APPLYING THE AFFECTIVE AWARE PSEUDO
ASSOCIATION METHOD TO ENHANCE THE TOP-N
RECOMMENDATIONS DISTRIBUTION TO USERS IN
GROUP EMOTION RECOMMENDER SYSTEMS
John Kalung Leung
1
, Igor Griva
2
and William G. Kennedy
3
1
Computational and Data Sciences Department, Computational Sciences and
Informatics, College of Science, George Mason University, 4400 University Drive,
Fairfax, Virginia 22030, USA
2
Department of Mathematical Sciences, MS3F2, Exploratory Hall 4114, George Mason
University,4400 University Drive, Fairfax, Virginia 22030, USA
3
Center for Social Complexity, Computational and Data Sciences Department, College
of Science, George MasonUniversity, 4400 University Drive, Fairfax, Virginia 22030,
USA
ABSTRACT
Recommender Systems are a subclass of information retrieval systems, or more succinctly, a class of
information filtering systems that seeks to predict how close is the match of the user’s preference to a
recommended item. A common approach for making recommendations for a user group is to extend
Personalized Recommender Systems’ capability. This approach gives the impression that group
recommendations are retrofits of the Personalized Recommender Systems. Moreover, such an approach
not taken the dynamics of group emotion and individual emotion into the consideration in making top-N
recommendations. Recommending items to a group of two or more users has certainly raised unique
challenges in group behaviors that influence group decision-making that researchers only partially
understand. This study applies the Affective Aware Pseudo Association Method in studying group formation
and dynamics in group decision making. The method shows its adaptability to group's moods change when
making recommendations.
KEYWORDS
User Behavioral Analysis, Text-based Group Emotion-aware Recommender System, Emotion Prediction,
Personality, Cross Information Domain Pseudo Users Association
1. INTRODUCTION
There are two keys design considerations in Group Recommender Systems (GRS). All
commercial Recommender Systems in the market are Group Recommender Systems
(\cite{pujahari2015group}). Group Recommender Systems would focus its design on making and
distributing personalized recommendations to users in the group, and regardless the grouping
formation is initiated by users or by the system. Some Group Recommender Systems provide
users with group formation functions to self-create, manage, maintain, and disband a group. In
most if not all Recommender Systems, for performance and throughput reasons, the system
groups individual users or groups of users with similar preferences and tastes into a group without
their awareness. When a Group Recommender System makes a top-N recommendation to an
active user in a group, the same top-N list will serve all group members. Such a system users’
grouping strategy is known as the system simulcast group (SSG) and is the most common type of
system users’ grouping strategy. Another system users’ grouping strategy is a system broadcast
group (SBG), which refers to distributing the same information to all SSG. SSG and SBG message
distribution is that SSG is intra-group oriented, whereas SBG is inter-group oriented.
The second Group Recommender Systems design point is taking from a user perspective about
user grouping types. A user can be a single member of a uni-group (UG) or a multigroup (MG)
member. Thus, the top-N list distribution could involve a network of SSG groups where the user
is a member. Making top-N recommendations for an active user is relatively straight forward for
a Recommender System. All it requires is to access the user’s preference profile to guide the top-
N recommendations-making process in evaluating and selecting recommendation candidates.
However, making recommendations for a group of users is more challenging than making
personalized recommendations. The Recommender System must adopt a group decision-making
strategy to guide the top-N recommendations making for all group members. In other words, the
Recommender System must consider the group preferences profile and the individual active
user’s preference profile when making the top-N recommendations for a group. Hence, regardless
of whether the preference profile is group-oriented or personal oriented, the profile values are
ever-changing as the group member consumed an item.
People love socializing. From a user’s perspective, a Recommender system should support group
formation management functions for group creation, group deletion, adding members to a group,
removing members from a group, listing members in a group, setting group member privileges,
notifying members in a group, and member vote accounting. Users should not be burden by the
system in preference profile maintenance nor respond to query from the system about the user’s
preferences. The ideal system should know all about its users’ likes and dislikes. It can derive the
needed users’ preference of information by mining the system’s implicit metadata, such as system
transactions and log files.
From a system perspective, it needs to find a viable solution to support all the above functions
and more. For practical purposes, the minimum number of users in a group is two (2). One must
find a means to compare the differences of profiles between a pair of users. This study draws the
lesson learned from prior works in (Leung, Griva, and William G. Kennedy 2020a) and (Leung,
Griva, and Kennedy 2021) that apply the Affective Aware Pseudo Association Method (AAPAM)
to facilitate the affective computing of pairwise object’s preference profile for the system and
users group related operations. This study intends to address what feasible a Recommender
System can do to provide the necessary support for the system’s and users’ group related
operations while addressing the ever-changing users’ profile issue.
2. RELATED WORK IN GROUP RECOMMENDER SYSTEMS
Two or more persons can form a group—for example, friends who meet every weekend for
dinner. A group can temporarily include random people to honor a March of Dimes walk for some
worthy causes. Regardless of the formation of a group is either a recurring basis or an ephemeral
basis, making a recommendation to a group faces two significant challenges: the semantics of
group recommendation and the efficient way to compute for group recommendation (Amer-
Yahia et al. 2009).
To solve problems such as cold start, data sparsity, and scalability, many commercial websites
that deployed Personalized Recommender Systems have incorporated methods by grouping users
with similar preferences to share a list of recommendations for each group and (Masthoff 2015).
Alternatively, in making a recommendation for groups, conventional approaches in information
retrieval and filtering extend Personalized Recommender Systems (Ricci, Rokach, and Shapira
2011). However, this approach gives the impression that group recommendations are retrofits of
their personalized counterparts. Although some researchers argue that recommendations for
groups using the extension approach perform well, such Group Recommender Systems are not
designed with the proper algorithms to solve problems specific to a group setting (Baltrunas,
Makcinskas, and Ricci 2010).
2.1. State of Group Recommender Systems
Group Recommender Systems (GRS), despite the lack of notoriety in the field of Recommender
Systems (RS), have been around as long as their glamorous sibling, the Personalized
Recommender Systems (PRS), which were introduced in the mid-1990s to expedite personalized
information retrieval on the Web which based on users' preferences (Adomavicius and Tuzhilin
2005). Later, when smartphones become prevalent, research on Mobile Recommender Systems
become a significant thrust (Adomavicius and Tuzhilin 2011). Recommender Systems use
different information sources to provide users with predictions and recommendations of items
while balancing factors like accuracy, novelty, dispersity, and stability in the process.
Collaborative Filtering is the favorite choice of method in recommendation processing, which
often couples with other information filtering techniques to form a Hybrid Recommender System
for achieving higher performance and better user experience (Adomavicius and Tuzhilin 2005),
and (Burke 2002). For example, other popular types of information filtering methods include
Content-Based, Knowledge-Based, Demographic-Based, and Social Networking Based Kompan
et al. (Kompan 2012), Adomavicius et al. (Adomavicius and Tuzhilin 2005), Kywe et al. (Kywe,
Lim, and Zhu 2012), De et al. (De Pessemier, Dooms, and Martens 2014), Bobadilla et al.
(Bobadilla et al. 2013), and Burke et al. (Burke 2002). Nonetheless, many of the personalized
recommended items are often (or mostly) used by groups rather than by individuals, for example,
FlyTrap for music (Crossen, Budzik, and Hammond 2002), Pocket RestaurantFinder for
restaurants (McCarthy 2002), e-Tourism for tourism (Garcia, Sebastia, and Onaindia 2011),
PolyLens for movies (O’connor et al. 2001), LET’S BROWSE for group web surfing (Lieberman,
Van Dyke, and Vivacqua 1998), and YuTV for TV shows (Yu et al. 2006) are all Group
Recommenders.
