=Paper=
{{Paper
|id=None
|storemode=property
|title=The Role of Emotions in Context-aware Recommendation
|pdfUrl=https://ceur-ws.org/Vol-1050/paper4.pdf
|volume=Vol-1050
|dblpUrl=https://dblp.org/rec/conf/recsys/ZhengMB13
}}
==The Role of Emotions in Context-aware Recommendation==
https://ceur-ws.org/Vol-1050/paper4.pdf
The Role of Emotions in Context-aware Recommendation
Yong Zheng, Robin Burke, Bamshad Mobasher
Center for Web Intelligence
School of Computing, DePaul University
Chicago, Illinois, USA
{yzheng8, rburke, mobasher}@cs.depaul.edu
ABSTRACT recommender system domain are the ones relevant to users’ subject
Context-aware recommender systems try to adapt to users’ pref- moods or feelings, for example, user mood when listening to music
erences across different contexts and have been proven to provide tracks (e.g. happy, sad, aggressive, relaxed, etc) [19, 9], or users’
better predictive performance in a number of domains. Emotion is emotions after seeing a movie (e.g. user may feel sad after seeing
one of the most popular contextual variables, but few researcher- a tragic movie) [13].
s have explored how emotions take effect in recommendations – Emotional context was first exploited by Gonzalez et al [8] for
especially the usage of the emotional variables other than the ef- the recommender system domain. This work was followed by oth-
fectiveness alone. In this paper, we explore the role of emotions in ers [9, 18, 13, 17] considering emotions as contexts in CARS re-
context-aware recommendation algorithms. More specifically, we search. While the effectiveness of emotions as contextual variables
evaluate two types of popular context-aware recommendation algo- is therefore well-established, there is little research that has exam-
rithms – context-aware splitting approaches and differential context ined specifically the role that these emotional variables play. We
modeling. We examine predictive performance, and also explore define "the role of emotions" as the concerns from two aspects –
the usage of emotions to discover how emotional features interact whether emotions are useful or effective to improve recommenda-
with those context-aware recommendation algorithms in the rec- tion performance? And, the usage of emotions in the recommen-
ommendation process. dation process – how emotions are used in the recommendation
algorithms, e.g. which emotional variables are selected? which
algorithm components are they applied to? and so forth. Current-
Categories and Subject Descriptors ly, most research are focused on demonstrating the effectiveness of
H.3.3 [Information Search and Retrieval]: Information filtering emotional variables, and few research further explore the usage of
the emotions in the recommendation process.
General Terms In this paper, we explore the role of emotions by two classes
of popular context-aware recommendation algorithms – context-
Algorithm, Experiment, Performance
aware splitting and differential context modeling. Since both of
these approaches require that the algorithm learn the importance
Keywords and utility of different contextual features, they help reveal how
Recommendation, Context, Context-aware recommendation, Emo- emotions work in each algorithm and what roles they can play in
tion, Affective recommender system recommendation.
The purpose of this study is therefore to address the following
1. INTRODUCTION research questions:
Advances in affective computing have enabled recommender sys- 1. Emotional Effect: Are emotions useful contextual variables
tems to take advantage of emotions and personality, leading to the for context-aware recommendation?
development of affective recommender systems (ARS) [18]. At
the same time, the emerging technique of context-aware recom- 2. Algorithm Comparison: What algorithms are best suited to
mender systems (CARS) takes contexts into consideration, convert- make use of emotional variables? Do they outperform the
ing a two-dimensional matrix of ratings organized by user and item: baseline algorithms?
Users × Items → Ratings, into a multidimensional rating space [1]. 3. Usage of Contexts: How do emotional variables compete
CARS have been demonstrated to be effective in a variety of appli- with other contextual variables, such as location and time?
cations and domains [4, 15, 10, 24, 13]. Emotional variables are
often included as contexts in CARS, which enables the further de- 4. Roles: How can we understand the specific roles emotional
velopment of both ARS and CARS. Typical emotional contexts in variables can play in those context-aware recommendation
algorithms?
2. RELATED WORK
RecSys’13, October 12–16, 2013, Hong Kong, China. Gonzalez et al [8] explored emotional context in recommender sys-
Paper presented at the 2013 Decisions@RecSys workshop in conjunc- tems in 2007. They pointed out that,"emotions are crucial for us-
tion with the 7th ACM conference on Recommender Systems. Copyright er’s decision making in recommendation process. The users always
c 2013 for the individual papers by the papers’ authors. Copying permitted
for private and academic purposes. This volume is published and copyright- transmit their decisions together with emotions." With the rapid de-
ed by its editors. velopment of context-aware recommender systems, emotion turns
21
out to be one of important and popular contexts in different kinds (the first two rows) in the training data and one unknown rating that
of domains, especially in the music and movie domain. For mu- we are trying to predict (the third row). There are three contextual
sic recommendation, the emotional context is appealing because it dimensions – time (weekend or weekday), location (at home or cin-
can be used to establish a bridge between music items and item- ema) and companion (friend, girlfriend or family). In the following
s from other different domains, and perform cross-media recom- discussion, we use contextual dimension to denote the contextu-
mendations [6, 1, 9]. Movie recommendation is another domain al variable, e.g. "Location" in this example. The term contextual
where emotion turns out to be popular in recent years. In 2010, condition refers to a specific value in a contextual dimension, e.g.
