=Paper=
{{Paper
|id=None
|storemode=property
|title=Context-Aware Recommender Systems: A Comparison Of Three Approaches
|pdfUrl=https://ceur-ws.org/Vol-771/paper6.pdf
|volume=Vol-771
|dblpUrl=https://dblp.org/rec/conf/aiia/PannielloG11
}}
==Context-Aware Recommender Systems: A Comparison Of Three Approaches==
https://ceur-ws.org/Vol-771/paper6.pdf
Context-Aware Recommender Systems: A
Comparison Of Three Approaches
Umberto Panniello and Michele Gorgoglione
Politecnico di Bari (Italy), Viale Japigia 182
Tel. +39 080 5962765
Fax: +39 080 5962766
u.panniello@poliba.it, m.gorgoglione@poliba.it
Abstract. Methods for generating context-aware recommendations were classified
into the pre-filtering, post-filtering and contextual modeling approaches. This pa-
per proposes a novel type of contextual modeling, that is called contextual neigh-
bors, based on the idea of using context to compute the neighborhood in a colla-
borative filtering approach, and introduces four variants of this method. In
addition, the paper presents the results of the comparison among these four ap-
proaches and among the contextual neighbors approach to the other contextual
approaches and to the un-contextual one. While some of these methods have been
studied independently, few prior research has compared their performance to de-
termine which of them is better.
1 Introduction
Most traditional Recommender Systems (RSs) provide recommendations of items
to users and vice versa and do not take into consideration the circumstances and
other contextual information when recommendations take place. For example,
when an online travel agency recommends a vacation package, it is important to
know when the person plans to go on vacation. Recently, some companies started
taking into account the contextual information (i.e., Sourcetone). In academia,
several studies, such as [1], demonstrated that context induces important changes
in a customer purchasing behavior. Experimental research on customer modeling
suggests that including context in a customer behavior model improves the ability
to predict her behavior in some cases because it allows the identification of more
homogeneous patterns in the data describing the purchasing history of a customer
[2]. Therefore, the accuracy of predicting consumer preferences should depend on
the degree to which we have incorporated the relevant contextual in-formation.
The usage of contextual information in context-aware recommender systems
(CARS) can be broadly classified into two groups: (1) recommendation via con-
text-driven querying and search, and (2) recommendation via contextual prefe-
2
rence elicitation and estimation. The approach based on the context-driven query-
ing and search has been used by several mobile and tourist recommender systems,
e.g., [3], [4], that typically use contextual information to query or search a certain
repository of resources (e.g., restaurants) and present the best matching resources
(e.g., nearby restaurants that are currently open) to the user.
The other approach to CARS, that we follow in this paper, is based on contextual
preference elicitation and estimation, e.g., [5], [6], [7,8], including elicitation and
estimation of ratings of various items provided by different users. This approach
can be traced back to the work in [6] and [9]. [9] hypothesized that the inclusion
of knowledge about the user’s task into the recommendation algorithm in certain
applications can lead to better recommendations. [6] described a way to incorpo-
rate the contextual information into recommender systems by using a multidimen-
sional approach in which the traditional 2-dimensional (2D) User/Item paradigm
was extended to support additional contextual dimensions, such as Time, Location
and Company. Since then, several contextual preference elicitation and estimation
approaches to CARS have been proposed, all of them emphasizing the need to
model and learn user’s context-sensitive preferences. Many of these methods are
reviewed in [10] and [7]. Once the context-sensitive preferences of users are
learned, recommendations are generated by either adapting the existing collabora-
tive filtering, content-based, or hybrid recommendation methods to context-aware
recommendation settings or by developing novel intelligent data analysis tech-
niques from data mining and machine learning. Furthermore, several scholars [5],
[7,8] have shown that adding contextual information helps to improve estimations
of unknown ratings in this approach. For example, [5] and [6] described a way to
include the contextual information by using a multidimensional approach, as it
was mentioned above. Also, it was shown in [5] that the multidimensional contex-
tual information does matter and can lead to better recommendations in compari-
son to the traditional un-contextual 2D recommender systems.
