0% found this document useful (0 votes)
407 views13 pages

Personality-Aware Product Recommendation System Based On User Interests Mining and Metapath Discovery

Uploaded by

sinduja
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
407 views13 pages

Personality-Aware Product Recommendation System Based On User Interests Mining and Metapath Discovery

Uploaded by

sinduja
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 13

86 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS, VOL. 8, NO.

1, FEBRUARY 2021

Personality-Aware Product Recommendation


System Based on User Interests Mining
and Metapath Discovery
Sahraoui Dhelim , Member, IEEE, Huansheng Ning , Senior Member, IEEE, Nyothiri Aung,
Runhe Huang , Senior Member, IEEE, and Jianhua Ma, Member, IEEE

Abstract— A recommendation system is an integral part of any


modern online shopping or social network platform. The product
recommendation system as a typical example of the legacy
recommendation systems suffers from two major drawbacks: rec-
ommendation redundancy and unpredictability concerning new
items (cold start). These limitations take place because the legacy
recommendation systems rely only on the user’s previous buying
behavior to recommend new items. Incorporating the user’s
social features, such as personality traits and topical interest,
might help alleviate the cold start and remove recommendation
redundancy. Therefore, in this article, we propose Meta-Interest,
a personality-aware product recommendation system based on
user interest mining and metapath discovery. Meta-Interest
predicts the user’s interest and the items associated with these
interests, even if the user’s history does not contain these items
or similar ones. This is done by analyzing the user’s topical
interests and, eventually, recommending the items associated with
Fig. 1. Collabortive filtering and content filtering.
the user’s interest. The proposed system is personality-aware
from two aspects; it incorporates the user’s personality traits online store is measured by their ability to match the right
to predict his/her topics of interest and to match the user’s user with the right product; here comes the usefulness of
personality facets with the associated items. The proposed system
was compared against recent recommendation methods, such as product recommendation systems. Generally speaking, product
deep-learning-based recommendation system and session-based recommendation systems are divided into two main classes.
recommendation systems. Experimental results show that the 1) Collaborative Filtering (CF): CF systems recommend
proposed method can increase the precision and recall of the new products to a given user based on his/her previous
recommendation system, especially in cold-start settings. (rating/viewing/buying) history and his/her neighbors
Index Terms— Big-five model, personality computing, product (similar users). For example, as shown in Fig. 1(a),
recommendation, recommendation system, social networks, social most of the people previously bought a football jersey,
computing, user interest mining, user modeling. and they have also bought a football; thus, the system
I. I NTRODUCTION predicates that the user might be interested in buying a
football.
W ITH the widespread of personal mobile devices and the
ubiquitous access to the internet, the global number
of digital buyers is expected to reach 2.14 billion people
2) Content Filtering or Content-Based Filtering (CBF):
CBF systems recommend new items by measuring their
similarity with the previously (rated/viewed/bought)
within the next few years, which accounts for one-fourth of
products. For example, as shown in Fig. 1(b), football is
the world population. With such a huge number of buyers and
recommended because it is semantically similar to the
the wide variety of available products, the efficiency of an
football jersey.
Manuscript received September 24, 2019; revised September 5, 2020 and Far from that, with the popularity of online social networks,
October 22, 2020; accepted November 7, 2020. Date of publication such as Facebook, Twitter, and Instagram, many users use
November 24, 2020; date of current version January 29, 2021. This work social media to express their feeling or opinions about different
was supported in part by the National Natural Science Foundation of China
under Grant 61872038, and in part by the Fundamental Research Funds for the topics or even explicitly expressing their desire to buy a spe-
Central Universities under Grant FRF-BD-18-016A. (Corresponding author: cific product in some cases, which made social media content
Huansheng Ning.) a rich resource to understand the users’ needs and interests [1].
Sahraoui Dhelim, Huansheng Ning, and Nyothiri Aung are with
the School of Computer and Communication Engineering, University On the other hand, the emerging of personality computing [2]
of Science and Technology Beijing, Beijing 100083, China (e-mail: has offered new opportunities to improve the efficiency of user
[email protected]). modeling in general and particularly recommendation systems
Runhe Huang and Jianhua Ma are with the Faculty of Computer and
Information Sciences, Hosei University, Tokyo 184-8584, Japan. by incorporating the user’s personality traits in the recommen-
Digital Object Identifier 10.1109/TCSS.2020.3037040 dation process. In this work, we propose a product recommen-
2329-924X © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 24,2021 at 05:56:37 UTC from IEEE Xplore. Restrictions apply.
DHELIM et al.: PERSONALITY-AWARE PRODUCT RECOMMENDATION SYSTEM 87

In Section IV, we evaluate the proposed system. Finally,


in Section V, we conclude the work and state some of the
future directions.
II. R ELATED W ORKS
In this section, we review the recent advances of
personality-aware recommendation system and interest mining
schemes as well.
A. Personality and Recommendation Systems
Many works have discussed the importance of incorpo-
rating the user’s personality traits in the recommendation
Fig. 2. Interest mining-based product recommendations.
systems. Yang et al. [4] proposed a recommendation system
of computer games to players based on their personality
dation system that predicts the user’s needs and the associated traits. They have applied text mining techniques to measure
items, even if his/her history does not contain these items or the players’ Big-five personality traits and classified a list
similar ones. This is done by analyzing the user’s topical inter- of games according to their matching with each dominant
est and, eventually, recommending the items associated with trait. They have tested their proposed system on 2050 games
the theses interest. The proposed system is personality-aware and 63 players from the Steam gaming network. While
from two aspects; it incorporates the user’s personality traits Wu et al. [5] presented a personality-based greedy reranking
to predict his/her topics of interest and to match the user’s per- algorithm that generates the recommended list, where the
sonality facets with the associated items. As shown in Fig. 2, personality is used to estimate the users’ diversity preferences,
the proposed system is based on a hybrid filtering approach Ning et al. [6] proposed a friend recommendation system
(CF and CBF) and personality-aware interest mining. that incorporates the big-five personality traits model and
Since we have multiple types of nodes (users, items, and hybrid filtering, where the friend recommended process is
topics), the system is modeled as a heterogeneous information based on personality traits and the users’ harmony rating.
network (HIN), which includes multiple types of nodes and Ferwerda et al. [7] studied the relationship between the user’s
links. In our case, product recommendation could be formu- personality traits and music genre preferences; they have
lated as link prediction in HIN [3]. For example, in Fig. 2, analyzed a data set that contains personality test scores and
given the user’s previous rating and topical interest represented music listening histories of 1415 Last.fm users. Similarly,
in an HIN, the problem is to predict whether or not a link in [8], they conducted an online user survey where the
exists between the user and the product (the ball). One of the participants were asked to interact with an application named
main challenges of link prediction in HIN is how to maintain Tune-A-Find and measured taxonomy choice (i.e., activity,
a reasonable balance between the size of information con- mood, or genre), individual differences (e.g., music expertise
sidered to make the prediction and the algorithm complexity factors and personality traits), and different user experience
of the techniques required to collect that information. Since, factors. Similarly, Hafshejani et al. [9] proposed a CF system
in practice, the networks are usually composed out of hundreds that clusters the users based on their big-five personality traits
of thousands or even millions of nodes, the method used using the K-means algorithm. Following that, the unknown
to perform link prediction in HIN must be highly efficient. ratings of the sparse user–item matrix are estimated based on
However, computing only local information could lead to poor the clustered users. Dhelim et al. [10] discussed the benefits
predictions, especially in very sparse networks. Therefore, of capturing the user’s social feature, such as personality traits
in our approach, we make use of metapaths that start from that are represented as cyberentities in cyberspace. Similarly,
user nodes and end up in the predicted node (product nodes Khelloufi et al. [11] showed the advantages of leveraging the
in our case), and we try to fuse the information from these user’s social features in the context of service recommendation
metapaths to make the prediction. in the Social Internet of Things (SIoT).
The contributions of this work are summarized as follows. B. Interest Mining
1) Propose a product recommendation system that infers Far from personality, many previous works have discussed
the user’s needs based on his/her topical interests. user interest mining from social media content. Piao et al.
2) The proposed system incorporates the user’s big-five [1] surveyed the literature of user interest mining from social
personality traits to enhance the interest mining process networks, and the authors reviewed all the previous works by
and perform personality-aware product filtering. emphasizing the following on four aspects: 1) data collection;
3) The relationship between the users and products is 2) representation of user interest profiles; 3) construction and
predicted using a graph-based metapath discovery; refinement of user interest profiles; and 4) the evaluation
therefore, the system can predict implicit and explicit measures of the constructed profiles. Zarrinkalam et al. [12]
interests. presented a graph-based link prediction scheme that operates
The remainder of this article is organized as follows. over a representation model built from three categories of
In Section II, we review the related works. In Section III, information: user explicit and implicit contributions to topics,
the system design of the proposed system is presented. relationships between users, and the similarity among topics.

Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 24,2021 at 05:56:37 UTC from IEEE Xplore. Restrictions apply.
88 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS, VOL. 8, NO. 1, FEBRUARY 2021

TABLE I
C OMPARISON W ITH R ELATED W ORKS

Trikha et al. [13] investigated the possibility of predicting the TABLE II


users’ implicit interests based on only topic matching using B IG -F IVE T RAITS AND A SSOCIATED C HARACTERS
frequent pattern mining without considering the semantic
similarities of the topics. While Wang et al. [14] proposed
a regularization framework based on the relation bipartite
graph that can be constructed from any kind of relationship,
they evaluated the proposed system from social networks
that were built from retweeting relationships. Dhelim et al.
[15] discussed the usage of the user’s interests to customize
the services offered by a cyber-enabled smart home. Faralli
et al. [16] proposed Twixonomy, a method for modeling of
Twitter users by a hierarchical representation based on their A. Big-Five Traits
interests. Twiconomy is built by identifying topical friends (a There are many personality theories that have tried to
friend represents an interest instead of a social relationship) explain human personality. The most prominent personality
and associate each of these users with a page on Wikipedia. theory is known as the five-factor model (FFM) or big-five
Dhelim et al. [17] used social media analysis to extract the personality traits. The FFM is based on a common language
user’s topical interest. Kang et al. [18] proposed a user description of personality, which makes it a compatible model
modeling framework that maps the user’ posted content in for computing tasks, such as machine learning personality
social media into the associated category in the news media recognition, natural language analysis, and semantic technolo-
platforms, and based on that, they used Wikipedia as a gies, to name a few. FFM is widely used for different purposes,
knowledge base to construct a rich user profile that represents such as mental disorders diagnosis or job recruitment. The
the user’ interests. Liu et al. [19] introduced iExpand, a new model defines the following five factors: neuroticism, openness
CF recommendation system based on user interest expansion to experience, extraversion, agreeableness, and conscientious-
via personalized ranking. iExpand uses a three-layer, user– ness, often denoted by the acronyms OCEAN or CANOE.
interest–item, representation scheme, which makes the recom- The big-five factors are shown in Table II along with their
mendation more accurate and with less computation cost and related personality facets. Many previous psychological studies
helps the understanding of the interactions among users, items, have proven the relationship between user’s interests and
and user interests. personality traits, such as the relationship between personality
Table I shows a comparison between the proposed system traits and Holland’s big-six domains of vocational interest
and some of the related works presented above. For the conve- (RIASEC) [24] and the relationship between hobby interests
nient manipulation of heterogeneous graphs, some works have and personality traits [25].
used metapaths embedding to represent the network infor- The purpose of Meta-Interest is to recommend the most
mation in lower dimensions, such as metapath2vec [20] and relevant items by detecting the user’s topical interests from its
Shi et al. [21]. However, in highly dynamic graphs, such as the social networking data. Fig. 3 shows the general system frame-
user–topic–product graph in our case, where the graph update work of Meta-Interest. The recommendation process includes
happens very frequently, computing the metapath embedding five steps. Step 1 is the personality traits’ measurement, which
all over again is very expensive in terms of computation. As we can be obtained by asking the user to take a personality
will discuss in the experimental section, our method requires measurement questionnaire or using automatic personality
more computational power to compute the initial metapaths recognition by analyzing the subject’s social network data.
compared with the metapath embedding methods but required The personality measurement phase is the only static part
less computing power for the update operation, which makes of the system, which is because personality traits have been
it favorable for highly dynamic graphs. proven to be relatively stable over time. Step 2 is mining the
III. S YSTEM D ESIGN user’s topical interests, including explicit and implicit interest
In this section, we will present the theoretical framework of minings. Explicit interest mining is performed by analyzing
the proposed system. the text shared by the user in social networks in order to detect

Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 24,2021 at 05:56:37 UTC from IEEE Xplore. Restrictions apply.
DHELIM et al.: PERSONALITY-AWARE PRODUCT RECOMMENDATION SYSTEM 89

TABLE III
N OTATIONS AND S YMBOLS

1) Users’ Representation: As mention earlier, one of the


Fig. 3. Meta-Interest recommendations process. most important aspects of the proposed system is that it
incorporates the user’s personality traits and their related
facets to detect the user’s interest and eventually in product
keywords that reflect its topical interests. Implicit interest recommendations. The users’ graph G U = (Vu , E u ) is
mining involves a more complex analysis of the social network constructed by measuring the similarity between its vertices.
structure and other latent factors that may influence the user’s In this regard, we consider three types of similarities: topic
topical interests. In Step 3, Meta-Interest matches the items interest similarity, product interest similarity, and personality
with the corresponding topics. The matching is in the form traits’ similarity, which we denote as SimT, SimI, and SimP,
of a many-to-many relationship that is to say that a topic respectively. Formally, let U = {u 1 , u 2 , . . . , u n } be the set
might be related to many items. Similarly, an item might be of all users and Pi = {PO , PC , PE , PA , PN } be the big-five
related to more than one topic. In Step 4, the set of most personality trait vector of the user u i ; Ti = {t1 , t2 , . . . , tm } is
similar users (neighbors) to the subject user is determined. the set of topical interest of u i , and Ii = {i 1 , i 2 , . . . , i k } is the
In this context, Meta-Interest uses three similarity measures, set of items that were previously viewed by u i
personality similarity, viewing/buying/rating similarity, and   i  i 
  i px − px p y − p y
common interest similarity. Finally, Step 5 is the item rec- ϕ u x , u y = α   2   2
i px − px i py − py
ommendation phase, and the recommendation is refined by i i
updating the neighbors’ set and the user’s topical interest     
 2|Tx Ty |   2|Ix I y | 
profile and topics–items matching. + (1 − α)      (1)
|T | + |T |   |I | + |I | 
x y x y