Yue et al [16] produced the overall winner in a recent challenge "home" and "cinema" are two contextual conditions for "Location".
on context-aware movie recommendation by mining mood-specific
movie similarity with matrix factorization. More recent research 3.1 Item Splitting
[18, 13] motivates the tendency of taking emotions as contexts Item splitting tries to find a contextual condition on which to split
to assist contextual recommendations. Research has demonstrat- each item. The split should be performed once the algorithm iden-
ed that emotions can be influential contextual variables in making tifies a contextual condition in which items are rated significantly
recommendations, but few of them explore how emotions interact differently. In the movie example above, there are three contextu-
with recommendation algorithm – the usage of emotional variables al conditions in the dimension companion: friend, girlfriend and
in the recommendation process. family. Correspondingly, there are three possible alternative con-
As described in [1], there are three basic approaches to devel- ditions: "friend and not friend", "girlfriend and not girlfriend",
op context-aware recommendation algorithms: pre-filtering, post- "family and not family". Impurity criteria [4] are used to deter-
filtering, and contextual modeling. A pre-filtering approach applies mine whether and how much items were rated differently in these
a context-dependent criterion to the list of items, selecting those alternative conditions, for example a t-test or other statistical met-
appropriate to a given context. Only the filtered items are con- ric can be used to evaluate if the means differ significantly across
sidered for recommendation. A post-filtering approach is similar conditions.
but applies the filter after recommendations have been computed. Item splitting iterates over all contextual conditions in each con-
Contextual modeling takes contextual considerations into the rec- text dimension and evaluates the splits based on the impurity cri-
ommendation algorithm itself. In this paper, we explore context- teria. It finds the best split for each item in the rating matrix and
aware splitting approaches (a class of pre-filtering algorithms), and then items are split into two new ones, where contexts are elimi-
differential context modeling (a contextual modeling approach.) nated from the original matrix – it transforms the original multi-
The remainder of this paper is organized as follows. In Sec- dimensional rating matrix to a 2D matrix as a result. Assume that
tion 3 and Section 4, we formally introduce the two types of recom- the best contextual condition to split item T1 in Table 1 is "Loca-
mendation algorithms we are studying, including their capability to tion = home and not home", T1 can be split into T11 (movie T1
capture the role of emotional contexts in recommendation process. being seen at home) and T12 (movie T1 being seen not at home).
Sections 5 and 6 discuss the experimental evaluations and mining Once the best split has been identified, the rating matrix can be
the role of emotions through those context-aware recommendation transformed as shown by Table 2(a).
algorithms, followed by the conclusions and future work in Sec- This example shows a simple split, in which a single contextual
tion 7. condition is used to split the item. It is also possible to perform a
complex split using multiple conditions across multiple context di-
3. CONTEXT-AWARE SPLITTING mensions. However, as discussed in [5], there are significant costs
of sparsity and potential overfitting when using multiple conditions.
APPROACHES We use only simple splitting in this work.
Contextual pre-filtering is popular as it is straightforward to imple-
ment, and can be applied with most recommendation techniques. Table 2: Transformed Rating Matrix
Item splitting [4, 5] is considered one of the most efficient pre-
filtering algorithms and it has been well developed in recent re- (a) by Item Splitting (b) by User Splitting
search. The underlying idea of item splitting is that the nature of
an item, from the user’s point of view, may change in different User Item Rating User Item Rating
contextual conditions, hence it may be useful to consider it as t- U1 T11 3 U12 T1 3
wo different items [5]. User splitting [2, 15] is based on a similar U1 T12 5 U12 T1 5
intuition – it may be useful to consider one user as two different U1 T11 ? U11 T1 ?
users, if he or she demonstrates significantly different preferences (c) by UI Splitting
across contexts. In this section, we introduce those two approaches
User Item Rating
and also propose a new splitting approach – UI splitting, a simple
combination of item and user splitting. U12 T11 3
To better understand and represent the splitting approaches, con- U12 T12 5
sider the following movie recommendation example: U11 T11 ?
Table 1: Movie Ratings in Contexts
3.2 User Splitting
User Item Rating Time Location Companion
Similarly, user splitting tries to split users instead of items. It can
U1 T1 3 Weekend Home Friend be easily derived from item splitting as introduced above. Similar
U1 T1 5 Weekend Cinema Girlfriend impurity criteria can also be used for user splits. Assume that the
U1 T1 ? Weekday Home Family best split for user U 1 in Table 1 is "Companion = family and not
family", U 1 can be split into U 11 (U 1 saw the movie with family)
In Table 1, there are one user U 1, one item T 1 and two ratings and U 12 (U 1 saw the movie with others). The rating matrix can
22
be transformed as shown by Table 2(b). The first two rows contain UBCF, where the four components we separate in UBCF can be
the same user U 12 because U 1 saw this movie with others (i.e. not described as follows:
family) rather than family as shown in the original rating matrix.