2 Pre-filtering, post-filtering and contextual modeling
In this paper, we use implicit ratings following the approach outlined by [11],
[12], [13]. We measure the utility of product j for user i with the purchasing fre-
quency xij specifying how often user i purchased product j. Unlike the traditional
two-dimensional (2D) recommender systems that try to estimate unknown ratings
in the Users × Items matrix, where Users and Items are the sets of users and items
respectively, context-aware recommender systems (CARS) also take into account
contextual information in a Users × Items × Context matrix. Context is a set of
contextual attributes C each having a hierarchical structure defined by a set of q
atomic attributes, i.e., C = (C1,…, Cq) [10]. Further, the values taken by attribute
Cq define finer (more granular) levels, while C1 coarser (less granular) levels of
3
(a) (b) (c)
Fig. 1 How to use context in the recommendation process
contextual knowledge [14]. For example, Fig. 2 presents the hierarchy for the con-
textual attributes used in our analysis (see Section 3).
In this paper we compare the three contextual approaches proposed in [10] (also
shown in Errore. L'origine riferimento non è stata trovata.), that start with
“Data” on users, items, ratings and contextual information (“Context”) and results
in generating context-specific recommendations, are presented. In the contextual
pre-filtering (PreF) the contextual information is used to filter out irrelevant rat-
ings before they are used for computing recommendations with standard (2D) me-
thods. In the contextual post-filtering (PoF) context is used after the classical (2D)
recommendation methods are applied to the standard recommendation data. In the
contextual modeling (CM) context is used inside the recommendation-generating
algorithms. The work presented in [10] helped researchers to understand different
aspects of using the contextual information in the recommendation process. How-
ever, [10] did not examine which of these methods are more effective for provid-
ing contextual recommendations, and therefore, left this important topic un-
addressed and without any prescriptive recommendations for which method to
use. In this paper we empirically compare the three contextual approaches, i.e.,
pre-filtering, post-filtering and contextual modeling among themselves to deter-
mine which one is better, and also with the un-contextual approach when contex-
tual information is ignored and not used in the recommendation methods. We also
present several types of contextual post-filtering and contextual modeling methods
and evaluate their performance in order to identify the best-performing methods in
these two categories. We show that one particular post-filtering method dominates
other approaches. However, this dominance comes with a certain complexity
“cost” that prevents that method to be a clear winner in all the practical settings.
(a) (b) (c)
C1
C2
Fig. 2 Hierarchical structure of the contextual attributes (a) Time Of The Year, (b) Intent
Of Purchase and (c) Store
4
2.1 The pre- and the post-filtering approaches
According to the pre-filtering approach (Errore. L'origine riferimento non è
stata trovata.(a) and [10]), the contextual information is used as a label for filter-
ing out those ratings that do not correspond to the specified contextual informa-
tion. This filtering is done before the main recommendation method is launched
on the remaining data that passed the filter. In other words, if a particular context
of interest is C, then the pre-filtering method selects from the initial set of all the
ratings only those corresponding to the specified context C. As a result, it gene-
rates the Users × Items matrix containing only the data pertaining to context C.
Then the 2D recommendation method (e.g., collaborative filtering) is launched on
this remaining dataset that passed the filter to generate recommendations for con-
text C. We call this approach Exact Pre-Filtering (EPF). According to the post-
filtering (PoF) approach (see Errore. L'origine riferimento non è stata trova-
ta.(b) and [10]), we first ignore all the contextual information in the data and ap-
ply a traditional 2D recommendation method, such as collaborative filtering, on
the whole un-contextual data set, where the contextual information is dropped.