In recent years, various Group Recommender Systems have emerged. Besides building from
scratch, these Group Recommenders most augment a Personalized Recommender to become a
Group Recommender System through one of the two recommendation strategies. The first group
recommendation strategy is "aggregating recommendations" of personalized recommendations
into recommendations for the whole group (Cantador and Castells 2012). It is similar to the
concept, "aggregated predictions," as denoted by Senot et al. (Senot et al. 2010), are the results of
aggregating predictions from an individual user into a group prediction. The second group
recommendation strategy is "aggregating preferences" of the users' preference model into a
group's preference model (Berkovsky and Freyne 2010). Similarly, "aggregated models," as
described by Cantador et al. (Cantador and Castells 2011), refer to aggregate individual user data
into group data.
From a user perspective, when a Group Recommender System allows users to create and manage
groups, the grouping behavior is explicit (Jameson and Smyth 2007). In contrast, a Group
Recommender System derives groups through aggregating recommendations or aggregating
preferences or aggregated predictions or aggregated models; such grouping behaviors denote as
the implicit grouping (Rashid, Karypis, and Riedl 2008). One of the benefits of grouping users
into groups is eradicating the cold start problem that all Recommender Systems face (Park and
Chu 2009).
2.2. Grouping Types
Since different groups exist, group recommender systems aim to manage the heterogeneity of
groups. Boratto et al. (Boratto and Carta 2010) speculated that the formation of a group would
affect its model and thus the predictive capability of a Recommender System. Building on the
work of Jameson et al. ((Jameson and Smyth 2007), which described the four tasks of a
recommender system in detail, Boratto et al. (Boratto and Carta 2010) further extended these four
tasks to four different variants of a group:
• Established group: people who share common interests and explicitly choose to join a
group;
• Occasional group: people who occasionally meet over some activities;
• Random group: people share a resource on occasion without their explicit consent;
• Auto group: People with shared preferences and automatically identified and grouped to
share some scared resources.
2.2.1. Established Group
O’Connor et al. (O’connor et al. 2001) described the established groups which they observed in
PolyLens, a movie Recommender for group use, have a persistent property that users who joined
a group not only shared common interests such as movie watching but also actively participated
in group activities such as rating watched movies. Once a user joined an established group, the
user tends to stay with the group for a long time. Many Group Recommender Systems, such as
PolyLens, apply the Collaborative Filtering method to construct user profiles for an individual
member in the group and then build the group profile by merging individual profiles from the
group members (Kim et al. 2010). To produce recommendations for each group member,
PolyLens uses a Collaborative Filtering algorithm to compute each movie's rating score that meets
the user's preferences. Movies with the highest recommended rates become the Top-N candidates
for recommendation. PolyLens uses a "least misery" selection strategy for making the group's
recommendation (O’connor et al. 2001), i.e., the recommended rating for a group is the lowest
predicted rating for a movie, to ensure every member is satisfied.
However, some Group Recommenders such as Jukola (O’Hara et al. 2004) and PartyVote
(Sprague, Wu, and Tory 2008) are two music recommenders able to make music
recommendations to an established social group of people attending a social event. These two
music recommenders work without requiring any user profiles. Instead, these recommenders
allow any event attendees to express their preferences by selecting a song, album, artist, or genre
from a digital music collection. The rest of the group votes to play songs from the selected list.
The Recommender computes the probability of the voted songs and plays the song which has the
highest probability.
2.2.2. Auto Group
Let v
i
be the vector of the ratings of user i for the items and v
j
be the vector of the ratings of user
j for the items. Cosine similarity measures the similarity of s
ij
between users i and j as expressed
in Boratto et al. (Boratto and Carta 2010) proposed group recommendation algorithm that works
in four steps:
1) Step 1. Cosine similarity between users' ratings matrixes measures users' similarity. The
evaluation procedure (Gfeller, Chappelier, and De Los Rios 2005) illustrates as follows.
Where “” is denoted as the dot product operator and “x” denoted as the multiplication operator.
2) Step 2. Communities detection algorithm proposed by (Blondel et al. 2008) can apply to
the user’s similarity network and generate partitions of different granularities.
3) Step 3. Rating prediction for items rated by enough users of a group by aggregating the
arithmetic mean of users involved for the group. So, for each item i, its rating r
i
is
expressed as:
4) Step 4. Ratings prediction for the remaining unrated items by considering both the rating
and the similarity (t
ij
) of its top similar items:
Where, “” is denoted as the dot product operator.
2.3. User Perspective in Grouping
From a user's perspective, groups' persistence is related to both usage patterns and privacy issues.
If users want to repeatedly receive recommendations for the same group of people, making the
group persistent saves time and effort. Moreover, if groups form and dissociate after a single use,
the ephemeral approach is better to meet the need. Whether groups are private, known only to
group members, or public and accessible to all. Users will behave differently in addressing their
privacy, security, net appearance, and perception needs.
2.3.1. Pseudo User
Pseudo user is not a new concept in the field of recommender systems. However, the pseudo-user
takes on different semantics and roles under different context as described by O'Connor et al.
(O’connor et al. 2001), Melville et al. (Melville, Mooney, and Nagarajan 2002), Resnick et al.
(Resnick and Varian 1997), Martinez et al. (Martínez et al. 2009), Vozalis et al. (Vozalis and
Margaritis 2003), and Lee et al. (Lee, Park, and Park 2008).
2.3.2. Pseudo Grouping
In this research study context, pseudo grouping and virtual grouping are synonyms. Research
studies use pseudo users in recommender systems, as in Martinez et al. (Martínez et al. 2009).
However, no study has investigated leverages pseudo groups as a grouping strategy and plan to
apply pseudo grouping to solve the cold start problem.
2.3.3. Combine Similar Groups to form Super Group
Several studies of recommender systems have applied Supergroup mostly in the study of the
semantic context under content filtering, such as Hu et al. (Hu, Rai12, and Carin 2016). The
supergroup concept is yet widely applied in the grouping strategy of recommender systems
research.
2.4. System Perspective in Grouping
There are times when a Recommender System is the initiator of the grouping strategy. For
example, to improve the performance of making recommendations to users, Recommender
Systems may select to partition users with similar taste into a group and make the same set of
recommendations to all the users in that group to minimize search through the entire rating
database or comparing the similarity of pair-users' profiles. In doing so, the Recommender System
sacrifices accuracy for performance gain by making similar recommendations to a grouped of
individual users.
2.5. Ephemeral and Persistence
The notion of ephemeral as a nature of group had been examined exhaustively by O'Connor et al.
(O’connor et al. 2001), Jameson (Jameson 2004), Schafer et al. (Schafer, Konstan, and Riedl
1999), and Good et al. (Good et al. 1999). On the other hand, Gartrell et al. (Gartrell et al. 2010),
O’Connor et al. (O’connor et al. 2001), Erickson (Erickson 2003), Smeaton et al. (Smeaton and
Callan 2005) had researched the persistence and related issues of groups.