3.3 UI Splitting
UI splitting is a new approach proposed in this paper – it applies
item splitting and user splitting together. Assuming that the best
split for item and user splitting are the same as described above,
the rating matrix based on UI splitting can be shown as Table 2(c).
Here we see that both users and items were transformed, creating
new users and new items.
Thus, given n items, m users, k contextual dimensions and d dis-
tinct conditions for each dimension, the time complexity for those
three splitting approaches is the same as O(nmkd) [5]. The pro- Figure 1: Algorithm Components in UBCF
cess in UI splitting is simply an application of item splitting fol-
lowed by user splitting on the resulting output. We keep the con-
textual information in the matrix after the transformation by item Figure 1 shows the original rating prediction by the well-known
splitting so that user splitting can be performed afterwards. Resnick’s UBCF algorithm [14], where a is a user, i is an item,
and N is a neighborhood of K users similar to a. The algorithm
3.4 Role of Emotions in Splitting calculates Pa,i , which is the predicted rating that user a is expected
Context-aware splitting approaches have been demonstrated to im- to assign to item i. Then we decompose it to four components:
prove predictive performance in previous research, but few of those
research results report on the details of the splitting process. More Neighborhood selection UBCF applies the well-known k nearest-
specifically, it is useful to know which contextual dimensions and neighbor (kNN) approach, where we can select the top-k
conditions are selected, and the statistics related to those selection- neighbors from users who have rated on the same item i. If
s. For our purposes, it is possible to explore how emotions interact contexts are taken into consideration, the neighborhood can
with the splitting process by the usage of contexts in the splitting be further restricted so that users have also rated the item i
process, which may help discover the role of emotions from this in the same contexts. This gives a context-specific recom-
perspective. In this paper, we try all three context-aware splitting mendation computation. However, the strict application of
approaches, empirically compare their predictive performance, and such a filter greatly increases the sparsity associated with us-
also explore the distributions of the usage of contexts (emotional er comparisons. There may only be a small number of cases
variables included) for splitting operations. in which recommendations can be made. DCM offers two
potential solutions to this problem. DCR searches for the
optimal relaxation of the context, generalizing the set of con-
4. DIFFERENTIAL CONTEXT MODELING textual features and contextual conditions to reduce sparsity.
Differential context modeling (DCM) is a general contextual rec- DCW uses weighting to increase the influence of neighbors
ommendation framework proposed in [23]. It can be applied to in similar contexts.
any recommendation algorithm. The "differential" part of the tech-
nique tries to break down a recommendation algorithm into differ- Neighbor contribution The neighbor contribution is the difference
ent functional components to which contextual constraints can be between a neighbor’s rating on an item i and his or her av-
applied. The contextual effect for each component is maximized, erage rating over all items. Context relaxation and contex-
and the joint effects of all components contribute the best perfor- t weighting can be applied to the computation of r̄u . This
mance for the whole algorithm. The "modeling" part is focused on computation is replaced by one in which this average is com-
how to model the contextual constraints. There are two approach- puted over ratings from a relaxed set of contexts in DCR, or
es: context relaxation and context weighting, where context relax- the average rating can be aggregated by a weighted average
ation uses an optimal subset of contextual dimensions, and context across similar contexts under DCW.
weighting assigns different weights to each contextual factor. Ac-
cordingly, we have two approaches in DCM: differential context User baseline The computation of r̄a is similar to the neighbor’s
relaxation (DCR) [21, 22, 20] and differential context weighting average rating and can be made context-dependent in the
(DCW) [24]. same way.
4.1 DCM in UBCF User similarity The computation of neighbor similarity sim(a, u)
We have successfully applied DCM to user-based collaborative fil- involves identifying ratings ru,i and ra,i where the users
tering (UBCF), item-based collaborative filtering (IBCF) and Slope- have rated items in common. For context-aware recommen-
One recommendation algorithms in our previous research [23], show- dation, we can additionally add contextual constraints to this
ing that DCM is an efficient approach to improve context-aware part: use ratings given in matching contexts with context re-
prediction accuracy. However, we did not explore much about the laxation or ratings weighted by contextual similarity using
modeling part in our previous work; that is, how contexts are se- context weighting.
lected, relaxed or weighted.
Recall that we separate different functional components from the With these considerations in mind, we can derive a new rating
recommendation algorithm, therefore, how the algorithms relax or prediction formula by applying DCR or DCW to UBCF. More de-
weigh contextual variables can tell the role of the contexts in diff- tails about the prediction equations and technical specifications can
erent components. In this paper, we continue to apply DCM to be found in [24].