Once the unknown ratings are estimated using the 2D method on this data and the
un-contextual recommendations are produced, we “contextualize” these recom-
mendations as follows. Although there exist various methods for contextualizing
the 2D recommendations [10], in this paper we consider the following two ap-
proaches, called “Weight” and “Filter”. Both approaches analyze data for a given
user in a given context to calculate the probability with which the user chooses a
certain item in the given context. After that, the recommendations obtained using
this 2D method are “contextualized” by using the contextual probabilities. For de-
tailed information on the methods, see [15].
2.2 The contextual modeling approach
According to the contextual modeling approach, we present a new method called
contextual-neighbors CM. This approach is based on user-based collaborative fil-
tering (CF) and works as follows. First, for each user i and context c, we define
the user profile in context c, i.e. the contextual profile Prof(i, c). For example, if
the contextual variable c has two values (e.g., Winter and Summer), then we have
two contextual profiles for each user, one for the Winter and the other for the
Summer. Note that these contextual profiles can be defined in many different
ways, some of which are presented in [2], and our approach does not depend on
any particular choice of a profiling method. However, in the experimental study
described in Section 4 we use the following specific contextual profiling tech-
nique. As explained in Section 2, we follow the transaction-based approach to RSs
and measure the utility rijc of product j for user i in context c with the purchasing
5
frequency xijc specifying how often user i purchased product j in context c. Then
we use this measure to define contextual profile as Prof(i, c) = (ri1c , …, rikc).
We use these profiles to define similarity among users and also to define and find
N nearest “neighbors” of user i in context c, where “neighbors” are determined us-
ing contextual profiles Prof(i’, c’) and similarity measures between the profiles.
Although this similarity can be defined in various ways, we use a popular CF ap-
proach and define the similarity measure d using the cosine measure as:
∑risc ri' sc' (1)
d(Prof(i' , c' ), Prof(i, c)) =
Prof(i, c) • Prof(i' , c' ) ∈
s S ii ' c
2 2 =
Prof(i, c) × Prof(i' , c' ) ∑risc2 ∑ri'2sc'
s∈S s∈S
ii ' c ii ' c
where risc’ and ri’sc’ are the ratings of item s assigned by user i and user i’ respec-
tively in context c and c’. Sii’c={s ∈ Items | risc ≠ ∅ ∧ ri’sc’ ≠ ∅} is the set of all
items co-rated by both user i and user i’ in context c.
Then we find N nearest neighbors for the (i, c) pair by identifying pairs (i’, c’)
such that d(Prof(i’, c’), Prof(i, c)) is the largest among all the candidate pairs (i’,
c’) subject to the following constraints:
• Mdl1: there are no constraints on the set of (i’, c’) pairs, and we select N pairs
that are the most similar to (i, c).
• Mdl2: we select an equal proportion of pairs (i’, c’) corresponding to each
context c (e.g., if the contextual variable has only two values, Winter and Summer
respectively, and the neighborhood size is 80, we select 40 neighbors from Winter
and 40 from Summer).
• Mdl3: we select N pairs (i’, c’) that are the most similar to (i, c) corresponding
to each context c at the same level of the context of interest (e.g., if the context of
interest is “Winter Holiday” in Fig. 2(a), we select the neighborhood by using only
profiles referred to level C2 of that contextual variable).
• Mdl4: we select an equal proportion of pairs (i’, c’) corresponding to each
context c at the same level of the context of interest (e.g., if the context of interest
is “Winter Holiday” in Fig. 2(a) and the neighborhood size is 80, we define the
neighborhood by using 20 users from the context “Winter Holiday”, 20 users from
the context “Winter Not Holiday”, 20 users from the context “Summer Holiday”
and 20 users from the context “Summer Not Holiday”).