B. Notations where px and p y is the average value of the personality traits


vector for user u x and u y , respectively, and pix and piy are
The notations and symbols used in the current work are
the i th trait in the personality traits vector of user u x and
explained in Table III.
u y respectively. α is the user similarity weight parameter that
tunes the contribution of item-topic similarity and personality
C. Representational Model similarity in the total similarity measure.
Let U = {u 1 , u 2 , . . . , u n } be the set of users, T = 2) Topics Representation: The interests of a given user are
{t1 , t2 , . . . , tm } the set of topics, and P = { p1 , p2 , . . . , pk } represented in the form of a set of topics. The topic space is
the set of all items. The system is modeled as a heterogenous represented by the graph G T = (Vt , E t ), where the vertices
graph that consists of three subgraphs G = (G U , G T , G P ), represent the topics and the edges represent the semantic
as shown in Fig. 4. G U = (Vu , E u ) is undirected graph similarity relationship between these topics. To associate these
where its node set Vu is the users set U , and the edges topics with items graph nodes, each topic node is associ-
set E u represents the similarity relationship between users. ated with a category of open directory project (ODP) [26]
In addition to online behaviors similarity, such as posting and (see Fig. 5). ODP is a public open directory for web sites’ clas-
follower/followee similarities, the personality traits’ similarity sifications. Currently, it contains 3.8 million websites that have
between users is also considered to compute the overall been categorized into 1 031 722 categories by 91 929 human
similarity between users. Similarly, the graphs G T = (Vt , E t ) editors. We have used the four-level subcategories to construct
and G P = (V p , E p ) represent the nodes and relationship the topics graph; these categories are used to match the interest
between topics and items, respectively. topics with the related items from the item graph.

Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 24,2021 at 05:56:37 UTC from IEEE Xplore. Restrictions apply.
90 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS, VOL. 8, NO. 1, FEBRUARY 2021

Fig. 4. User–topic–item heterogeneous information network.

Algorithm 1 Interest_mining
Input u x , sx , Fx Output Ix
1: if (sx > C S) then
2: Semantic_Annotation(sx )
3: Topics_Extraction(sx )
4: else
5: for f ∈ Fx do
6: Ix ← Ix ∪ {Per sonali t y_ f acet_topi cs( f )}
7: end for
8: end if

similarity measure, β = 0, when the item has no views and


never been bought before (item cold start)
     
   2|C x C y |   
ϑ Px , Py = β   + (1 − β)  2|Vx Vy |  .
 |C | + |C |   |V | + |V | 
x y x y
Fig. 5. OPD root categories.
(2)
3) Item Representation: Similar to the users and interest
topics, the items are represented as a graph data structure D. Interest Mining
G P = (V p , E p ), where the nodes represent the items and the The main advantage of our approach is that the proposed
edges represent the similarity between the items. The similarity system makes use of the user’s interests along with the
between items is computed from two similarity measures, user’s personality information to optimize the accuracy of
content similarity and collaborative similarity. The content system recommendations and alleviate the cold-start effects.
similarity is measured by common item’s metadata tags, while By analyzing the user’s social network posted data, we can
the collaborative similarity is calculated by measuring the ratio infer his/her topical interests. The task can be achieved by
of common buyers/viewers between the two items to the total applying automatic topic extraction techniques, such as latent
buyers/viewers of each item. Formally, let C x : {c0 , c1 , . . . , cn } Dirichlet allocation (LDA) [27] or frequency-inverse category
and C y : {c0 , c1 , . . . , cm } denote the content tags of item Px frequency (TFICF) [28]. However, such techniques are sup-
and Py , respectively, and Vx and Vy represent the sets of their posed to be applied to long articles, and they do not yield
viewing/buying users. The similarly between Px and Py is good results if applied on the user’s short sparse noisy posts,
computed using the function ϑ, as shown in (2), where β is the such as tweets [29]. Therefore, to overcome this problem,
item similarity threshold, and it is used to tune the contribution we have enriched each post from the user’s data using semantic
of content similarity and collaborative similarity to the overall annotators, which could help to reduce the noise and alleviate

Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 24,2021 at 05:56:37 UTC from IEEE Xplore. Restrictions apply.
DHELIM et al.: PERSONALITY-AWARE PRODUCT RECOMMENDATION SYSTEM 91