2.6. Merging Strategies
O’Connor et al. (O’connor et al. 2001) developed PolyLens, a collaborative filtering
recommender system designed to recommend movies for a group of viewers rather than
individuals. PolyLens merges the individual user's profile among similar taste users to form the
group profile and applies the similarity function against the rating matrix for suitable items for
making a recommendation to the group. This dissertation research will evaluate other merging
strategies besides the simple aggregation approach.
2.6.1. Weight Settings for Influential Members
Kim et al. (Kim and Srivastava 2007) articulated that social influence impacts e-Commerce
decision-making. Few studies have considered a social influence in an e-Commerce decision
support system because until recently, data about social interaction does not adequately capture
in e-Commerce. It becomes apparent that the customer decision process is influenced by
information from trusted people, not from product manufacturers or recommendation systems.
The social influence from high-quality reviews written by previous consumers can have a direct,
positive effect on potential consumers' decision making, and this effect can propagate through a
social network Huang et al. (Huang and Benyoucef 2013). One of the aims of this dissertation
research is to examine a balanced method in extracting the degree of influence by weighing that
members exert among each other through decision-making.
2.7. Group Recommendation Making Strategies
There are several ways of extending a Personalized Recommender System to a Group
Recommender System as described by De et al. (De Pessemier, Dooms, and Martens 2012).
Merging strategy and Virtual User strategy are two conventional approaches for grouping.
According to Kagita et al. (Kagita, Padmanabhan, and Pujari 2013), there are three ways to
implement the Merging strategy: merged profiles, merging recommendation, and merging score.
Incidentally, Kagita et al. (Kagita, Pujari, and Padmanabhan 2013) also illuminated the Virtual
User strategy. The aspect of grouping strategies for a system is different from that of a user. Next,
to examine are the differences between them. Specifically, an examination on the nature of
ephemeral group versus persistence group from a user perspective and the following grouping
strategies from a system perspective:
1) Making group recommendations for a clustered of users,
2) Making group recommendations by considering transitive precedence relation through a
virtual user model,
3) Adapting personalized recommendation for a group through aggregation strategies,
4) Automatic identification groups of users with similar interests, and
5) Making group recommendations through group behavior modeling.
2.7.1. Making Group Recommendations for a Clustered of Users
To avoid an exhaustive search through the entire user preference database to match a particular
user's preferences, Ntoutsi et al. (Ntoutsi et al. 2012) advocated a group recommendation system
that enhances recommendations partitioning users into clusters based on similar preferences.
Aggregated preferences of clustered members drive the decision making of recommendations for
users. Popescu et al. also studied aggregated user preference (Popescu and Exam 2011) and
Masthoff et al. (Masthoff 2015). The algorithms to estimate the relevance of an item for a user
deployed by Ntoutsi et al. (Ntoutsi et al. 2012) for the group recommendations framework come
in two flavors:
1) In studies of Amer et al. (Amer-Yahia et al. 2009) and Konstan et al. (Konstan et al.
1997). both have illustrated personalized recommendations produced from evaluating
relevance scores for unrated items of an unusually active user based on collaborative
filtering techniques. However, users typically rate only a few items against the vast
amount of the available items. Thus, the notion of support measures the percentage of the
active user's friends who have expressed preferences for the item to minimize the rating
matrix's skewness. The relevance and support scores are then combined to estimate a
recommendation worthiness score of an item for a user.
2) In addition to personalized recommendations, group recommendations deploy a model
built on context information of users in the group as illustrated in Amer et al. (Amer-
Yahia et al. 2009) by combining all individual users' preferences. Moreover, Ntoutsi et
al. (Ntoutsi et al. 2012) further refined these group preferences by three aggregation
design methods that carry different semantics. Firstly, the Least Misery design will
capture cases where strong user preferences act as a veto. Next, the Fair design will
capture more common cases where most of the group members are satisfied. Lastly, the
Most Optimistic design will capture cases where the most satisfied member of the group
acts as the most influential member. After applying these three design methods
appropriately, a Top-K list then uses in making recommendations to users in the group.
2.7.2. Making Group Recommendations through Aggregation Functions
Two dominant strategies for making a Personalized Recommender to become a Group
Recommender (Berkovsky and Freyne 2010). The first grouping strategy is to aggregate
individual preferences into a recommendation list. In effect, this approach creates a pseudo-user
for a group based on its group members and then makes recommendations based on the pseudo
user's preferences. The second grouping strategy is to aggregate the individual member's
recommendation list to form the group recommendation list. In other words, every single user in
the group will receive an individual recommendation list. All individual recommendation lists
combine to form a recommendation list for the group. The aim of making a group
recommendation is to compute a recommendation score for each candidate item that reflects the
interests and preferences of all group members. An acceptable approach to obtain a consensus of
group ranking recommendation score for a candidate item is through some aggregation functions.
Popular aggregation functions are as follows.
Least-misery:
Average:
Average without Misery:
where
2.7.3. Adapting Personalized Recommendation for Group through Aggregation Strategies
The main problem of group recommendation stems from adopting personalized recommendation
based on individual user preference. Recommender Systems have to learn users' preferences from
user rating records. After learning individual user preferences, recommender systems then
aggregate user rating information to provide group recommendations. Aggregation strategies
research is ongoing. Researchers such as Mastho et al. (Masthoff 2015) developed various
aggregation strategies that later research built-on. Highlight below is a few of their aggregation
strategies. Merging individual user profiles as the group's preferences is the superior strategy.
The utilitarian strategy considers utility values for group recommendation. The utility values are
of two types, i.e., additive or multiplicative. An example of a movie recommender system first
adds/multiply all the movie ratings separately for a user group. Then display the movies list in the
highest aggregate utility value as recommended items.
Most pleasure strategy generates a group rating list based on the maximum of individual ratings.
For example, movies with the highest universal rating values will add to the recommended list.
The least misery strategy generates a group rating list based on the minimum of individual ratings.
Then the item with a high minimum individual rating will be recommended. The idea behind this
strategy is that a group is as happy as its least happy member.
Padmanabhan et al. (Padmanabhan, Seemala, and Bhukya 2011) implemented a rule-based
aggregation strategy called RTL strategy to group recommender systems by combining all of the
above three strategies. This algorithm performs better regarding accuracy in group
recommendation as compared to the preceding three grouping strategies. However, RTL has no
provision for creating a better group (regrouping). Pujahari et al. (Pujahari and Padmanabhan
2015) solved the regrouping deficiency by applying Predictive Rule Mining algorithms (Machado
2003), which featured learning rules to learn individual users' preferences applying a new
algorithm for making group recommendation.
3. DATASETS
Developing an Emotion Aware Recommender System requires an emotion labeled dataset to
work with various machine learning algorithms. However, there is no readily available emotion
labeled dataset in any publicly accessible repository. One can overcome the deficiency by mining
emotion features from existence metadata in the dataset. One can demonstrate making top-N
movie recommendations to users in a group through an Emotion Aware Recommender System.
Thus, it is necessary to mine the needed emotion label from the subjective text in movie branch
mark datasets from the MovieLens repository and The Movie Database (TMDb) website, where
the movie overview contains the subjective text metadata. One can download a variety of
MovieLens movie datasets stored in the GroupLens repository. However, one needs to scrape the
TMDb website for needed metadata. Before performing any emotion modeling on a dataset, one
needs to develop a text-based emotion classifier to classify the movie overview's emotion features.