3 Experimental Setup
We compared the pre-filtering, post-filtering, contextual modeling and un-
contextual recommendations across a wide range of experimental settings. First,
we selected three different data sets. The first dataset (DB1) comes from an e-
commerce website commercially operating in an European country which sells
electronic products to approximately 120,000 users and contains about 220,000
6
purchasing transactions during an observation period of three years. For this data-
set, we selected the time of the year as a contextual variable. Its hierarchical struc-
ture is presented in Fig. 2(a). The classification into Summer or Winter and Holi-
day or Not Holiday is based on the experiences of the CEO of the e-commerce
website that we used in our study. The data was pre-processed by excluding cus-
tomers who made only one single transaction or had abnormal behavior or had
transactions either only in the first two years or only in the third year. The result-
ing dataset contained about 1,500 users and about 10,000 transactions thus having
about 7 transactions per user.
The second dataset (DB2) is taken from the study described in [2]. First, a special
purpose browser was developed to help users navigate Amazon.com website and
purchase products on its site. This browser was made available to a group of stu-
dents who were asked to navigate and simulate purchases on Amazon.com during
a period of four months. Once a product was selected by a student to be pur-
chased, the browser recorded the selected item, the purchasing price and other use-
ful characteristics of the transaction and this information was stored in the data-
base. In addition, the student was asked at the beginning of each browsing session
to specify its context, which was the intent of a purchase in our case. The structure
of this contextual variable IntentOfPurchase is presented in Fig. 2(b). Further, the
data was pre-processed by excluding the students who made less than 40 transac-
tions and eliminating the students who had any kind of misleading or abnormal
behavior. The resulting number of students was 556, and the total number of pur-
chasing transactions for the students was 31,925 thus having about 57 transactions
per user.
The third dataset (DB3) comes from an e-commerce website which sells comics
and comics-related products, such as T-shirts, DVDs and gadgets. It contains
about 50,000 transactions and 5,000 users thus having about 10 transactions per
user. We used the store (i.e., the section in the Web site where products are
bought), as a contextual variable (see Fig. 2(c)). The importance of this contextual
variable comes from the expectation that customers’ behavior changes when navi-
gating and buying products in different sections of the Web site.
In our study, we recommend product categories instead of individual items be-
cause the e-commerce applications that we consider have very large numbers of
items (hundreds of thousands or even millions). We tried different item aggrega-
tion strategies and found that the best results are for 14 categories for DB1, 24 cat-
egories for DB2 and 136 for DB3. We aggregated items into categories according
to the classification provided by the Web sites product catalogue. Estimations of
unknown utilities were done by using a standard user-based collaborative filtering
(CF) method [16]. According to the CF approach, the neighborhood was formed
using the cosine similarity [17]. The neighborhood size N was set to N = 80 users,
which proved to be the optimal size in our experiments.
The experiments were performed for datasets DB1, DB2 and DB3 for all the le-
vels of contextual knowledge (un-contextual, C1 and C2), as presented in Fig. 2.
We have performed t-tests in order to determine if the chosen contextual variables
7
matter. The results of these tests were statistically significant (at 95% level) and
demonstrated that the contextual variables TimeOfTheYear, IntentOfPurchase and
Store matter. We used Precision, Recall, F-Measure, Mean Absolute Error (MAE)
and Root Mean Square Error (RMSE) [18] as the performance measures in our ex-
periments, as done in [5]. We computed MAE and RSME in a standard way by
taking absolute and squared differences between the estimated and actual values
of estimated utilities since the utilities are measured using discrete variables [19],
[20], [21]. Finally, we divided each dataset into the training and the validation
sets, the training set containing 2/3 and the validation set 1/3 of the whole dataset.
For the DB1 dataset, the first two years were the training set and the third year
was the validation set. For the DB2 dataset, we randomly split it in 2/3 for the
training set and the remaining 1/3 for the validation set (in this case, it was im-
possible to make a good temporal split because all the transactions were made
within a couple of months). For the DB3 dataset, the first nine months were the
training set and the last three months were the validation set.
4 Results
Firstly, we compare the traditional un-contextual 2D collaborative filtering and the
exact pre-filtering (EPF) methods. This type of comparison has been studied in [5]
before, where it was shown that in certain cases the un-contextual approach domi-
nates, while in other cases the contextual one dominates the un-contextual method.