23
4.2 The Role of Emotions in DCM Usually, a threshold for the splitting criteria should be set so that
One advantage of DCM is that it allows us to explore the role of users or items are only be split when the criteria meets the signifi-
contexts in each algorithm component. DCM seeks to optimize the cance requirement. We use an arbitrary value of 0.2 in the tIG case.
contribution of context in each component, and so, the output of For tmean , tchi and tprop , we use 0.05 as the p-value threshold. A
the optimization phase, in which contextual dimensions are select- finer-grained operation is to set another threshold for each impurity
ed and/or weighted, can reveal the relative importance of different value and each data set. We deem it as a significant split once the
contextual features in different algorithm components. In this pa- p-value is no larger than 0.05. We rank all significant splits by the
per, we provide comparisons and also explore the role of emotions impurity value, and we choose the top first (highest impurity) as the
in these two classes of context-aware recommendation algorithms. best split. Items or users without qualified splitting criteria are left
unchanged.
In DCM, we used the same configuration in our previous work
5. EXPERIMENTAL SETUP [24]: we select the top-10 neighbors in UBCF, choose 100 as the
In this section, we introduce the data sets, evaluation protocols and maximal iteration in the optimization process, and Pearson Corre-
the specific configurations in our experiments. lation is used as the user similarity measure. See our previous work
for more details.
5.1 Data Sets
We examine context-aware splitting approaches and DCM with the 5.3 Evaluation Protocols
real-world data set LDOS-CoMoDa, which is a movie dataset col- For the evaluation of predictive performance, we choose the root
lected from surveys and used by Odic et al [12, 13, 17]. After fil- mean square error (RMSE) evaluated using the 5-fold cross vali-
tering out subjects with less than 5 ratings and rating records with dation. The data set is relatively small and some subjects in the
incomplete feature information, we got the final data sets contain- survey were required to rate specific movies, thus precision and re-
ing 113 users, 1186 items, 2094 ratings, 12 contextual dimensions call may be not a good metric, but we plan to evaluate them and
and the rating scale is 1 to 5. We created five folds for cross val- other metrics in our future work.
idation purposes and all algorithms are evaluated on the same five We applied DCM only to user-based collaborative filtering. For
folds. Specific descriptions of all the contextual dimensions and splitting approaches, there are more options – we evaluate the per-
conditions can be found in previous work [12] and the description formance of three recommendation algorithms: user-based collab-
of the data set online 1 . orative filtering (UBCF), item-based collaborative filtering (IBCF)
In our experiments, we use all 12 contextual dimensions, includ- and matrix factorization (MF) for each splitting approach. Ko-
ing three emotional contexts: Mood, DominantEmo and EndEmo, ren [11] introduced three MF techniques: MF with rating bias (Bi-
and non-emotional contexts, including Location, time, etc. The asMF), Asymmetric SVD (AsySVD) and SVD++; we found that
goal is to compare emotional contexts with others. For example, BiasMF was the best choice for this data.
in DCR, only influential contextual variables for a specific compo- We used the open source recommendation engine MyMediaLite
nent are selected – whether emotional contexts are selected or not v3.07 [7] to evaluate UBCF, IBCF and BiasMF in our experiments.
can indicate the significance of their impacts for this component. (We choose K=30 for KNN-based UBCF and IBCF.) In order to
For a comprehensive comparison, we also performed experiments better evaluate MF techniques, we tried a range of different fac-
without emotional contexts to better evaluate the importance of the tors (5 ≤ N ≤ 60, increment 5) and training iteration T (10 ≤
emotional dimensions. T ≤ 100, increment 10). Other parameters like learning and reg-
In this data, there are the three contextual dimensions that con- ularization factors are handled by MyMediaLite, where stochastic
tain emotional information. "EndEmo" is the emotional state expe- gradient descent is used as the optimization method.
rienced at the end of the movie. "DominantEmo" is the emotion- For comparison purposes, we choose the well-known context-
al state experienced the most during watching, i.e. what emotion aware matrix factorization algorithm (CAMF) [3], i.e. CAMF_C,
was induced most times during watching. "Mood" is what mood CAMF_CI and CAMF_CU 2 as the baseline. We tried our best
the user was in during that part of the day when the user watched to fine-tune the configurations (e.g. learning parameters, training
the movie. Mood has lower maximum frequency than emotions, iterations, etc) for CAMF in order to make a reasonable and com-
it changes slowly, so we assumed that it does not change during prehensive comparison.