3.1. MovieLens Datasets
In the article Leung et al. (Leung, Griva, and William G. Kennedy 2020b), it worked with (Riedl
and Konstan 2015) advocated the four MovieLens movie benchmark datasets, namely, ml-latest-
small (a.k.a. mlsm), ml-20m (a.k.a. ml20m), ml-25m (a.k.a. ml25m), and ml-latest (a.k.a. ml27m
hereafter) datasets. (Leung, Griva, and William G. Kennedy 2020b) also illustrated the need to
scrape the TMDb for movie overview metadata to use as input to the Tweets Affective Classifier
(TAC) to classify the scraped TMDb dataset's emotion labels before merging with MovieLens
datasets. Table 1 illustrates the statistics of each mentioned MovieLens dataset after merging with
the emotion labeled TMDb dataset. The name of a MovieLens dataset reflects the number of
ratings it contains. Please note that the Number of Users column does not match up the Number
of Overviews column. The reason is that the collection of movies in MovieLens and TMDb is not
identical.
Table 1. Statistics of MovieLens datasets.
Attribute
Dataset
No. Users
No. Movies
No. Ratings
No. Overviews
mlsm
610
9742
100836
9625
ml20m
138493
27278
20000263
26603
ml25m
162541
62423
25000095
60494
ml27m
283228
58098
27753444
56314
In each of the four MovieLens datasets, there are two data files named ratings, and tags contain
user ids as a unique identifier. MovieLens stated that user ids found in ratings and tags data files
are consistent within the same dataset but are not uniform across different datasets (grouplens.org
2019). For example, in (Leung, Griva, and William G. Kennedy 2020b), user id 400 is only
compatible within the same MovieLens MLSM dataset and is not across ml20m, ml25m, ml27m
datasets. In other words, user id 400 in other MovieLens datasets are not the same user id 400 as
in the MLSM dataset. However, (Leung, Griva, and William G. Kennedy 2020b) has
demonstrated by using the Affective Aware Pseudo Association Method (AAPAM), the disjoint
user id 400 in MLSM can correctly connect to the proper user id in MovieLens datasets, as
depicted in Table 2. For example, user id 400 of MLSM can Pseudo Associate Connect (PAC) to
user id 66274 of MLSM or PAC to user id 95450 of ml25m or PAC to user id 89195 of ml27m.
All mentioned user id are disjoint users of each other. In the context of this study, disjoint users
refer to two or more individual mutually exclusive users whose user id are different and reside in
different datasets within the same or different information domains whose emotion profiles,
UVEC, are highly similar or even identical. One can consider the disjoint users as identical pseudo
users.
3.2. Affective Aware Pseudo Association Method
The Affective Aware Pseudo Association Method (AAPAM) computes the Affective Index
Indicator (AII) using the Cosine Similarity algorithm (Bigdeli and Bahmani 2008), as depicted in
Equation 2, Cosine Similarity expresses the closeness of the emotion profiles between two users
or items. When using AAPAM to compare pairwise between User id 400 of MLSM against users
in other MovieLens datasets, AII reveals, as depicted in Table 2, the closest other users’ emotion
profiles that match the candidate user. User id 400 in MLSM can make a pseudo associate
connection (PAC) to user id 66274 with AII 0.999916 in ml20m or to user id 95449 with AII
0.999999 ml25m, or user id 89195 with AII 0.999999 in ml27m, respectively.
AAPAM also worked with The Movie Database (TMDb) (TMDb 2018). By scraping from
TMDb, it yields the movie metadata for movie overviews, poster images, and other metadata.
AAPAM applied the Tweets Affective Classifier (TAC), a method developed in (Leung, Griva,
and William G Kennedy 2020), to classify a movie emotion profile. A movie emotion profile is
also known as a movie vector, MVEC, which represents a multi-dimensional embedding of a
probability distribution of seven primary human emotions: neutral, happiness, sadness, hate,
anger, disgust, and surprise. Each user in (Leung, Griva, and William G Kennedy 2020) also has
a user emotion profile, UVEC, where it contains the average value of all movies MVECs the user
has watched.
Table 2. Pseudo Association Connection of ml-latest-small user id 400 to other users in different
datasets through affective index indicator.
Dataset
mlsm
ml20m
ml25m
ml27m
User1
ID
PAC
400
66274
95459
89195
User1
Movie
Count
43
22
43
43
User1
Watched
List
movieID
6
47
50
260
…,
122886
134130
164179
168252
47
260
300
307
...,
2628
2797
3418
3481
6
47
50
260
…,
122886
134130
164179
168252
6
47
50
260
…,
122886
134130
164179
168252
User1
UVEC
Neutral
Happiness
Sadness
Hate
Anger
Disgust
Surprise
0.16353
0.08874
0.12709
0.20332
0.11934
0.15881
0.13918
0.16250
0.08609
0.12654
0.20701
0.11776
0.16005
0.14005
0.16353
0.08874
0.12709
0.20332
0.11934
0.15881
0.13918
0.16353
0.08874
0.12709
0.20332
0.11934
0.15881
0.13918
User1
Affective
1.0
0.99992
0.99999
0.99999
Index
Indicator
As illustrated in Table 2, user id 400 in the rating data file of the MLSM dataset has watched 43
movies; taking the average of all the 43 movies’ MVECs yields the UVEC for user id 400. As
mentioned in (Leung, Griva, and William G Kennedy 2020), a movie's MVEC is static and stays
unchanged throughout the film’s life; whereas, a user's UVEC changes its value each time the
user watches a movie. The user’s UVEC reflects the up-to-date movie taste and preference of the
user. A movie MVEC is unique, while a UVEC may not be unique when two users watched the
same movie set.
Besides, the AAPAM method can PAC connect disjoint users from different datasets within the
same domain; this study believes the same technique can PAC connect disjoint users and items
among different datasets across different domains. Unlike MovieLens datasets, some other movie
datasets such as TMDb highlight the average voting score, the sentiment rating value on a scale
of 1 (lowest) to 10 (highest), of a movie by a group of users who have watched and rated the
movie through the voting count attribute in the data file instead of individual user’s sentiment.
TMDb does not maintain user information, and neither contains a user-id field in the dataset.
When applying the AAPAM to connect MovieLens and TMDb domains, the PAC connection
applies to movie items between MovieLens and TMDb. Here, the PAC connection between movie
A in MovieLens to movie B in TMDb indicates how similar the two movies’ emotion profiles,
MVECs, form a one-to-one relationship. However, when applying the PAC to connect a user in
MovieLens and a movie item in TMDb, the movie item MVEC in TMDb must first be normalized
with the respective voting count. The normalized MVEC represents the average UVEC of the
group of users who have rated it. Thus, the PAC connection between user A in MovieLens to the
normalized MVEC of movie B in TMDb indicates how similar user A to a group of users B is in
the form of a one-to-many relationship.
4. METHODOLOGY
This study utilized affective features in two data sources that it uses. The affective features applied
to users’ emotion profiles, UVECs, and items’ emotion profiles, IVECs, or MVECs equivalent.
No disjoint users and items can interconnect without adding affective features across data sources
of MoieLens, and TMDb. Affective aware features are added to the data sources through Tweets
Affective Classifier (TAC) as developed in (Leung, Griva, and William G Kennedy 2020). Table
3 depicts samples after added affective features UVEC and MVEC in this study’s data sources.