Algorithm 2 Item_mapping
Input pz , U pz
Output I pz
1: if (vi ews( pz ) > C S) then
2: I pz ← O P D_T opi cs( pz )
3: else
4: for f ∈ Fx and u x ∈ U pz do
|U | Fig. 6. Network scheme and length 2 metapath samples.
5: if (|u y , f ∈ Fy | > 2pz ) then
6: I pz ← I pz ∪ {Per sonali t y_ f acet_topi cs( f )}
metapaths [21] to predict the matching score between a given
7: end if
user node in G U and an item node in G P .
8: end for
A metapath is a sequence of relations between nodes defined
9: end if
over a heterogeneous network, which can be used to define
a topological structure with various semantics. In our case,
we investigate the metapaths that start from a user node and
ambiguity of the post and increase the topic detection accuracy,
end with an item node P : {u → x → · · · → x → i }.
as shown in the proposed framework in [18]. Algorithm 1
Each metapath is characterized by the number of links between
shows the pseudocode of interest mining steps. When the user
the source and destination nodes, and it is called the path
is during the cold-start phase or completely did not view any
length Pl . For example, the possible metapath with path length
articles (lines 1–4), Meta-Interest estimates the topical interest
P2 from a user node to an item node is presented in Fig. 6.
based on the interests of users with similar personality facets.
For a given metapath P : {s → x → · · · → x → d}, any
Otherwise, it crawls the viewed news articles and extracts the
path in the network that connects nodes s and d following
labels of each news article to serve as the topical interest of
the same intermediate node types as defined by P is called a
the user, as we will see in the experimental section.
path instance of P. For a given metapath P, the path count is
the number of all path instances Pc = |{ p : p ∈ P}|. In our
E. Item Mapping
case, we consider all metapaths that start with a user node
After populating the topics public space using ODP ontol- and end with an item node with maximum metapath length to
ogy categories, the items are matched with these topics. Each lmax = 2. We have made the maximum length to 3 because
item is associated with one or more topics and, subsequently, short metapaths are semantically more important than long
recommended for users that have these topics within their ones, and they are good enough for capturing the structure
topical interests. Algorithm 2 shows the pseudocode of the of the network. Besides that, it is computationally expensive
item interest mapping process. With newly added items that to explore longer metapath because the path count increases
have not been viewed by any user, the item is directly exponentially with the increase in the path length Pl [33].
associated with the corresponding topic category in ODP By exploring all metapaths with length constraint, we could
ontology, whereas items that have passed the cold-start phase holistically extract all relationships between nodes with dif-
are associated with the interest of those that are related to the ferent filtering combinations. We are interested in three types
personality facets that are shared among the users who bought of metapaths: first, the interest metapaths (IP) of the format
this item. U-T-P [see Fig. 6(b)] that represents metapaths that are based
on interest mining and item matching; second, the friendship
F. Metapath Discovery metapaths (FP) of the format U-U-P [see Fig. 6(c)] that
After building the users–topics–items heterogeneous graph represents metapaths that are based on CF (users’ similarity);
G = (G U , G T , G P ) that incorporates the users, topics, and and finally, content metapaths (CP) of the format U-P-P
items subgraphs and their interrelationships. At this stage, [see Fig. 6(d)] that represents metapaths that are based on
the objective is to predict for a given user the N-most recom- the content filtering (items similarity). Similarly, by exploring
mended items that match his/her topical interests and previous longer metapath, we get more hybrid filtering paths (based
buying/viewing behaviors. Predicting the users’ recommended on both CF and CBF, in addition to interest mining and
items is formulated as a graph-based link prediction problem. item mapping); for example, metapaths of length Pl = 3
Link prediction problem has been investigated in many works could be based on CBF (i.e., U-P-P-P), CF (i.e., U-U-U-P),
before, and many schemes have been proven to achieve high hybrid filtering (i.e., U-U-P-P), or a combination of fil-
accuracy in their predictions, such as Adamic/Adar [30], Katz tering and interest mining (i.e., U-T-T-P, U-T-P-P, and
[31], and Jaccard [32]. However, these schemes are supposed U-U-T-P).
to work on homogeneous graphs where all nodes represent the The importance of each metapath is characterized by its
same type of entities and all the edges are connecting these weight w p . The path weight is computed by the sum of
entities, which is not the case with our heterogeneous graph. its edges’ weight over its length Pl . Formally, let P n :
Since, in our representation model G = (G U , G T , G P ), nodes {v 1 , v 1 , . . . , v n } be a metapath with a length of Pl = n, and
can represent different entities (users, topics, and items) and the path weight of P is denoted as w p , which is the sum
the links can connect different nodes (user–user, user–topic, of all the links’ weights within P, as shown in (3), where
user–item, topic–item, item–item, and topic–topic). We use wvi ,vi+1 represents the weight of link that connects the nodes

Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 24,2021 at 05:56:37 UTC from IEEE Xplore. Restrictions apply.
92 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS, VOL. 8, NO. 1, FEBRUARY 2021

Algorithm 3 DiscoverMetaPaths The pseudocode shown in Algorithm 3 presents the steps of


Input u s , lmax , ε metapath discovery. The algorithm takes as input the user
Output F N L source node u s , the maximum metapath length to explore lmax ,
1: V I ST ← Ø
and the link prediction score threshold ε. We denote P as the
2: P ← Ø
set of the temporarily explored metapaths, P is updated by
3: F N L ← Ø
adding new explored paths or removing dead paths (paths that
4: for i = 1 to lmax do
have no neighbors or paths that do not end with an item node),
5: if (i = 1) then
and FNL is the set of the final metapaths. The set of visited
6: V I ST ← V I ST ∪ {u s } nodes is denoted as VIST, and v denotes the set of neighbors
7: for N G B ∈ u s do of node v. NODE and CURN are temporary variables used to
8: P ← P ∪ {u s → N G B} denote the current node and current path respectively in each
9: V I ST ← V I ST ∪ {N G B} iteration. In Lines 5–11, a path from the source node u s to
10: end for every neighbor node is created and inserted into the set of
11: else
metapaths P, and node u s and its neighbors are marked as
12: T EMP ← Ø visited nodes VIST. In lines 13—25, for each path CURN
13: for CU R N ∈ P do from P, the last node of these paths is visited and added to
14: N O D E ← pc [i ] the final metapaths if it is an item node, and recursively, all
15: if (N O D E = i tem) and (w pc > ε) then the nodes that have not been visited before are added as a
16: F N L ← F N L ∪ { pc } potential metapaths. Algorithm 4 shows the pseudocode of
17: end if recommendation process. Initially, if the user is still in the
18: if ( N O D E−V I ST = Ø) then cold-start phase (lines 2–7), the recommended items are to
19: for N G B ∈  N O D E−V I ST do be filtered based on the topical interests that were extracted
20: T E M P ← T E M P ∪ {CU R N → N G B} from the user’s social media data and by associating these
21: V I ST ← V I ST ∪ {N G B} topics with the related items according to their OPD categories.
22: end for Otherwise, the metapaths starting from the source user u s
23: end if are discovered and grouped according to the metapath types
24: P ← P − CU R N (interest metapath, friendship metapath, and content metapath),
25: end for and the items that are in the intersection of these metapaths
26: P ← T EMP sets are given propriety in the recommended items’ set.
27: end if
Lines 7–10 enumerate all the neighbors of the source
28: end for
node and lines 13–25 (and eventually lines 19–22) are the
primary computational blocks in Algorithm 3. If we study
the worst case graph structure, which is a complete graph
(fully connected graph), where every user is similar to all other
v i and v i+1
n users and interested in all topics and also connected to all the
wvi ,vi+1 available products (even though, in this case, we do not have
wp = i=1
. (3)
Pl to run Algorithm 3, there is no unknown link that we need to
In order to predict a possible recommendation for a given predict). Algorithm 3 still runs in linear time complexity. Let
user node, we explore all the instances of metapath with a G be a complete graph (fully connected) with n nodes and
maximum path length lmax = 3. Because short metapaths n = x + y + z (x: user nodes; y: topic nodes; and z: product
are more semantically significant compared with longer meta- node). The run time of the block (lines 7–10) is O(x +y+z−1)
paths. Therefore, we prioritize shorter metapath by consid- to add all the graph nodes to the visited nodes group VIST
ering that the contribution of a path weight to the overall and their generated paths to P. The block (12–25) also runs
link prediction score is inversely proportional to the meta- in linear time of O(x + y + z − 1) as well; Even, it includes
path length Pl . The link prediction score between user u i a nested loop (lines 19–22) that could result a quadratic time,
and item p j with the metapath maximum link constrain as lines 19–22 will never be reached due to the if-condition block
lmax = l is computed using (4). To predict the N-most in lines 18–23 (as the studied graph is a complete graph,
recommended items for a given user, we extract all metapaths and VIST will contain all the graph nodes [added in block
by exploring the interest graph with a fixed length and link (7–10)]; therefore, NODE − VIST = ∅). Hence, the overall
prediction score constraints time complexity of Algorithm 3 is O(n).
 IV. S YSTEM E VALUATION

l
r∈Pi, j (k) wr
δi, j =
l
. (4) In this section, we present the details of the collected data
k=2
k −1 set, evaluation metrics and baselines, and the analysis of the
obtained results.
To compute the recommended items for a given user,
we extract all metapaths instances between the user and poten- A. Baselines
tial recommended items by exploring the user–interest–item To evaluate the performance of the proposed product rec-
graph with a fixed length and link prediction score constraints. ommendation system, we have compared it with different

Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 24,2021 at 05:56:37 UTC from IEEE Xplore. Restrictions apply.
DHELIM et al.: PERSONALITY-AWARE PRODUCT RECOMMENDATION SYSTEM 93

Algorithm 4 RecommendProducts influence. DGRec is open source, and its implementation code
Input u s , ls is available on Github.3
Output R 4) LightFM (Cold Start) [34]: LightFM is a cold-start
1: R ← Ø alleviation framework that uses a hybrid matrix factorization
2: if (C S(u s )) then model to represent items (products in our case) and users as
3: for t ∈ Is do linear combinations of their content features’ latent factors.
4: P R ← Pr oduct_i nter est (t) LightFM is parameterized in terms of d-dimensional user and
5: R ← R ∪ PR item feature embeddings e U f and e I f for each feature f.
6: end for Every feature is also modeled by a scalar bias term (b U f
7: else for user and b I f for item features). The model’s prediction
8: P = Di scover Meta Path(u s ) for user u and item i is then given by the dot product of user
9: I P = I nter est Paths(P) and item representations, adjusted by the user and item feature
10: F P = Fri end Paths(P) biases. LightFM is open source, and the implementation code
11: C P = Content Paths(P) is available on GitHub.4
12: Rec Paths = T op N Paths(I P ∩ F P ∩ C P, F P ∩ 5) CF-CBF: This is the hybrid filtering system that com-
C P, C P ∩ I P) bines the users’ viewing similarity and product similarity to
13: for Path ∈ Rec Paths do determine the neighborhood set and recommends new items.
14: P R ← Path[lastn ode]
15: R ← R ∪ PR B. Evaluation Metrics
16: end for Any product recommendation system is evaluated by mea-
17: end if suring the accuracy and coverage of its recommended items.
To test the efficiency of Meta-Interest and compare it to the
afore-mentioned baselines, we determine the recommended
items by each baseline, display it in the user’s feed along
baselines that use various recommendation techniques, such
with other irrelevant items, and measure the accuracy rate of
as deep learning, metapath analysis, network embedding, and
the relevant items. Formally, let F = R∪I be the set that repre-
session-based. The proposed system is compared with the
sents all items in user u’s feeds, where R = { p1 , p2 , . . . , pr }
following baselines:
is the set relevant items, and I = { p1 , p2 , . . . , pi } is the
1) GNN-SEAL (Graph Neural Networks) [22]: GNN-
set of irrelevant items. After showing F in user u’s feeds,
SEAL is a link prediction framework that formulates link
we denote V = { p1 , p2 , . . . , pv } as the set of viewed items.
prediction problem as a subgraph classification problem. For
In this context, we are interested in the following values:
every predicted link (user–item link in our case), GNN-SEAL
1) true positives: the group of relevant  items that have been
determines its h-hop enclosing subgraph A and computes
viewed by the user TP = {x / x ∈ R V }; 2) false positives:
its node information matrix X (which contains structural
the group of irrelevant items
 that have been viewed by the
labels, latent embeddings, and the explicit attributes of
user FP = {x / x ∈ I V }; and 3) false negatives: the
nodes). After that, the framework feeds (A, X) into a graph
group of relevant items that have not been viewed by the user
neural network (GNN) to classify the link existence so
FN = {x / x ∈ R, x ∈ / V }. We have used the following
that it can learn from both graph structure features (from
metrics.
A) and latent/explicit features (from X) simultaneously for
Precision: The portion of relevant viewed items in the total
link prediction. The framework is open source, and the code
viewed items, and it is computed as follows:
is available on GitHub.1
2) metapath2vec (Metapath and Network Embedding) [20]: TP
Precision = . (5)
metapath2vec formalizes metapath-based random walks to TP + FP
build the heterogeneous neighborhood of a node and then
Recall: The portion of relevant viewed items in the total
uses the heterogeneous skip-gram model to perform node
relevant items, and it is computed as follows:
embeddings and, subsequently, user–item link prediction.
metapath2vec is open source, and its implementation code is TP
Recall = . (6)
available on Github.2 TP + FN
3) DGRec (Session-Based) [23]: DGRec is a session-based
F-Measure: It is also called the balanced F-Score; it is
recommendations’ framework that employs dynamic-graph-
the harmonic average of the precision and recall; and it is
attention neural network to model the context-dependent social
computed as follows:
influence and recurrent neural network to model dynamic
user interest. Finally, DGRec gives the recommendation by 2 P R
F= . (7)
merging the user’s interests and preferences and his/her social P+R

3 github.com/DeepGraphLearning/RecommenderSystems/tree/master/
1 github.com/muhanzhang/SEAL socialRec
2 ericdongyx.github.io/metapath2vec/m2v.html 4 github.com/lyst/lightfm

Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 24,2021 at 05:56:37 UTC from IEEE Xplore. Restrictions apply.
94 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS, VOL. 8, NO. 1, FEBRUARY 2021

TABLE IV
D ATA S ET S TATISTICS

C. Data Set Description


We have integrated the Meta-Interest product recom-
mendation system with a social network platform called
Newsfulness5 that we have implemented earlier for automatic
personality recognition projects. Newsfulness enables the user
to view and shares news articles from various news publishers.
During registration, the users go through the TIPI Big-Five
personality questionnaire [35] to capture their personality Fig. 7. Users’ similarity parameter tuning.
traits. Newsfullness collects published articles from different
English-speaking news websites, and the collected articles are
from the following outlets (BBC, CNN, Aljazeera, France24,
Russia-Today, Reuters, The Guardian, and The New York
Times). The gathered articles are from all the news classes
(politics, business, sports, health, travel, education, entertain-
ment, art, science, and technology) from different geographic
regions categories. The products’ recommendation system was
implemented by fetching products from different online stores
(mainly JD, Banggood, and Amazon). The statistical details
of the used data set are presented in Table IV.