watching. "EndEmo" and "DominantEmo" contain the same sev-
en conditions: Sad, Happy, Scared, Surprised, Angry, Disgusted,
Neutral, where "Mood" only has simple three conditions: Positive, 6. EXPERIMENTAL RESULTS
Neutral, Negative. In this section, the comparisons of predictive performance are in-
troduced first, followed by the discussion of emotional roles dis-
5.2 Configurations covered in our experiments.
To evaluate the performance of context-aware splitting approach-
es, we used four splitting criteria described in [5]: tmean , tchi , 6.1 Prediction
tprop and tIG . tmean estimates the statistical significance of the We use three treatments of the context information – one is the
difference in the means of ratings associated to each alternative data with All Contexts where both emotional contexts and non-
contextual condition using a t-test. tchi and tprop estimates the emotional variables are included, and the second one is the data
statistical significance of the difference between two proportions – omitting the emotional context dimensions marked by No Emo-
high ratings (>R) and low ratings (≤R) by chi square test and z- tions in the table below. The third one is the data with Emotions
test respectively, where we choose R = 3 as in [5]. tIG measures Only emitting all non-emotional variables. The overall experimen-
the information gain given by a split to the knowledge of the item i tal results are shown in Table 3, where the numbers in underlined
rating classes which are the same two proportions as above. in italic are the best RMSEs by each approach (i.e. the best one in
1 2
http://212.235.187.145/spletnastran/raziskave/um/comoda/ CAMF_CU is a CAMF approach which utilizes the interaction
comoda.php between contexts and users. [12]
24
Table 3: Overall Comparison of RMSE
(The results for splitting approaches are based on BiasMF.)
Algorithms All Contexts No Emotions Emotions Only
CAMF_C 1.012 1.066 0.968
CAMF [3]
CAMF_CI 1.032 1.083 1.019
CAMF_CU 0.932 1.021 0.902
DCR 1.043 1.057 1.046
DCM [24]
DCW 1.017 1.037 1.036
Item Splitting 1.011 1.014 1.014
Splitting
User Splitting 0.913 0.971 0.932
Approaches
UI Splitting 0.892 0.942 0.903
each row), and the bold numbers are the best RMSEs by each data Table 4: Comparison of Predictive Performance (RMSE) Among
forms (i.e. the best one in each column). Splitting Approaches
The comparison of performances of those approaches over the
three forms of data can be visualized by Figure 2. We see that in- LDOS-CoMoDa
cluding emotions as contexts can improve RMSE compared with Algorithms tmean tchi tprop tIG
the situation we only use non-emotional contexts (i.e. No Emotion- UBCF 1.040 1.021 1.028 1.043
Item
s). The results differ when we switch our attention to the "Emo- IBCF 1.030 1.024 1.026 1.034
Splitting
tions Only" one. In CAMF, it helps achieve the lowest RMSE with BiasMF 1.020 1.011 1.016 1.020
Emotions Only data. But, including all of the contextual informa- UBCF 1.026 0.987 0.999 1.052
User
tion yielded the best RMSE for the other approaches: DCM and IBCF 0.985 0.967 0.985 1.039
Splitting
context-aware splitting algorithms. BiasMF 0.934 0.913 0.928 1.011
UBCF 1.012 0.956 0.989 1.042
1.11 UI
IBCF 0.972 0.946 0.972 1.020
1.08 Splitting
1.05
BiasMF 0.927 0.892 0.915 0.998
1.02
0.99
0.96 tion algorithm with tchi as the splitting criteria.
0.93
0.9
LDOS−CoMoDa
0.87
CAMF_C CAMF_CI CAMF_CU DCR DCW Item User UI Splitting
Splitting Splitting
1.02
All Contexts No Emotions Emotions Only
1.00
Figure 2: Comparison of RMSE Over Three Contextual Situations
0.98
This result is not surprising because both DCM and splitting ap-
proaches choose among contextual features, deciding which to ap-
RMSE
0.96
ply in recommendation. CAMF, on the other hand, uses all avail-
able context information in performing its factorization. Without
0.94
the benefit of feature selection, adding additional features to CAM-
F may increase noise.