Table 3. TMDb movie emotion profile example.
tmdbId
2
525662
movieId
4470
189111
Mood
Disgust
Hate
Neutral
0.15705037
0.11876434
Happiness
0.08608995
0.05086204
Sadness
0.15583897
0.12669845
Hate
0.07506061
0.3391073
Anger
0.08469571
0.13069303
Disgust
0.26612538
0.13746719
Surprise
0.17513901
0.096407644
4.1. Tweets Affective Classifier
Leung et al. in (Leung, Griva, and William G Kennedy 2020) illustrated the Tweets Affective
Classifier (TAC) development, a text-based tweets emotion classifier capable of detecting and
recognizing six basic human emotions advocated by Paul Ekman. The six emotions are happiness,
sadness, fear, anger, disgust, and surprise. For affective computing convenience, one designs TAC
capable of detecting and recognizing natural emotion. One then feeds TMDb’s movie overview
to TAC to classify the overview’s moods or the movie’s emotion profile, MVEC depicted in Table
4. After joining the emotion-labeled TMDb data file with the rating data file in MovieLens, it
yields an emotion-labeled MovieLens data file depicted in Table 5. With the joined emotion-
labeled data file, one computes the users’ emotion profiles, UVEC, by taking the MVEC from all
the movies that the user has watched.
Table 4. First few rows of the cleansed emotion labeled MovieLens data file.
Index
tid
mid
iid
mood
neutral
happy
sad
hate
anger
disgust
surprise
1
2
4470
94675
disgust
0.157
0.086
0.156
0.075
0.085
0.266
0.175
2
5
18
113101
disgust
0.121
0.060
0.098
0.128
0.133
0.244
0.216
3
6
479
107286
hate
0.075
0.114
0.054
0.433
0.095
0.128
0.100
4
11
260
76759
neutral
0.299
0.262
0.079
0.030
0.017
0.083
0.230
5
12
6377
266543
surprise
0.150
0.080
0.055
0.083
0.103
0.153
0.376
Table 5. Sample if users' emotion profiles derived from taking the average of the user’s moods
values from all movies the user has watched
mlsm
id
neutral
joy
sadness
hate
anger
disgust
surprise
400
0.163529
93
0.088735
25
0.1270899
8
0.203318
40
0.119338
19
0.158812
87
0.1391753
8
414
0.166351
88
0.097305
81
0.1180924
0.164195
1
0.115177
99
0.172503
15
0.1663736
7
474
0.168858
31
0.099746
59
0.1187206
0.160877
16
0.112612
72
0.171919
68
0.1672649
5
448
0.172833
09
0.096858
13
0.1160457
3
0.161207
33
0.112276
07
0.170985
78
0.1697938
9
4.2. System Grouping
This paper advocates a method for the system to group users in a system simulcast group (SSG)
according to the similarity in users’ emotion profiles, UVEC, so that all SSG users share the same
top-N recommendations list. However, to simulate the personalized recommendation
functionality before pushing the SSG top-N list to an active user in the SSG, the system reranked
the top-N list’s MVEC against the active user’s UVEC, thereby ensuring all users in the SSG
receives personalized individual top-N recommendations list. The benefit of grouping users in
SSG is improving the system’s throughput while performing fewer top-N recommendations made
by the Recommender System. For example, a system hosts n million users. The system must run
a top-N recommendation process per-user or n million top-N recommendations operations to
make
. personalized recommendations. By grouping users into m users per SSG, the system
only needs to make
top-N or
. It is an m fold decrease in top-N recommendations making.
Thus, by grouping users in SSG, the system can expand the handling of more users without
upsizing the system.
This paper has devised the following scheme to demonstrate an SSG’s working in an Emotion
Aware Recommender System. MovieLens ml-latest-small (MLSM) ratings data file contains
100,836 ratings of 9743 movies by 610 users. Some users rated 20 movies, the minimum number
of movies a user must have rated in MovieLens sampled datasets, while some users have rated
thousands of movies. By sorting the MLSM rating data file in descending order with users rated
most movies at the top, draw out the top ten users to act as the dominant user in one of the ten
SSG groups. Table 7 depicts ten dominant users to act as anchor user for the ten SSG group. Table
7 also shows the sample list of SSG group members. With each SSG’s dominant user’s UVEC, it
computed against all undraw users through pairwise Cosine Similarity of UVEC. It ranked the
result in descending order. In effect, the result shows the Affective Index Indicator (AII) of each
member’s UVEC relative to the UVEC of the dominant user. Draw out the top 60 users who have
the most similar emotion profile of the dominant user to join the SSG. Continue the same step to
draw out members for the other nine SSG. The grouping scheme forming ten SSG each has a
dominant user who rated most movies and 60 members with closely matched emotion profiles
against the dominant user’s UVEC.
IMDb periodically publishes a list containing the “IMDb 100 Greatest Movies of All Time.”
Using the list as top-N recommendations, one can feed to all ten SSG. Doing so, one simulates
the system broadcast group (SBG) broadcast message function. Before sending the top-N
recommendations to an SSG member, the system reranked the user’s UVEC against all movie
emotion profiles, MVEC, on the list through the pairwise Cosine Similarity. This paper picks
three users from one of the SSG to demonstrate the reranking process. Hence the SSG organized
by ranking all members’ UVEC by computing their Affective Index Indicator (AII). The dominant
member occupies the top of the AII list. The least-misery member occupies the bottom of the AII
list, and the 30th member in the AII list is the average member of the SSG.
Table 6 showing the reranked top 10 of the “IMDb 100 Greatest Movies of All Time” for
MovieLens ml-latest-small user ID 414 using UVEC of 20% and 100% movie watching history
to compute the pairwise AII of Cosine Similarity. The two reranked lists of user ID 414 show a
different rank order from the movie ranking of IMDb and the more movies the user has watched.
Users with different UVEC affect the reranked result.
Table 6. User ID 414 Reranked IMDb 100 Greatest Movies of All Time using 20% and 100%
Movie Watching UVEC
Ranking
100 All
Time
Greatest
Movies
Moive
ID
Movie Title
UserId414
Reranked
(UVEC
20%
watched)
UserId414 Reranked
Movie Title (UVEC
20% Watched)
UserId414
Reranked
(UVEC
100%
watched)
UserId414 Reranked
(AII - UVEC 100%
Watched)
1
858
The Godfather (1972)
1252
Chinatown (1974)
1252
Chinatown (1974)
2
1221
The Godfather: Part II
(1974)
1213
Goodfellas (1990)
1213
Goodfellas (1990)
3
2019
Seven Samurai (1954)
899
Singin' in the Rain
(1952)
318
The Shawshank
Redemption (1994)
4
296
Pulp Fiction (1994)
318
The Shawshank
Redemption (1994)
899
Singin' in the Rain
(1952)
5
1203
12 Angry Men (1957)
912
Casablanca (1942)
8125
Sunrise (1927)
6
5618
Spirited Away (2001)
8125
Sunrise (1927)
26150
Andrei Rublev (1966)
7
527
Schindler's List (1993)
26150
Andrei Rublev (1966)
912
Casablanca (1942)
8
912
Casablanca (1942)
89759
A Separation (2011)
89759
A Separation (2011)
9
1219
Psycho (1960)
1251
8½ (1963)
1237
The Seventh Seal
(1957)
10
1213
Goodfellas (1990)
1237
The Seventh Seal
(1957)
1251
8½ (1963)
4.3. User Grouping
One significant difference between a Group Recommender System and Personalized
Recommender System is that Group Recommender System supports users with functions to
initiate group formation, and Personalized Recommender System cannot A Group Recommender
System allows the system and users to create a multi-group (MG) and explicitly join the group.