However, the work reported in [5] was done in the context of multi-dimensional
recommendations and only for one small dataset. In this study, we wanted to pro-
vide a more extensive comparison of the un-contextual and the pre-filtering me-
thods across several and bigger datasets and across numerous other experimental
settings. Furthermore, we also needed to conduct this comparison for the unifor-
mity reasons because we also do the comparison of the post-filtering and the con-
textual modeling approaches to the un-contextual approach in this paper.
We compared the un-contextual and the EPF contextual methods for the case of
user-based collaborative filtering, across the datasets DB1, DB2 and DB3, across
multiple levels of classification hierarchy (C1 and C2) and different performance
measures. We split the data into the training and testing sets as described in Sec-
tion 4. For the sake of brevity we only present a summary of these results in Table
1 where only the F-measure, averaged across the experimental setting, is reported
and Table 2 where only the MAE, averaged across the experiment settings, is re-
ported. Table 1 and Table 2 report all the accuracy gains (in terms of F-measure
and MAE, namely) across each recommender systems for DB1, DB2 and DB3
(negative values mean performance reduction). For example, the first row of Table
1 shows the performance gains (reductions), in terms of F-measure, for the un-
contextual RS vis-à-vis the EPF, Filter PoF, Weight PoF and Mdl1 methods. The
matrixes in Table 1 and Table 2 are anti-symmetric, as should be the case when
8
two methods are compared in terms of their relative performance. However, the
comparison among the other performance measures used in the experiments is
discussed throughout this section.
The EPF dominates the un-contextual case in some cases in terms of the F-
measure and is dominated in other cases, consistently with [5]. The intuitive ex-
planation is that when the contextual information is very granular, this results in a
more homogeneous rating prediction model but also leads to data sparsity issues
(only few context-specific ratings are used to build the model). In contrast, when
the contextual information is too coarse, we can have enough data to build the
model, but it is more heterogeneous. This conflict between homogeneity of the
model and rating sparsity produces mixed results when comparing EPF and the
un-contextual model. Furthermore, EPF outperforms the un-contextual model in
terms of MAE and RMSE (see Table 2).
Secondly, we compare the un-contextual and the post-filtering methods. Unlike
exact pre-filtering, there exist many post-filtering methods. The comparison can
depend very significantly on the choice. We compared the Weight PoF and Filter
PoF and the un-contextual method across various experimental conditions de-
scribed in Section 3. The Filter PoF dominates the un-contextual case across all
the levels of context for the F-measure: for DB1, the difference between contex-
tual and un-contextual models is 20% on average, for DB2 it is 27% on average
and for the DB3 it is 90%. On the contrary, Weight PoF is always dominated by
the un-contextual model. All this makes us to conclude that the comparison of the
post-filtering and un-contextual methods depends very significantly on the type of
the post-filtering method being used. This observation is not surprising because
the post-filtering method takes the results of the same 2D un-contextual method
and reorders them based on the context-based post-filtering heuristic. If the heuris-
tic is “good” (such as Filter), then this re-ordering improves recommendation
quality; otherwise, it can make the results even worse (as for heuristic Weight).
Thirdly, we compare the un-contextual and the contextual modeling methods. We
consider four methods Mdl1, Mdl2, Mdl3, Mdl4 in our study and compare them to
the un-contextual method across the experimental conditions. The experiments
proved that the performances of the four CM approaches are very similar and the
differences are not statistically significant. For this reason we only present the per-
formance of Mdl1 because it slightly dominates other CM methods in some cases.
This makes sense because the N neighbors are selected for Mdl1 in an uncon-
strained manner, whereas they are selected based on various constraints for the
other three approaches. As Table 1 and Table 2 show, Mdl1 always outperforms
the un-contextual model.