0.92
From an overall view, UI splitting has the best RMSE across all
data treatments among those context-aware recommendation algo-
0.90
rithms. Table 4 shows more details of the predictive performance
among the three splitting approaches if we include emotions as the Item Splitting User Splitting UI Splitting
contexts (i.e. using all contexts). The numbers are shown as RMSE
values, where the numbers in bold are the best performing RMSE
values for each recommendation algorithm across all three splitting
approaches. The numbers in underlined in italic are the best RMSE Figure 3: Boxplot of RMSE Among Splitting Approaches
values achieved for the data set using each splitting approach. The
numbers in underlined in bold are the global best RMSE for the Generally, MF is the best performing recommendation algorith-
data set. m. This is not surprising because splitting increases sparsity and
These tables show that adding emotions to contexts is able to MF approaches are designed to handle sparsity data. Because the
provide improvement in terms of RMSE, thus answering research difference in RMSE is small, we show the boxplot of RMSEs a-
question 1 in the affirmative. It also suggests an answer to question mong the best performing item splitting, user splitting and UI s-
2: that the UI splitting approach outperforms other two splitting ap- plitting approaches (i.e. the configuration as the underlined values
proaches if it is configured optimally. In particular, the best RMSE in Table 4) in Figure 3. The data in the figure are the 120 RMSE
values are achieved by UI splitting using MF as the recommenda- values which comes from the training iterations – different factors
25
30%
28% 36%
Time
25% 33%
DayType 30%
23%
Season 27%
20%
24% Time
18% Location 21% Season
15%
Weather 18%
13% EndEmo
15%
10% Social DominantEmo
12%
8% EndEmo 9% Mood
5% 6%
DominantEmo
3% 3%
0% Interaction
0%
Tmean Tchi Tprop Tmean Tchi Tprop
Figure 4: Item Splitting Figure 5: User Splitting
Table 5: Context Relaxation and Weighting by DCM
Algorithm Components Context Relaxation By DCR Context Weighting By DCW
Neighbor Selection N/A Day, Mood
Neighbor Contribution Movie Year, Genre Movie Genre
User Baseline DominantEmo, EndEmo, Movie Language DominantEmo, EndEmo, Interaction
User Similarity EndEmo, Location DominantEmo
(5 ≤ N ≤ 60, with 5 increment in each step) and training iter- EndEmo denotes the emotion of the users after seeing the movie,
ation T (10 ≤ T ≤ 100, with 10 increment in each step). The and this result is consistent with previous work [12] on this da-
figure confirms the comparative effectiveness of UI splitting – the ta. Obviously, emotion is a personal quality and can be considered
box is significantly lower than the other two approaches with the as more dependent with users other than items – we conjecture that
same training iterations. this may be the underlying clue explaining why user splitting works
better than item splitting for the LDOS-CoMoDa data. And in user
6.2 The Role of Emotions splitting, the top two selected contextual dimensions are the two
As mentioned before, one reason we sought to explore the usage emotional variables: EndEmo and DominantEmo – emotions are
of emotional contexts to discover how emotions interact within generated and owned by users, therefore they are more dependent
context-aware recommendation algorithms. Both splitting and D- with users other than items. This pattern is confirmed by the com-
CM offer insights into how contextual dimensions and features are parison between two CAMF approaches: CAMF_CI and CAM-
influential for recommendation. F_CU – contexts are more dependent with users than with items,
which results in better performances by CAMF_CU.
6.2.1 Context-aware Splitting In short, the statistics based on splitting approaches reveal the
In splitting approaches, the splitting statistics help discover how importance of emotions – at least the top first selected contextual
contexts were applied in our experiments. In Figure 4 and 5, we dimension is the "EndEmo" for both item splitting and user split-
show the top selected contextual dimensions for item splitting and ting.
user splitting3 . The right legend indicates the contextual dimen-
sions, and the y axis denotes the percentage of splits (item splits 6.2.2 Differential Context Modeling
or user splits) using each dimension. For a clearer representation in In DCM, the context selection and context weighting can be ex-
the figures, we only show contextual dimensions used on more than amined and show, in a detailed way, the contribution of each con-
5% (item splitting) or 6% (user splitting) of the recommendations. textual dimension in the final optimized algorithm. The results are
We do not show results of tIG because this splitting criterion had shown in Table 5, where the four components in UBCF were de-
the worst performance. scribed in the previous section. "N/A" in the table indicates no
In general, the top two dimensions are consistent across those contextual constrains were placed. For clearer representation, we
three impurity criteria: EndEmo and Time for item splitting and did not show specific weights of contexts in DCW; instead, we on-
EndEmo and DominantEmo for user splitting. However, the per- ly list variables which were assigned weights above a threshold of
centages as y axis in the figures are different, not to mention that 0.7. The weights are normalized to 1, and 0.7 therefore represents
the selected condition in each dimension differs too, which result- a very influential dimension.
s in different performance by using various impurity criteria. In We can see in the table that emotional variables are selected in
terms of the specific selected emotional conditions, the results are DCR and weighted significantly in DCW and for which compo-
not consistent – the top context dimension for item and user split- nents. Emotion is influential for a specific component but may not
ting is the same – EndEmo, but the most frequent selected condi- for other ones. In DCR for example, EndEmo turns out to be influ-
tion in this dimension is "Happy" for item splitting and "Neutral" ential when measuring user similarities and user baselines, but it is
for user splitting. not that significant in computing the neighbor contribution.
3
The actual splitting is based on a specific contextual condition It is not surprising that the results from DCW are not fully consis-
in a dimension, but the results of selected contextual conditions tent with ones from DCR. DCW is a finer-grained approach. Emo-
are fuzzy, thus the distribution of selected contextual dimensions is tional variables are assigned to neighbor selection in DCW but not
explored and reported here. for the same component in DCR, and the specific selected emotions
26
Table 6: Comparison of Context-aware Splitting and Differential Context Weighting
Answers Context-aware Splitting Approaches Differential Context Weighting
Q1. Emotional Effect Yes. RMSE is improved with emotional context Yes. RMSE is improved with emotional context di-
dimensions. mensions.