The system provides functions for the user to self manages group administration and group
maintenance. An MG's joining can be by invitation only known as private multi-group (PMG) or
open to public multi-group (OMG). When the system makes a top-N recommendation to an MG,
all MG members share the top-N list, and no need to re-ranked the top-N is necessary for
individual users in the MG. However, the Group Recommender System may apply a decision-
making support strategy in top-N recommendations to assist the group decision making in
selecting a recommendation in the top-N list. This study focuses on applying affective awareness
recommendations to support group decision-making by the most dominant member strategy,
average members’ mood strategy, and the Least-misery member strategy. The most dominant
member in an MG has the largest number of movies a member has watched among the group.
The least-misery member is the member whose emotion profile is most dissimilar to the MG’s
most dominant member. The least-misery member of the group determines by the pairwise
Affective Index Indicator (AII) of all members’ emotion profiles in the MG against the most
dominant member emotion profile, then rank the AII list in the descending order, the least-misery
member occupies the lowest rank. Depicts in Table 8 is a user MG of size 5. User ID 195 is the
dominant member of the group, and user ID 463 is the least-misery member.
5. Results
5.1. Group Formation by System
One needs to determine the number of simulcast group (SSG), g, that wants to form. The system
can make a rank list on user interaction and pick a reasonable number for g. In the MovieLens
dataset, the system can compute the number of movies watching history as the system interaction
criteria. Each SSG will anchor with a member pick from the top of the interaction rank list. The
system computes the pairwise Cosine Similarity between the anchor member of the SSG against
other ungrouped users. After ranking the pairwise list, the system will move m number of users
from the top of the ranked list to the SSG. Depicted in Table 7 is 10 SSG on MovieLens ml-latest-
small dataset. Each SSG anchors with a dominant user with most movie watching history. Each
SSG contains 61 members, including the anchor member. The system picks members by the
Affective Index Indicator (AII) values highly similar to the anchor member.
Table 7. 10 System Simulcast Groups Formed by Using MovieLens ml-latest-small
Rank
Dominant
User ID
Movie
Count
System Simulcast Group Member User ID
Member
Count
1
414
2698
(1) 414, (2) 232, …, (31) 212, …, (61) 167
61
2
599
2478
(1) 599, (2) 477, …, (31) 198, …, (61) 474
61
3
474
2108
(1) 474, (2) 560, …, (31) 260, …, (61) 350
61
4
448
1864
(1) 448, (2) 226, …, (31) 453, …, (61) 389
61
5
274
1346
(1) 274, (2) 330, …, (31) 424, …, (61) 477
61
6
610
1302
(1) 610, (2) 160, …, (31) 405, …, (61) 218
61
7
68
1259
(1) 68, (2) 414, …, (31) 212, …, (61) 593
61
8
380
1218
(1) 380, (2) 160, …, (31) 218, …, (61) 318
61
9
606
1115
(1) 606, (2) 177, …, (31) 6, …, (61) 66
61
10
288
1055
(1) 288, (2) 483, …, (31) 555, …, (61) 167
61
5.2. Grouping by Users
Once a Group Recommender System provides users with group formation functionality, users
can use the group formation to create and maintain a multi-users group (MG). In the following
illustration, all members subsampled randomly from the derived emotion labeled movie dataset
MLSM. Depicted below is a five-member multi-user group. The group average UVEC derived
from the mean of five members’ UVEC. User ID 195 watched most movies, thus the dominant
member, whereas user ID 463 is the least-misery member.
Table 8. A Five Member Multi-User Group
Rank
UserId
Watched
UVEC
1
195
187
0.1639455, 0.0902557, 0.1176815, 0.1726736, 0.1185870,
0.1777129, 0.1591437
2
602
135
0.1639545, 0.0869860, 0.1168919, 0.16947266, 0.1156349,
0.1817310, 0.1653290
3
190
66
0.1603803, 0.0849701, 0.1254172, 0.17182250, 0.1135154,
0.1797844, 0.1641099
4
521
40
0.1574143, 0.0944750, 0.1240710, 0.14589457, 0.1083259,
0.1795868, 0.1902323
5
463
33
0.1558253, 0.0968441, 0.1140474, 0.19975860, 0.1226243,
0.1571110, 0.1537890
Average
Group
UVEC
0.1603040, 0.0907061, 0.1196220, 0.17192440, 0.1157376,
0.1751852, 0.1665208
5.3. Group Decision Making Strategies
A reasonable way to handle a group decision making using the dominant member strategy is to
generate the top-N recommendations based on the dominant user’s preference. Similarly, to
support the group decision making using the least-misery member strategy, the top-N
recommendations are generated based on the least-misery member’s preference. Instead of using
a movie database to generate a disparate top-N list based on different decision-making strategies,
this paper limits the overall movie selection range to generate top-N listed from the “IMDb 100
Greatest Movies of All Time”. Each of the movies in the IMDb list has computed an emotion
profile MVEC. By computing the Affective Index Indicator (AII) pairwise between the UVEC of
the preferred user group decision-making strategy and the IMDb MVEC and reranked the result
in descending order will yield the desire top-N list. Depicted in Table 9 listed the top-10 of
dominant user ID 195, least-misery user ID 463, and using the five-member MG average UVEC.
Table 9 shows the top-10 recommendations list generated to support MG users’ decision-making
strategies of the dominant member, least-misery member, and average group user profile.
Table 9. Top-10 Generated by Dominant Member, Least-misery Member and Average Group
User Profile Decision-making Strategies
Top-
10
Rank
UserId195
Dominant
User Strategy
Top-10
UserId195
Dominant
user Top-10
Movie Title
UserId463
Least-misery
User Strategy
Top-10
UserId463
Least-misery
User Top-10
Movie Title
Average
Group User
Strategy
Top-10
Average
Group User
Top-N Movie
Title
1
858
The Godfather
(1972)
1252
Chinatown
(1974)
1252
Chinatown
(1974)
2
1221
The
Godfather:
Part II (1974)
1213
Goodfellas
(1990)
1213
Goodfellas
(1990)
3
2019
Seven
Samurai
(1954)
899
Singin' in the
Rain (1952)
318
The Shawshank
Redemption
(1994)
4
296
Pulp Fiction
(1994)
318
The Shawshank
Redemption
(1994)
899
Singin' in the
Rain (1952)
5
1203
12 Angry Men
(1957)
912
Casablanca
(1942)
8125
Sunrise (1927)
6
5618
Spirited Away
(2001)
8125
Sunrise (1927)
26150
Andrei Rublev
(1966)
7
527
Schindler's
List (1993)
26150
Andrei Rublev
(1966)
912
Casablanca
(1942)
8
912
Casablanca
(1942)
89759
A Separation
(2011)
89759
A Separation
(2011)
9
1219
Psycho (1960)
1251
8½ (1963)
1237
The Seventh
Seal (1957)
10
1213
Goodfellas
(1990)
1237
The Seventh
Seal (1957)
1251
8½ (1963)
6. FUTURE WORK
The author et al. plans to widen the study of applying the Affective Aware Pseudo Association
Method in other aspects of Group Recommender Systems besides group formation and group
decision-making support. The current study relied on pre-processed users’ and items’ emotion
profiles. Future studies will investigate the real-time processing of objects’ emotion profile to
accommodate the new arrival of objects to the Recommender Systems.