Finally, we compare the performance of the pre-filtering, post-filtering and con-
textual modeling methods. Table 1 shows that Weight PoF is the worst recom-
mender system (it never outperforms any other RS) followed by the un-contextual
one (it outperforms only the Weight PoF while it is dominated by all the other
context-aware RSs). On the other side, the Filter PoF is the best recommender
system (it outperforms all the other systems and it is dominated by the Mdl1 only
9
for the DB1) followed by the Mdl1 (it generally outperforms all the other RSs with
the exception of Filter PoF). The EPF always outperforms the un-contextual re-
commender system, while in some cases dominates the other context-aware RSs
and in other cases is dominated by them. Table 2 shows similar results. In sum-
mary, the answer to the question of which contextual approach provides the best
performance depends on the type of the post-filtering method used. In fact, al-
though the Filter PoF does not outperform all the other approaches in every set-
ting, it does in most settings, while the Weight PoF is often outperformed by all
the other context-based approaches. Overall, the Filter PoF is the best approach in
78% of comparisons in terms of F-measure (i.e., seven times out of 9), 67% in
terms of Precision, 100% in terms of MAE and RMSE (i.e., three times out of
three). In terms of F-measure, the CM approach (Mdl1) is the best in 44% of com-
parisons (percentage values do not sum to 100% because in some cases approach-
es are equivalent), while EPF only 11%. The alternative post-filtering method,
Weight PoF, has never the highest performance in terms of F-measure. In contrast,
it has the lowest F-measure and Precision in 78% of cases, the lowest Recall in
33% of comparisons, and it is always the worst in terms of MAE and Recall.
Although we demonstrated that the post-filtering method Filter PoF dominates al-
ternative contextual approaches considered in the paper in terms of better recom-
mendation performance, it does not mean that it should always be used in practice
for the following reasons. First, using Filter PoF entails finding the right parame-
ters for the method, such as the size of the neighborhood N and the threshold val-
ue. Determining good values of these parameters in a specific application can be
time consuming and expensive. Also, different post-filtering methods, such as
Weight vs. Filter, have different types of parameters, and this complicates the se-
lection of the best post-filtering method even further.
Table 1 F-measure gains (reductions) across recommender systems.
Un-contextual EPF Filter PoF Weight PoF Mdl1
Un-contextual 0% -2% -20% 27% -22%
EPF 2% 0% -19% 29% -21%
DB1 Filter PoF 20% 19% 0% 59% -3%
Weight PoF -27% -29% -59% 0% -39%
Mdl1 22% 21% 3% 39% 0%
Un-contextual 0% -2% -27% 14% -7%
EPF 2% 0% -26% 16% -5%
DB2 Filter PoF 27% 26% 0% 57% 28%
Weight PoF -14% -16% -57% 0% -18%
Mdl1 7% 5% -28% 18% 0%
Un-contextual 0% -88% -90% 8% -88%
EPF 88% 0% -12% 819% 5%
DB3 Filter PoF 90% 12% 0% 942% 19%
Weight PoF -8% -819% -942% 0% -89%
Mdl1 88% -5% -19% 89% 0%
10
Table 2 MAE gains (reductions) across recommender systems.
Un-contextual EPF Filter PoF Weight PoF Mdl1
Un-contextual 0% -57% -257% -9% -88%
EPF 57% 0% -127% 31% -20%
DB1 Filter PoF 257% 127% 0% 70% 47%
Weight PoF 9% -31% -70% 0% -73%
Mdl1 88% 20% -47% 73% 0%
Un-contextual 0% -66% -214% -14% -76%
EPF 66% 0% -89% 31% -6%
DB2 Filter PoF 214% 89% 0% 64% 44%
Weight PoF 14% -31% -64% 0% -55%
Mdl1 76% 6% -44% 55% 0%
Un-contextual 0% -205% -250% 0% -233%
EPF 205% 0% -15% 67% -9%
DB3 Filter PoF 250% 15% 0% 71% 5%
Weight PoF 0% -67% -71% 0% -233%
Mdl1 233% 9% -5% 233% 0%
Finally, post-filtering methods are computationally more expensive than the pre-
filtering method EPF because only a relatively small subset of data is used in es-
timations of unknown ratings for EPF. Therefore, the selection of the best contex-
tual modeling method in a given application is more complicated in practice than a
simple rule “always use Filter PoF in all the settings.” The CM approach is a good
alternative in practice, because it proved to be the second-best and the perfor-
mance is stable across various possible CM methods unlike the post-filtering ap-
proach.