Q2. Algorithm Comparison UI splitting is the best and outperforms DCM and DCW is better than DCR and also works better than
CAMF approaches. some CAMF approaches in specific contextual situ-
ations.
Q3. Usage of Contexts Emotional dimension is the top selected dimension Emotional dimensions get more weight than other
for context splitting. contextual variables in most algorithm components.
Q4. Roles EndEmo is used as the top first selected context for Emotions are significantly influential for some algo-
both user and item splitting. DominantEmo is the rithm components, e.g. EndEmo and DominantEmo
top second one in user splitting. for user similarities and baselines in DCW.
are different too, e.g. it is DominantEmo selected in DCW for user question, by pointing out which components make good use of d-
similarity calculation, but it is the EndEmo in DCR. In DCW, the ifferent contextual variables – showing here that DominantEmo is
weights for EndEmo and DominantEmo are close to 1 (the weights important for calculating the user’s baseline. Note that the neigh-
are all above 0.92) in the component of user baseline, which im- bor contribution component in DCW is alone in making significant
plies significant emotional influence on this component. Emotional use of non-emotional context dimensions. This is interesting be-
contexts are important in DCM, where it can be further confirmed cause in this component we are determining which movies to use
by Table 3 – the RMSE values are increased if we remove emotions to compute the neighbor’s baseline for prediction. It appears that
from the contexts. the system prefers to use non-emotional considerations in making
The result by DCM is useful for further applications, such as af- use of others’ ratings, even though the user’s baseline is computed
fective computing or marketing purposes. Take the results of DCW based on emotional dimensions.
in Table 5 for example, Mood is influential for selecting neighbors, In addition, we can get extra information from our splitting ex-
which implies that if two users rated the same item under the same periments because splitting is performed based on specific contex-
mood situation, it is highly possible that they are the good neighbor tual conditions. As described above, we found in particular that
candidates for each other (though neighbor selection also depends "Happy" vs not-"Happy" was the important split on the EndEmo
on the user similarities). Similarly, DominantEmo is useful to mea- dimension for item splitting. Essentially, the algorithm is indicat-
sure user similarities, which infers that users rated items similarly ing that there is an important difference between users who feel
with the same dominant emotions are probably the good neighbors happy at the conclusion of a given movie and those that do not.
in user-based collaborative filtering. In our future work, we plan to continue our exploration of these
algorithms and more data using additional metrics, such as recall
7. CONCLUSION AND FUTURE WORK and/or normalized discounted cumulative gain. We are also look-
ing at additional data sets to see if similar effects are found with
In conclusion, both context-aware splitting approaches and DCM
respect to emotions, as well as the empirical comparison among
are able to reveal how emotions interact with algorithms to im-
those context-aware recommendation algorithms. We also plan to
prove recommendation performance by exploring the usage of con-
examine the effects by the correlations between different contex-
texts in the recommendation or splitting process. More specifically,
tual variables, e.g. how emotional effects change if emotions are
in context-aware splitting approaches, the percentage of emotional
significantly dependent with other contexts or features. In addition,
contexts used by item splits or user splits can tell the importance
it is interesting to explore the association among emotions, user
of emotions in distinguishing different user rating behavior. DCM
profiles, item features and users’ ratings. For example, user may
provides a way to see very specifically which emotional contexts
feel "sad" after seeing a tragedy movie, but the emotion could be
are influential for which components in the recommendation pro-
"happy" because he or she saw such a good movie even if it is a
cess.
tragedy. Therefore, which specific emotions will result in a higher
Table 6 examines how our research questions are answered for
rating? the "sad" or the "happy"?
each type of context-aware recommendation. As discussed above,
we find that contextual dimensions keyed to emotions are very use-
ful in recommendation (Question 1) and that UI splitting outper- 8. REFERENCES
forms the DCM and CAMF approaches (Question 2). [1] G. Adomavicius, B. Mobasher, F. Ricci, and A. Tuzhilin.
Question 3 asks about how emotional dimensions compare with Context-aware recommender systems. AI Magazine,
other contextual information. Our results show that emotion-linked 32(3):67–80, 2011.
context makes an important contribution to context-aware recom- [2] L. Baltrunas and X. Amatriain. Towards time-dependant
mendation, although other dimensions are also certainly important. recommendation based on implicit feedback. In ACM
For example, we see that EndEmo is the top selected contextual RecSys’ 09, Proceedings of the 4th International Workshop
dimension in item and user splitting, with Time running the sec- on Context-Aware Recommender Systems (CARS 2009),
ond in item splitting, and DominantEmo is the top second selected 2009.
contextual dimension in user splitting. [3] L. Baltrunas, B. Ludwig, and F. Ricci. Matrix factorization
Our fourth research question asks about how those two class- techniques for context aware recommendation. In
es of context-aware recommendation algorithms can tell the roles Proceedings of the fifth ACM conference on Recommender
of emotions by usage of contexts. Context-aware approaches are systems, pages 301–304. ACM, 2011.
able to infer the roles by the distribution of usage of emotions se- [4] L. Baltrunas and F. Ricci. Context-based splitting of item
lected for the splitting process. DCM techniques help answer this ratings in collaborative filtering. In Proceedings of the third
27
ACM conference on Recommender systems, pages 245–248. factors affecting music recommender success. In ISMIR,
ACM, 2009. volume 2, pages 204–208, 2002.