7. CONCLUSION
Human preference in decision-making is highly influenced by their moods. By capturing and
modeling emotion features in users and items can better reflect human preferences. This paper
advocated applying the Affective Index Indication (AII) developed by the Affective Aware
Pseudo Association Method (AAPAM) to obtain pairwise similarity metrics for comparing the
closeness between two objects. In the study, the AII was applied to compare two users’ emotion
profiles UVEC to determine the closeness between users. The same technique was also applied
to find the closeness between a user’s emotion profile UVEC and a list of movies’ emotion
profiles MVEC. Provided that movies and users have precomputed with MVEC and UVEC, to
make movie top-N recommendations by AII technique is straight forward. The AII generated top-
N list also capable of adapting to the dynamic of the user’s mood change. AII technique can
support various decision-making strategies from group users and support system and user-
oriented group formation.
REFERENCES
Adomavicius, Gediminas and Alexander Tuzhilin (2005), “Toward the next generation of recommender
systems: A survey of the state-of-the-art and possible extensions,” IEEE transactions on
knowledge and data engineering, 17 (6), 734–49.
Adomavicius, Gediminas and Alexander Tuzhilin (2011), “Context-aware recommender systems,” in
Recommender systems handbook, Springer, 217–53.
Amer-Yahia, Sihem, Senjuti Basu Roy, Ashish Chawlat, Gautam Das, and Cong Yu (2009), “Group
recommendation: Semantics and efficiency,” Proceedings of the VLDB Endowment, 2 (1), 754–
65.
Baltrunas, Linas, Tadas Makcinskas, and Francesco Ricci (2010), “Group recommendations with rank
aggregation and collaborative filtering,” in Proceedings of the fourth ACM conference on
Recommender systems, 119–26.
Berkovsky, Shlomo and Jill Freyne (2010), “Group-based recipe recommendations: analysis of data
aggregation strategies,” in Proceedings of the fourth ACM conference on Recommender systems,
ACM, 111–18.
Bigdeli, Elnaz and Zeinab Bahmani (2008), “Comparing accuracy of cosine-based similarity and
correlation-based similarity algorithms in tourism recommender systems,” in Management of
Innovation and Technology, 2008. ICMIT 2008. 4th IEEE International Conference on, IEEE,
469–74.
Blondel, Vincent D, Jean-Loup Guillaume, Renaud Lambiotte, and Etienne Lefebvre (2008), “Fast
unfolding of communities in large networks,” Journal of statistical mechanics: theory and
experiment, 2008 (10), P10008.
Bobadilla, Jesús, Fernando Ortega, Antonio Hernando, and Abraham Gutiérrez (2013), “Recommender
systems survey,” Knowledge-Based Systems, 46, 109–32.
Boratto, Ludovico and Salvatore Carta (2010), “State-of-the-art in group recommendation and new
approaches for automatic identification of groups,” in Information retrieval and mining in
distributed environments, Springer, 1–20.
Burke, Robin (2002), “Hybrid recommender systems: Survey and experiments,” User modeling and user-
adapted interaction, 12 (4), 331–70.
Cantador, Iván and Pablo Castells (2011), “Extracting multilayered Communities of Interest from
semantic user profiles: Application to group modeling and hybrid recommendations,” Computers
in Human Behavior, 27 (4), 1321–36.
Cantador, Iván and Pablo Castells (2012), “Group recommender systems: new perspectives in the social
web,” in Recommender systems for the social web, Springer, 139–57.
Crossen, Andrew, Jay Budzik, and Kristian J Hammond (2002), “Flytrap: intelligent group music
recommendation,” in Proceedings of the 7th international conference on Intelligent user
interfaces, ACM, 184–85.
De Pessemier, Toon, Simon Dooms, and Luc Martens (2012), “Design and evaluation of a group
recommender system,” in Proceedings of the sixth ACM conference on Recommender systems,
ACM, 225–28.
De Pessemier, Toon, Simon Dooms, and Luc Martens (2014), “Comparison of group recommendation
algorithms,” Multimedia Tools and Applications, 72 (3), 2497–2541.
Erickson, Thomas (2003), “Designing visualizations of social activity: six claims,” in CHI’03 extended
abstracts on Human factors in computing systems, ACM, 846–47.
Garcia, Inma, Laura Sebastia, and Eva Onaindia (2011), “On the design of individual and group
recommender systems for tourism,” Expert systems with applications, 38 (6), 7683–92.
Gartrell, Mike, Xinyu Xing, Qin Lv, Aaron Beach, Richard Han, Shivakant Mishra, and Karim Seada
(2010), “Enhancing group recommendation by incorporating social relationship interactions,” in
Proceedings of the 16th ACM international conference on Supporting group work, ACM, 97–
106.
Gfeller, David, Jean-Cédric Chappelier, and Paolo De Los Rios (2005), “Finding instabilities in the
community structure of complex networks,” Physical Review E, 72 (5), 056135.
Good, Nathaniel, J Ben Schafer, Joseph A Konstan, Al Borchers, Badrul Sarwar, Jon Herlocker, John
Riedl, and others (1999), “Combining collaborative filtering with personal agents for better
recommendations,” in AAAI/IAAI, 439–46.
grouplens.org (2019), “Movielens ml-20m dataset readme,” (accessed October 8, 2019), [available at
http://files.grouplens.org/datasets/movielens/ml-20m-README.html].
Hu, Changwei, Piyush Rai12, and Lawrence Carin (2016), “Non-negative Matrix Factorization for
Discrete Data with Hierarchical Side-Information,” in Proceedings of the 19th International
Conference on Artificial Intelligence and Statistics, 1124–32.
Huang, Zhao and Morad Benyoucef (2013), “From e-commerce to social commerce: A close look at
design features,” Electronic Commerce Research and Applications, 12 (4), 246–59.
Jameson, Anthony (2004), “More than the sum of its members: challenges for group recommender
systems,” in Proceedings of the working conference on Advanced visual interfaces, ACM, 48–
54.
Jameson, Anthony and Barry Smyth (2007), “Recommendation to groups,” in The adaptive web,
Springer, 596–627.
Kagita, Venkateswara Rao, Vineet Padmanabhan, and Arun K Pujari (2013), “Precedence Mining in
Group Recommender Systems,” in Pattern Recognition and Machine Intelligence, Springer,
701–7.
Kagita, Venkateswara Rao, Arun K Pujari, and Vineet Padmanabhan (2013), “Group Recommender
Systems: A Virtual User Approach Based on Precedence Mining,” in AI 2013: Advances in
Artificial Intelligence, Springer, 434–40.
Kim, Jae Kyeong, Hyea Kyeong Kim, Hee Young Oh, and Young U Ryu (2010), “A group
recommendation system for online communities,” International Journal of Information
Management, 30 (3), 212–19.
Kim, Young and Jaideep Srivastava (2007), “Impact of social influence in e-commerce decision making,”
in Proceedings of the ninth international conference on Electronic commerce, ACM, 293–302.
Kompan, Michal (2012), “Group Inspired Single-User Collaborative Filtering,” Information Sciences and
Technologies Bulletin of the ACM Slovakia, 4 (2), 35–37.
Konstan, Joseph A, Bradley N Miller, David Maltz, Jonathan L Herlocker, Lee R Gordon, and John Riedl
(1997), “GroupLens: applying collaborative filtering to Usenet news,” Communications of the
ACM, 40 (3), 77–87.
Kywe, Su Mon, Ee-Peng Lim, and Feida Zhu (2012), “A survey of recommender systems in twitter,” in
Social Informatics, Springer, 420–33.