5 Conclusions
In this paper, we compared the un-contextual and contextual recommender sys-
tems, for which we considered the pre-filtering, the post-filtering and the contex-
tual modeling methods of generating contextual recommendations. In particular,
we used the exact pre-filtering (EPF) and the Weight and the Filter post-filtering
methods for the first two approaches. Moreover, we proposed a new type of con-
textual modeling, that we called contextual neighbors CM, and four specific types
of contextual neighbors methods, called Mdl1, Mdl2, Mdl3 and Mdl4, each of them
selecting contextual neighborhoods in a somewhat different way.
We compared the un-contextual with the contextual methods across various expe-
rimental settings, including three datasets, different level of item aggregation, dif-
ferent neighborhood sizes, seven recommendation engines (un-contextual, EPF,
Filter PoF, Weight PoF, Mdl1, Mdl2, Mdl3 and Mdl4), different contextual levels
11
(C1 and C2) and several performance measures (Precision, Recall, F-Measure,
MAE and RMSE). We showed that the contextual Filter method dominates the un-
contextual one and that the un-contextual method dominates the Weight method.
We also showed that EPF dominates the un-contextual method in some cases and
is inferior in other cases on the datasets (and the corresponding applications) used
in our study. We also compared the contextual neighbors methods to identify the
best performing one. Although Mdl1 slightly outperforms the others, there are no
significant performance differences among them. This result is not surprising be-
cause different ways of selecting contextual neighborhood do not fundamentally
change recommendation results. We have then selected Mdl1 as the best-of-breed
contextual modeling method and compared it with the pre-, post- filtering and un-
contextual methods. We showed that Mdl1 dominates the traditional un-contextual
approach and is comparable to the pre-filtering method (EPF). We have also
shown that Mdl1 dominates some of the less advanced post-filtering methods (such
as Weight PoF) but is inferior to the best post-filtering methods (such as Filter
PoF). This implies, among other things, that, in order to decide which approach
should be used in a particular recommendation application, various post-filtering
methods should be carefully compared which is a laborious and a time-consuming
strategy. However, the CM approach is a good alternative in practice, because it
proved to be the second-best and the performance is stable across various possible
CM methods unlike the post-filtering approach.
As future work we intend to analyze deeper different aspects of the comparisons
between pre-filtering, post-filtering and contextual modeling approaches. In par-
ticular, we plan to analyze the impact of different approaches in terms of generat-
ed profit and sales, diversity of the recommended items and trust of the users on
the different recommender systems. Moreover, we will compare the contextual
neighbor approach to other contextual modeling approaches.
References
1. Bettman, J., Luce, M., & Payne, J. (1998). Constructive Consumer Choice Processes. Journal of
Consumer Research, 25(3), 187-217.
2. Palmisano, C., Tuzhilin, A., & Gorgoglione, M. (2008). Using Context to Improve Predictive Mod-
eling of Customers in Personalization Applications. IEEE Transactions on Knowledge and
Data Engineering, 20(11), 1535-1549.
3. Cena, F., Console, L., Gena, C., Goy, A., Levi, G., Modeo, S., et al. (2006). Integrating heterogene-
ous adaptation techniques to build a flexible and usable mobile tourist guide. AI Commun.,
19(4), 369-384.
4. van Setten, M., Pokraev, S., & Koolwaaij, J. (2004). Context-Aware Recommendations in the Mo-
bile Tourist Application COMPASS. In Adaptive Hypermedia and Adaptive Web-Based
Systems (pp. 235-244).