[5] L. Baltrunas and F. Ricci. Experimental evaluation of [20] Y. Zheng, R. Burke, and B. Mobasher. Assist your
context-dependent collaborative filtering using item splitting. decision-making in various situations: Differential context
User Modeling and User-Adapted Interaction, pages 1–28, relaxation for context-aware recommendations. In Research
2013. Colloquium, School of Computing, DePaul University,
[6] S. Berkovsky, T. Kuflik, and F. Ricci. Cross-domain Chicago IL, USA, 2012.
mediation in collaborative filtering. In User Modeling 2007, [21] Y. Zheng, R. Burke, and B. Mobasher. Differential context
pages 355–359. Springer, 2007. relaxation for context-aware travel recommendation. In 13th
[7] Z. Gantner, S. Rendle, C. Freudenthaler, and International Conference on Electronic Commerce and Web
L. Schmidt-Thieme. Mymedialite: A free recommender Technologies (EC-WEB 2012), pages 88–99, 2012.
system library. In Proceedings of the fifth ACM conference [22] Y. Zheng, R. Burke, and B. Mobasher. Optimal feature
on Recommender systems, pages 305–308. ACM, 2011. selection for context-aware recommendation using
[8] G. Gonzalez, J. L. de la Rosa, M. Montaner, and S. Delfin. differential relaxation. In ACM RecSys’ 12, Proceedings of
Embedding emotional context in recommender systems. In the 4th International Workshop on Context-Aware
Data Engineering Workshop, 2007 IEEE 23rd International Recommender Systems (CARS 2012). ACM, 2012.
Conference on, pages 845–852. IEEE, 2007. [23] Y. Zheng, R. Burke, and B. Mobasher. Differential context
[9] B.-j. Han, S. Rho, S. Jun, and E. Hwang. Music emotion modeling in collaborative filtering. In School of Computing
classification and context-based music recommendation. Research Symposium. DePaul University, USA, 2013.
Multimedia Tools and Applications, 47(3):433–460, 2010. [24] Y. Zheng, R. Burke, and B. Mobasher. Recommendation
[10] N. Hariri, B. Mobasher, R. Burke, and Y. Zheng. with differential context weighting. In The 21st Conference
Context-aware recommendation based on review mining. In on User Modeling, Adaptation and Personalization (UMAP
IJCAI’ 11, Proceedings of the 9th Workshop on Intelligent 2013), pages 152–164, 2013.
Techniques for Web Personalization and Recommender
Systems (ITWP 2011), pages 30–36, 2011.
[11] Y. Koren. Factorization meets the neighborhood: a
multifaceted collaborative filtering model. In Proceedings of
the 14th ACM SIGKDD international conference on
Knowledge discovery and data mining, pages 426–434.
ACM, 2008.
[12] A. Odic, M. Tkalcic, J. F. Tasic, and A. Košir. Relevant
context in a movie recommender system: Users’ opinion vs.
statistical detection. In ACM RecSys’ 12, Proceedings of the
4th International Workshop on Context-Aware Recommender
Systems (CARS 2012), 2012.
[13] A. Odić, M. Tkalčič, J. F. Tasič, and A. Košir. Predicting and
detecting the relevant contextual information in a
movie-recommender system. Interacting with Computers,
25(1):74–90, 2013.
[14] P. Resnick, N. Iacovou, M. Suchak, P. Bergstrom, and
J. Riedl. Grouplens: an open architecture for collaborative
filtering of netnews. In Proceedings of the 1994 ACM
conference on Computer supported cooperative work, pages
175–186. ACM, 1994.
[15] A. Said, E. W. De Luca, and S. Albayrak. Inferring
contextual user profiles – improving recommender
performance. In ACM RecSys’ 11, Proceedings of the 4th
International Workshop on Context-Aware Recommender
Systems (CARS 2011), 2011.
[16] Y. Shi, M. Larson, and A. Hanjalic. Mining mood-specific
movie similarity with matrix factorization for context-aware
recommendation. In Proceedings of the Workshop on
Context-Aware Movie Recommendation, pages 34–40. ACM,
2010.
[17] M. Tkalčič, U. Burnik, A. Odić, A. Košir, and J. Tasič.
Emotion-aware recommender systems–a framework and a
case study. In ICT Innovations 2012, pages 141–150.
Springer, 2013.
[18] M. Tkalcic, A. Kosir, and J. Tasic. Affective recommender
systems: the role of emotions in recommender systems. In
ACM RecSys Workshop on Human Decision Making, 2011.
[19] A. L. Uitdenbogerd and R. G. van Schyndel. A review of
28