Lee, Tong Queue, Young Park, and Yong-Tae Park (2008), “A time-based approach to effective
recommender systems using implicit feedback,” Expert systems with applications, 34 (4), 3055–
62.
Leung, John Kalung, Igor Griva, and William G. Kennedy (2020a), “An Affective Aware Pseudo
Association Method to Connect Disjoint Users Across Multiple Datasets – an enhanced
validation method for Text-based Emotion Aware Recommender,” International Journal on
Natural Language Computing (IJNLC), 9, 11–31.
Leung, John Kalung, Igor Griva, and William G. Kennedy (2020b), “An Affective Aware Pseudo
Association Method to Connect Disjoint Users Across Multiple Datasets – an enhanced
validation method for Text-based Emotion Aware Recommender,” International Journal on
Natural Language Computing (IJNLC) Vol, 9 (4).
Leung, John Kalung, Igor Griva, and William G Kennedy (2020), “Using Affective Features from Media
Content Metadata for Better Movie Recommendation Making,” in Proceedings of the 12th
International Joint Conference on Knowledge Discovery, Knowledge Engineering and
Knowledge Management, Science and Technology Publications, LDA, 161–68.
Leung, John Kalung, Igor Griva, and William G. Kennedy (2021), “Making Cross-Domain
Recommendations by Associating Disjoint Users and Items Through the Affective Aware
Pseudo Association Method,” in Proceedings of 8th International Conference on Artificial
Intelligence and Applications (AIAP 2021), Zurich, Switzerland: Computer Science &
Information Technology (CS & IT), 113–29.
Lieberman, Henry, Neil W Van Dyke, and Adrian S Vivacqua (1998), “Let’s browse: a collaborative
Web browsing agent,” in Proceedings of the 4th international conference on Intelligent user
interfaces, ACM, 65–68.
Machado, FP (2003), “CPAR: Classification based on predictive association rules,” IEICE Transactions
on Information and Systems.
Martínez, Ana Belén Barragáns, José J Pazos Arias, Ana Fernández Vilas, Jorge García Duque, and
Martin Lopez Nores (2009), “What’s on TV tonight? An efficient and effective personalized
recommender system of TV programs,” IEEE Transactions on Consumer Electronics, 55 (1),
286–94.
Masthoff, Judith (2015), “Group Recommender Systems: Aggregation, Satisfaction and Group
Attributes,” in Recommender Systems Handbook, Springer, 743–76.
McCarthy, Joseph F (2002), “Pocket restaurantfinder: A situated recommender system for groups,” in
Workshop on Mobile Ad-Hoc Communication at the 2002 ACM Conference on Human Factors
in Computer Systems.
Melville, Prem, Raymond J Mooney, and Ramadass Nagarajan (2002), “Content-boosted collaborative
filtering for improved recommendations,” in Aaai/iaai, 187–92.
Ntoutsi, Irene, Kostas Stefanidis, Kjetil Norvag, and Hans-Peter Kriegel (2012), “grecs: A group
recommendation system based on user clustering,” in International Conference on Database
Systems for Advanced Applications, Springer, 299–303.
O’connor, Mark, Dan Cosley, Joseph A Konstan, and John Riedl (2001), “PolyLens: a recommender
system for groups of users,” in ECSCW 2001, Springer, 199–218.
O’Hara, Kenton, Matthew Lipson, Marcel Jansen, Axel Unger, Huw Jeffries, and Peter Macer (2004),
“Jukola: democratic music choice in a public space,” in Proceedings of the 5th conference on
Designing interactive systems: processes, practices, methods, and techniques, ACM, 145–54.
Padmanabhan, Vineet, Siva Krishna Seemala, and Wilson Naik Bhukya (2011), “A rule based approach
to group recommender systems,” in Multi-disciplinary Trends in Artificial Intelligence, Springer,
26–37.
Park, Seung-Taek and Wei Chu (2009), “Pairwise preference regression for cold-start recommendation,”
in Proceedings of the third ACM conference on Recommender systems, ACM, 21–28.
Popescu, Gheorghe and EDIC Candidacy Exam (2011), “Preference Aggregation in Group and Social
Recommender Systems,” signature.
Pujahari, Abinash and Vineet Padmanabhan (2015), “A New Grouping Method Based on Social Choice
Strategies for Group Recommender System,” in Computational Intelligence in Data Mining-
Volume 1, Springer, 325–32.
Rashid, Al Mamunur, George Karypis, and John Riedl (2008), “Learning preferences of new users in
recommender systems: an information theoretic approach,” Acm Sigkdd Explorations Newsletter,
10 (2), 90–100.
Resnick, Paul and Hal R Varian (1997), “Recommender systems,” Communications of the ACM, 40 (3),
56–58.
Ricci, Francesco, Lior Rokach, and Bracha Shapira (2011), Introduction to recommender systems
handbook, Springer.
Riedl, J and J Konstan (2015), “MovieLens Latest Datasets (2015),” (accessed April 19, 2018), [available
at https://grouplens.org/datasets/movielens/latest/].
Schafer, J Ben, Joseph Konstan, and John Riedl (1999), “Recommender systems in e-commerce,” in
Proceedings of the 1st ACM conference on Electronic commerce, ACM, 158–66.
Senot, Christophe, Dimitre Kostadinov, Makram Bouzid, Jérôme Picault, Armen Aghasaryan, and Cédric
Bernier (2010), “Analysis of strategies for building group profiles,” in International Conference
on User Modeling, Adaptation, and Personalization, Springer, 40–51.
Smeaton, Alan F and Jamie Callan (2005), “Personalisation and recommender systems in digital
libraries,” International Journal on digital libraries, 5 (4), 299–308.
Sprague, David, Fuqu Wu, and Melanie Tory (2008), “Music selection using the PartyVote democratic
jukebox,” in Proceedings of the working conference on Advanced visual interfaces, ACM, 433–
36.
TMDb (2018), “TMDb About,” (accessed May 11, 2018), [available at
https://www.themoviedb.org/about?language=en].
Vozalis, Emmanouil and Konstantinos G. Margaritis (2003), “Analysis of recommender systems
algorithms,” in The 6th Hellenic European Conference on Computer Mathematics & its
Applications, 732–45.
Yu, Zhiwen, Xingshe Zhou, Yanbin Hao, and Jianhua Gu (2006), “TV program recommendation for
multiple viewers based on user profile merging,” User modeling and user-adapted interaction,
16 (1), 63–82.
Authors
John K. Leung is a Ph.D. candidate in Computational and Data Sciences Department,
Computational Sciences and Informatics at George Mason University in Fairfax,
Virginia. He has over twenty years of working experience in information technology
research and development capacity. Formerly, he worked in the T. J. Watson
Research Center at IBM Corp. in Hawthorne, New York. John has spent more than a
decade working in Greater China, leading technology incubation, transfer, and new
business development.
Igor Griva is an Associate Professor in the Department of Mathematical Sciences at
George Mason University. His research focuses on the theory and methods of
nonlinear optimization and their application to problems in science and engineering.
William G. Kennedy, PhD, Captain, USN (Ret.) is an Associate Professor in the
Department of Computational and Data Sciences and is a Co-Director of the Center
for Social Complexity at George Mason University in Fairfax, Virginia. He has over
10-years’ experience in leading research projects in computational social science with
characterizing the reaction of the population of a mega-city to a nuclear WMD event
being his most recent project. His teaching, research, and publication activities are in
modeling cognition and behavior from individuals to societies.