12
5. Adomavicius, G., Sankaranarayanan, R., Sen, S., & Tuzhilin, A. (2005). Incorporating Contextual
Information in Recommender Systems Using a Multidimensional Approach. ACM Transac-
tions on Information Systems, 23, 103-145.
6. Adomavicius, G., & Tuzhilin, A. (2001). Multidimensional Recommender Systems: A Data Ware-
housing Approach. In Electronic Commerce (pp. 180-192).
7. Anand, S., Mobasher, B., Berendt, B., Hotho, A., Mladenic, D., & Semeraro, G. (2007). Contextual
Recommendation. In From Web to Social Web: Discovering and Deploying User and Con-
tent Profiles (Vol. 4737, pp. 142-160): Springer Berlin / Heidelberg.
8. Yu, Z., Zhou, X., Zhang, D., Chin, C.-Y., Wang, X., & Men, J. (2006). Supporting Context-Aware
Media Recommendations for Smart Phones. Pervasive Computing, IEEE, 5(3), 68-75.
9. Herlocker, J., & Konstan, J. (2001). Content-Independent Task-Focused Recommendation. IEEE In-
ternet Computing, 5(6), 40-47.
10. Adomavicius, G., & Tuzhilin, A. (2010). Context-Aware Recommender Systems. In Springer
(Ed.), Handbook on Recommender Systems.
11. Chen, L. S., Hsu, F. H., Chen, M. C., & Hsu, Y. C. (2008). Developing recommender systems with
the consideration of product profitability for sellers. Information Sciences, 178(4), 1032-
1048.
12. Huang, Z., Chung, W., & Chen, H. A graph model for e-commerce recommender systems. In
Journal of the American Society for Information Science and Technology, 2004 (Vol. 55,
pp. 259-274, Vol. 3)
13. Huang, Z., Li, X., & Chen, H. Link prediction approach to collaborative filtering. In JCDL '05:
Proceedings of the 5th ACM/IEEE-CS joint conference on Digital libraries, 2005 (pp. 141-
142): ACM Press
14. Kwon, O., & Kim, J. (2009). Concept lattices for visualizing and generating user profiles for con-
text-aware service recommendations. Expert Systems with Applications, 36(2), 1893-1902.
15. Panniello, U., Tuzhilin, A., Gorgoglione, M., Palmisano, C., & Pedone, A. Experimental compari-
son of pre- vs. post-filtering approaches in context-aware recommender systems. In RecSys
'09: Proceedings of the third ACM conference on Recommender systems, New York, New
York, USA, 2009 (pp. 265-268): ACM
16. Resnick, P., & Varian, H. R. (1997). Recommender systems. Commun. ACM, 40(3), 56-58,
doi:10.1145/245108.245121.
17. Breese, J., Heckerman, D., & Kadie, C. Empirical analysis of predictive algorithms for collabora-
tive filtering. In Proceedings of the 14th conference on Uncertainty in Artificial Intelli-
gence 1998 (Vol. 461, pp. 43-52)
18. Herlocker, J. L., Konstan, J. A., Terveen, L. G., & Riedl, J. T. (2004). Evaluating collaborative fil-
tering recommender systems. ACM Transactions on Information Systems, 22(1), 5-53,
doi:10.1145/963770.963772.
19. Nichols, D. M. Implicit rating and filtering. In Proceedings of the Fifth DELOS Workshop on Fil-
tering and Collaborative Filtering, 1998 (pp. 31-36)
20. Oard, D., & Kim, J. Implicit Feedback for Recommender Systems In Proceedings of the AAAI
Workshop on Recommender Systems, 1998 (pp. 81-83)
21. Adomavicius, G., Huang, Z., & Tuzhilin, A. (2008). Personalization and Recommender Systems.
In Z.-L. C. a. S. Raghavan (Ed.), State-of-the-Art Decision Making Tools in the Informa-
tion-Intensive Age (pp. 55-100): Tutorials in Operations Research.