptive
DOI
IVAPP2013-InternationalConferenceonInformationVisualizationTheoryandApplications
540

Jannach, D., Karakaya, Z., Gedikli, F., 2012. Accuracy
Improvements for Multi-Criteria Recommender
Systems. In Proc. of the EC’12, 674-689.
Pombinho, P., Carmo, M. B., Afonso, A. P., 2009.
Evaluation of Overcluttering Prevention Techniques
for Mobile Devices. In Proc. of the IV’09, 127-134.
Pombinho, P., Carmo, M. B., Afonso, A. P., 2012. Context
Aware Point of Interest Adaptive Recommendation. In
Proc. of the CARR 2012 workshop, in Conjunction
with IUI 2012.
Reichenbacher, T., 2008. Mobile Usage and Adaptive
Visualization. Encyclopedia of GIS. Part 16.
Heidelberg: Springer Verlag, 677-682.
UnderstandingtheRoleofHistoricalContextinaPointofInterestRecommendationSystem
541