D. Result Discussion
To tune the optimal value of the users’ similarity parame-
ter α and products’ similarity parameter β, we observe the
optimal value of α and β that maximize the F-Measure of Fig. 8. Products’ similarity parameter tuning.
the proposed system. Figs. 7 and 8 show the optimal value
of α and β in different topics of interest count and viewed
items count, respectively. As we can observe from Fig. 7,
during the cold-start phase with no topic of interest at all,
α = 1, and at this point, the users’ similarity is based only
on personality similarity measurement. With the increase in
previously detected topics of interest, the value of α gradually
decreases and finally stabilizes with α = 0.2 when the user
passes the cold-start phase and had enough topical interest
and previously viewed items. Similarly, the optimal value of
β during the cold-start phase for the new item with no views
is β = 1, and with the increase in the number of views,
the value of β decreases to finally stabilize with β = 0.5,
as shown in Fig. 8. For the size of Top-n recommended Fig. 9. Top-n recommendation parameter tuning.
products, in our experiment, we set N = 20, as choosing
a larger value will lead to uncertainty of whether the users did
not view the products’ feed because they are not interested in F-Measure of the proposed system and all the studied baselines
them, or they did not view them because there are too many have decreased dramatically when the value of N is over 22.
items in the products’ feed. If we ignore this uncertainty and The precision, recall, and F-measure of Meta-Interest com-
just consider that the user did not view the product out of pared with the baseline schemes are shown in Fig. 10.
his disinterest, this will lead to an increase in false positives As we can observe, the proposed system, Meta-Interest, and
and false negatives as well, hence the decrease in the overall the session-based system, DGRec, clearly have the highest
system performance. As we can observe from Fig. 9, the precision (0.854 and 0.845) and recall (0.868 and 0.855),
respectively. The superiority of the proposed system is because
5 www.newsfullness.live/data set of the personality biased approach that filters the relevant

Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 24,2021 at 05:56:37 UTC from IEEE Xplore. Restrictions apply.
DHELIM et al.: PERSONALITY-AWARE PRODUCT RECOMMENDATION SYSTEM 95

Fig. 10. Overall system evaluation. Fig. 12. System evaluation under cold start (new items).

2) Cold-start items’ test, in which we consider only the new


items that have not been viewed or rated by any user.
Figs. 11 and 12 show the results of the cold-start users’
test and cold-start items’ test, respectively.
As we can observe, the proposed system and the session-based
system, DGRec, still have the upper hand even in the
cold-start phase as both systems are robust in cold-start
settings, as explained early. However, we can notice that the
LightFM is ranked third and obviously outperforms meta-
path2vec and GNN-SEAL because LightFM was originally
designed to mitigate the cold-start effects, and as mentioned
early, LightFM has a poor performance when the amount of
Fig. 11. System evaluation under cold start (new users). the data increases. To further study the relationship between
the amount of available data and the performance of the
items that are related to the personality facets of the user, proposed system compared with the baselines, we measure the
while other approaches view the user’s personality traits just performance of Meta-Interest and the other baseline systems
as additional information that helps find the similarity and while changing the percentage of training set size from 10%
construct the network embeddings or features. The second to 100%. Fig. 13 shows the precision, recall, and F-measure
reason for the superiority of Meta-Interest (and also DGRec values of the studied systems with different training set sizes.
compared with other baselines) is the ability of Meta-Interest We can clearly observe that Meta-Interest outperforms the
and DGRec to alleviate the cold-start effects, and hence other baselines with only a small training set size, with only
maintain stable precision, and recall values all over the phases. 10% of the training set and Meta-Interest scores 0.768 and
Unlike the network representation method, metapath2vec, and 0.765 in precision and recall, respectively. With the increase
the deep-learning method, GNN-SEAL, that come third and in training set size, Meta-Interest steadily improves to reach
fourth with 0.84 and 0.828 of precision value and 0.835 and 0.854 and 0.868 in precision and recall using 100% of the
0.825 of recall value, respectively, LightFM performs quite training set (around 10.07% improvement compared with 10%
well in the cold-start phase (as we will see later in other training). LightFM ranks second in terms of precision and
figures); however, it fails to cope with a large amount of recall with 10% of the training set; however, it ends in
diverse data in later stages, which leads to a drop in its the fifth place (better only than the conventional CF-CBF)
precision and recall values. with full training set 100%, whereas deep-learning-based
One of the main reasons for incorporating the user’s per- schemes (GNN-SEAL) and network embeddings approach
sonality in the product recommendation systems and interest (metapath2vec) have a low performance with small training
mining schemes is to alleviate the effects of the cold-start data. When trained with 10% of the data set, GNN-SEAL
problem [6]. In this regard, we have tested the performance scores 0.63 and 0.64 in precision and recall, respectively. How-
of Meta-Interest and the studied baselines under the cold-start ever, GNN-SEAL and metapath2vec performance increase
settings. The cold-start settings include two tests. dramatically with the increase in the training data size. For
1) Cold-start users’ test, in which only the new users are instance, when trained with the full training data, meta-
considered in the precision and recall measurements. path2vec scores 0.84 and 0.835 (around 23% improvement
In our experiment, a user is considered in the cold-start compared with 10% training). That is because metapath2vec
phase if the number of viewed articles and items is less uses network embedding, which requires the presence of dense
than 20. node links to capture the network structure.

Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 24,2021 at 05:56:37 UTC from IEEE Xplore. Restrictions apply.
96 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS, VOL. 8, NO. 1, FEBRUARY 2021

Fig. 13. System evaluation with different sizes of the training set.

TABLE V metapath discovery, and the system predicts the user’s needs
S PEED C OMPARISON (s) and the associated items. Products’ recommendation is com-
puted by analyzing the user’s topical interest and, eventually,
recommending the items associated with those interests. The
proposed system is personality-aware from two aspects: first,
because it incorporates the user’s personality traits to predict
his topics of interest; second, it matches the user’s personality
facets with the associated items. Experimental results show
that the proposed system outperforms the state-of-art schemes
in terms of precision and recall especially in the cold-start
In a practical situation with a large graph of millions of phase for new items and users.
nodes and links that require intensive computational power, However, Meta-Interest could be improved in different
the speed of the recommendation system is crucial to keep aspects.
a reasonable response time. Therefore, it is important to 1) In this work, the users’ personality traits’ measurement
analyze the speed and time complexity of the proposed system was conducted through questionnaires. Integrating an
compared with the compared baselines. Table V shows the automatic personality recognition system, which can
time complexity of the proposed system compared with the detect the users’ personality traits based on their
studied baselines. The shown values in Table V are the average shared data, into Meta-Interest is one of our future
of 100 times testing. The time complexity of Meta-Interest and directions.
all the baselines were tested on Dell Inspiron 173000 Laptop, 2) The proposed system uses big-five to model the user’s
with tenth-generation Intel Core i7-1065G7 Processor (8- personality. Extending Meta-Interest to include other
MB Cache, up to 3.9 GHz) and 16-GB RAM (2 × 8 GB, personality traits models, such as the Myers–Briggs type
DDR4, 2666 MHz), running Ubuntu 19.04 operating system. indicator, is a future direction.
As we can observe from Table V, when it comes to the 3) The proposed system could be further improved by
total computational time required for the system to compute integrating a knowledge graph and infer topic–item
the recommendation for all users, Meta-Interest is not the association using semantic reasoning.
fastest system, metapath2vec has the lowest computational
time of 14.6 s, and GNN-SEAL ranks second with 15.5. ACKNOWLEDGMENT
However, for the update operation where we add a new block The authors would like to thank all the active users of
of users and items and compute the time required for the Newsfullness that agreed to be a part of the Meta-Interest
system to compute the recommendation of these new users, experiment.
metapath2vec needs to recalculate the network embeddings in
lower dimensional space all over again, which costs as high as R EFERENCES
the initial time required to compute all the recommendations, [1] G. Piao and J. G. Breslin, “Inferring user interests in microblog-
which is not the case with Meta-Interest, as we just need to ging social networks: A survey,” User Model. User-Adapted Interact.,
recalculate the weights of the newly added metapaths. vol. 28, no. 3, pp. 277–329, Aug. 2018. [Online]. Available: http://link.
springer.com/10.1007/s11257-018-9207-8
V. C ONCLUSION [2] A. Vinciarelli and G. Mohammadi, “A survey of personal-
ity computing,” IEEE Trans. Affect. Comput., vol. 5, no. 3,
In this article, we have proposed a personality-aware pp. 273–291, Jul. 2014. [Online]. Available: http://ieeexplore.ieee.org/
product recommendation system based on interest mining and document/6834774/

Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 24,2021 at 05:56:37 UTC from IEEE Xplore. Restrictions apply.
DHELIM et al.: PERSONALITY-AWARE PRODUCT RECOMMENDATION SYSTEM 97

[3] V. Martínez, F. Berzal, and J.-C. Cubero, “A survey of link prediction [21] C. Shi, B. Hu, W. X. Zhao, and P. S. Yu, “Heterogeneous information
in complex networks,” ACM Comput. Surv., vol. 49, no. 4, pp. 1–33, network embedding for recommendation,” IEEE Trans. Knowl. Data
Feb. 2017. Eng., vol. 31, no. 2, pp. 357–370, Feb. 2019. [Online]. Available:
[4] H.-C. Yang and Z.-R. Huang, “Mining personality traits from https://ieeexplore.ieee.org/document/8355676/
social messages for game recommender systems,” Knowl.-Based [22] M. Zhang and Y. Chen, “Link prediction based on graph
Syst., vol. 165, pp. 157–168, Feb. 2019. [Online]. Available: neural networks,” in Proc. Adv. Neural Inf. Process. Syst., 2018,
https://linkinghub.elsevier.com/retrieve/pii/S095070511830577X pp. 5165–5175.
[5] W. Wu, L. Chen, and Y. Zhao, “Personalizing recommendation diversity [23] W. Song, Z. Xiao, Y. Wang, L. Charlin, M. Zhang, and J. Tang,
based on user personality,” User Model. User-Adapted Interact., vol. 28, “Session-based social recommendation via dynamic graph attention
no. 3, pp. 237–276, Aug. 2018. networks,” in Proc. 12th ACM Int. Conf. Web Search Data Mining.
[6] H. Ning, S. Dhelim, and N. Aung, “PersoNet: Friend recommen- New York, NY, USA: ACM, Jan. 2019, pp. 555–563. [Online]. Avail-
dation system based on big-five personality traits and hybrid filter- able: https://dl.acm.org/doi/10.1145/3289600.3290989
ing,” IEEE Trans. Comput. Social Syst., vol. 6, no. 3, pp. 394–402, [24] P. I. Armstrong and S. F. Anthoney, “Personality facets and RIASEC
Jun. 2019. [Online]. Available: https://ieeexplore.ieee.org/document/ interests: An integrated model,” J. Vocational Behav., vol. 75, no. 3,
8675299/ pp. 346–359, Dec. 2009. [Online]. Available: https://linkinghub.elsevier.
[7] B. Ferwerda, M. Tkalcic, and M. Schedl, “Personality traits and com/retrieve/pii/S0001879109000657
music genres: What do people prefer to listen to?” in Proc. [25] U. Wolfradt and J. E. Pretz, “Individual differences in creativity:
25th Conf. User Model., Adaptation Personalization, Jul. 2017, Personality, story writing, and hobbies,” Eur. J. Personality, vol. 15,
pp. 285–288. no. 4, pp. 297–310, 2001.
[8] B. Ferwerda, E. Yang, M. Schedl, and M. Tkalcic, “Personality and [26] J.-H. Lee, J. Ha, J.-Y. Jung, and S. Lee, “Semantic contextual advertising
taxonomy preferences, and the influence of category choice on the based on the open directory project,” ACM Trans. Web, vol. 7, no. 4,
user experience for music streaming services,” Multimedia Tools Appl., p. 24, 2013.
vol. 78, no. 14, pp. 20157–20190, 2019. [27] D. M. Blei, A. Y. Ng, and M. I. Jordan, “Latent Dirichlet allocation,”
[9] Z. Y. Hafshejani, M. Kaedi, and A. Fatemi, “Improving sparsity and new J. Mach. Learn. Res., vol. 3, pp. 993–1022, Mar. 2003.
user problems in collaborative filtering by clustering the personality [28] J. Han and H. Lee, “Characterizing user interest using heteroge-
factors,” Electron. Commerce Res., vol. 18, no. 4, pp. 813–836, neous media,” in Proc. 23rd Int. Conf. World Wide Web, 2014,
Dec. 2018. [Online]. Available: http://link.springer.com/10.1007/ pp. 289–290.
s10660-018-9287-x [29] N. Naveed, T. Gottron, J. Kunegis, and A. C. Alhadi, “Searching
[10] S. Dhelim, N. Huansheng, S. Cui, M. Jianhua, R. Huang, and microblogs: Coping with sparsity and document quality,” in Proc. 20th
K. I.-K. Wang, “Cyberentity and its consistency in the cyber-physical- ACM Int. Conf. Inf. Knowl. Manage., 2011, pp. 183–188.
social-thinking hyperspace,” Comput. Electr. Eng., vol. 81, Jan. 2020, [30] L. A. Adamic and E. Adar, “Friends and neighbors on the Web,” Soc.
Art. no. 106506. [Online]. Available: https://linkinghub.elsevier.com/ Netw., vol. 25, no. 3, pp. 211–230, 2003.
retrieve/pii/S0045790618334839 [31] L. Katz, “A new status index derived from sociometric analysis,”
[11] A. Khelloufi et al., “A social relationships based service rec- Psychometrika, vol. 18, no. 1, pp. 39–43, 1953.
ommendation system for SIoT devices,” IEEE Internet Things J., [32] S. Dhelim, N. Aung, and H. Ning, “ComPath: User interest
early access, Aug. 14, 2020. [Online]. Available: https://ieeexplore. mining in heterogeneous signed social networks for Internet of
ieee.org/document/9167284/, doi: 10.1109/JIOT.2020.3016659. people,” IEEE Internet Things J., early access, Nov. 10, 2020.
[12] F. Zarrinkalam, M. Kahani, and E. Bagheri, “Mining user interests over [Online]. Available: https://ieeexplore.ieee.org/document/9253614, doi:
active topics on social networks,” Inf. Process. Manage., vol. 54, no. 2, 10.1109/JIOT.2020.3037109.
pp. 339–357, Mar. 2018. [33] Y. Sun, J. Han, X. Yan, P. S. Yu, and T. Wu, “PathSim: Meta
[13] A. K. Trikha, F. Zarrinkalam, and E. Bagheri, “Topic-association min- path-based top-K similarity search in heterogeneous information net-
ing for user interest detection,” in Proc. Eur. Conf. Inf. Retr. Basel, works,” Proc. VLDB Endowment, vol. 4, no. 11, pp. 992–1003,
Switzerland: Springer, 2018, pp. 665–671. Aug. 2011.
[14] J. Wang, W. X. Zhao, Y. He, and X. Li, “Infer user interests via link [34] M. Kula, “Metadata embeddings for user and item cold-start rec-
structure regularization,” ACM Trans. Intell. Syst. Technol., vol. 5, no. 2, ommendations,” in Proc. 2nd Workshop New Trends Content-Based
p. 23, 2014. Recommender Syst. Co-Located 9th ACM Conf. Recommender Syst.
[15] S. Dhelim, H. Ning, M. A. Bouras, and J. Ma, “Cyber-enabled (RecSys), Vienna, Austria, vol. 1448, 2015, pp. 14–21. [Online]. Avail-
human-centric smart home architecture,” in Proc. IEEE SmartWorld, able: http://ceur-ws.org/Vol-1448/paper4.pdf
Ubiquitous Intell. Comput., Adv. Trusted Comput., Scalable Comput. [35] S. D. Gosling, P. J. Rentfrow, and W. B. Swann, “A very brief measure
Commun., Cloud Big Data Comput., Internet People Smart of the Big-Five personality domains,” J. Res. Pers., vol. 37, no. 6,
City Innov. (SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI), pp. 504–528, 2003.
Oct. 2018, pp. 1880–1886. [Online]. Available: https://ieeexplore.ieee.
org/document/8560294/
[16] S. Faralli, G. Stilo, and P. Velardi, “Automatic acquisition of a taxonomy
of microblogs users’ interests,” J. Web Semantics, vol. 45, pp. 23–40,
Aug. 2017.
[17] S. Dhelim, N. Aung, and H. Ning, “Mining user interest based
on personality-aware hybrid filtering in social networks,” Knowl.-
Based Syst., vol. 206, Oct. 2020, Art. no. 106227. [Online]. Available:
https://linkinghub.elsevier.com/retrieve/pii/S0950705120304354
[18] J. Kang and H. Lee, “Modeling user interest in social media
using news media and wikipedia,” Inf. Syst., vol. 65, pp. 52–64, Sahraoui Dhelim (Member, IEEE) received the B.S.
Apr. 2017. degree in computer science from the University of
[19] Q. Liu, E. Chen, H. Xiong, C. H. Q. Ding, and J. Chen, “Enhanc- Djelfa, Djelfa, Algeria, in 2012, the master’s degree
ing collaborative filtering by user interest expansion via personal- in networking and distributed systems from the Uni-
ized ranking,” IEEE Trans. Syst., Man, Cybern. B. Cybern., vol. 42, versity of Laghouat, Laghouat, Algeria, in 2014, and
no. 1, pp. 218–233, Feb. 2012. [Online]. Available: http://ieeexplore. the Ph.D. degree in computer science and technology
ieee.org/document/6006538/ from the University of Science and Technology
[20] Y. Dong, N. V. Chawla, and A. Swami, “Metapath2vec: Scal- Beijing, Beijing, China, in 2020.
able representation learning for heterogeneous networks,” in Proc. His current research interests include social
23rd ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining. computing, personality computing, user model-
New York, NY, USA: ACM, Aug. 2017, pp. 135–144. [Online]. Avail- ing, interest mining, recommendation systems, and
able: https://dl.acm.org/doi/10.1145/3097983.3098036 intelligent transportation systems.

Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 24,2021 at 05:56:37 UTC from IEEE Xplore. Restrictions apply.
98 IEEE TRANSACTIONS ON COMPUTATIONAL SOCIAL SYSTEMS, VOL. 8, NO. 1, FEBRUARY 2021

Huansheng Ning (Senior Member, IEEE) received Runhe Huang (Senior Member, IEEE) received the
the B.S. degree from Anhui University, Hefei, China, Ph.D. degree in computer science and mathematics
in 1996, and the Ph.D. degree from Beihang Uni- from the University of the West of England, Bristol,
versity, Beijing, China, in 2001. He is currently a U.K., in 1993.
Professor and the Vice Dean of the School of Com- She is currently a Full Professor with the Faculty
puter and Communication Engineering, University of Computer and Information Sciences, Hosei Uni-
of Science and Technology, Beijing. versity, Japan. She received a Sino-Britain Friend-
He is the Founder and the Chair of the Cyberspace ship Scholarship for her study in the U.K. Her
and Cybermatics International Science and Tech- research fields include cognitive computing, brain
nology Cooperation Base. He has presided many modeling, computational intelligence computing, big
research projects, including the Natural Science data, machine learning.
Foundation of China and the National High Technology Research and Devel-
opment Program of China (863 Project). He has authored or coauthored more
than 100+ journal/conference articles and authored five books. His current Jianhua Ma (Member, IEEE) received the B.S.
research focuses on the Internet of Things and general cyberspace. and M.S. degrees from the National University of
Dr. Ning serves as an Associate Editor for the IEEE S YSTEMS J OURNAL Defense Technology, Changsha, China, in 1982 and
(2013-currently), the IEEE I NTERNET OF T HINGS J OURNAL from 2014 to 1985, respectively, and the Ph.D. degree from Xidian
2018, and as a Steering Committee Member of the IEEE I NTERNET OF University, Xi’an, China, in 1990.
T HINGS J OURNAL (2016-currently). He is currently a Full Professor with the Faculty
of Computer and Information Sciences, Hosei
University, Japan.
Dr. Ma is an Editor-in-Chief of Journal of Ubiq-
uitous Computing and Intelligence (JUCI), Journal
of Mobile Multimedia (JMM), Journal of Auto-
nomic and Trusted Computing (JoATC), and International Journal of
u- and e-Service, Science and Technology (IJUNESST), and an Assistant
Nyothiri Aung received the M.E. degree in informa- Editor-in-Chief of International Journal of Pervasive Computing and Com-
tion technology from Mandalay Technological Uni- munications (JPCC). He is an Area Editor of the International Journal
versity, Mandalay, Myanmar, 2012, and the Ph.D. of Computer Processing of Oriental Languages (JCPOL), an Associated
degree in computer science and technology from Editor of the International Journal of Distance Education Technologies
the University of Science and Technology Beijing, (JDET) and Journal of Human-centric Computing (HCIS) (Springer), on the
Beijing, China, in 2020. Advisory Board of the International Journal of Smart Home (IJSH), and
She worked as tutors with the Department of Infor- on the Editorial Board of the International Journal of Wireless and Mobile
mation Technology, Technological University of Computing (IJWMC), International Journal of Security and Its Applications
Meiktila, Meiktila, Myanmar, from 2008 to 2010 and (IJSIA), International Journal of Database Theory and Application (IJDTA),
a System Analyst of ACE Data System, Yangon, International Journal of Computing & Information Technology (IJCIT), and
Myanmar, from 2012 to 2015. Her research interests International Journal On Advances in Software. He is the Founder Chair of the
include social computing, personality computing, and intelligent transportation IEEE CIS Smart World Technical Committee and the IEEE SMC Cybermatics
systems. Technical Committee.

Authorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 24,2021 at 05:56:37 UTC from IEEE Xplore. Restrictions apply.

You might also like