Decision Support Systems 63 (2014) 52–63

Contents lists available at ScienceDirect

Decision Support Systems

journal homepage: www.elsevier.com/locate/dss

Object typicality for effective Web of Things recommendations

Yi Cai a, Raymond Y.K. Lau b,⁎, Stephen S.Y. Liao b, Chunping Li c, Ho-Fung Leung d, Louis C.K. Ma e
a
School of Software Engineering, South China University of Technology, China
b
Department of Information Systems, City University of Hong Kong, Hong Kong Special Administrative Region
c
School of Software, Tsinghua University, 100084 Beijing, China
d
Department of Computer Science and Engineering, The Chinese University of Hong Kong, China
e
SCOPE, City University of Hong Kong, Hong Kong Special Administrative Region

a r t i c l e i n f o a b s t r a c t

Available online 23 September 2013 With the rapid growth of “Web of Things” (WoT), there is a pressing need to develop effective mechanisms for
the intelligent discovery and selection of these things (items). Recommender systems are viable solutions to ad-
Keywords: dress the issue of WoT discovery and selection. However, classical recommender systems are weak in handling
Object typicality sparse recommendation spaces which characterize most WoT recommendations. Moreover, classical recom-
Recommender systems
mender systems may not be able to scale up to efficiently process a large number of things on the Web, and
Situation awareness
yet these systems may produce big-error recommendations that diminish users' trusts on utilizing WoT. The
Web of Things
main contribution of our research is the design and development of a novel recommendation method which is
underpinned by the principle of object typicality verified in the field of cognitive psychology to address the afore-
mentioned issues related to WoT recommendations. Based on the MovieLens benchmark data set, our experi-
mental results show that the proposed recommendation method is effective and produces the least big-errors.
Since the proposed method exploits data generalization by operating at item group and user group level during
recommendation time, it is more effective and efficient than other baseline methods given sparse training data.
Based on the Netflix benchmark data set that simulates a large WoT recommendation space, the proposed meth-
od also significantly outperforms state-of-the-art recommendation methods in terms of Mean Absolute Error
(MAE). The business implication of our research is that the proposed recommendation method can enhance
the situation awareness of WoT applications which facilitate the reuse of enterprise resources and the interoper-
ability among enterprises.
© 2013 Elsevier B.V. All rights reserved.

1. Introduction
The transition from the Social Web to the Semantic Web has trig-
gered increasingly more physical devices (i.e., smart things) such as
RFID chips, wireless sensors, actuators, and mobile phones to be
connected to the digital world for the development of useful real-
world applications on the Internet (i.e., the “Internet of Things”) [12].
More recently, the vision of “Web of Things” (WoT) leads to the design
of more sophisticated applications or services by interconnecting any
objects through the Web layer (e.g., interconnecting via the HTTP proto-
col) [8,12]. For example, a movie recommendation service (i.e., a smart
thing) is connected to a sensor installed at a cinema; a new WoT appli-
cation is then composed to timely inform users once tickets of the rec-
ommended movies are nearly sold out at that cinema.
However, it has been pointed out that the emerge of a large number
of smart things on the Web causes great difficulties to computers as well
⁎ Corresponding author.
E-mail addresses: ycai@scut.edu.cn (Y. Cai), raylau@cityu.edu.hk (R.Y.K. Lau),
issliao@cityu.edu.hk (S.S.Y. Liao), cli@tsinghua.edu.cn (C. Li), lhf@cuhk.edu.hk
(H.-F. Leung), islma@cityu.edu.hk (L.C.K. Ma).
as humans to find, select, and utilize smart things in an effective way
[11,25]. Due to the problem of information overload [19,38,47], it is ex-
tremely difficult for Web users to develop the situation awareness about
the huge number of things initiated on the Web everyday. Accordingly,
there is a pressing need to examine the issue of smart things discovery
and selection on the Web. Since recommender systems have been
shown to be viable solutions for the discovery and selection of services
on the Web [49], this paper focuses on the design and development of a
novel recommendation method that can facilitate users (humans or ma-
chines) to discover and select relevant smart things. For example, the
proposed WoT recommendation service can autonomously suggest
some useful things to users based on their previous usage experience
and the preferences of other like-minded users.
A large body of research about recommender systems has been
performed in the past two decades. Existing recommender systems
are usually classified under one of the three broad categories, namely
collaborative filtering (CF) [17,36], content-based recommendation
(CB) [28,33], and hybrid recommendation [2,24,27]. Content-based
recommendation methods suggest items to a user based on similar
items s/he having consumed before. On the other hand, collaborative fil-
tering methods recommend items to a user based on the preferences of

0167-9236/$ – see front matter © 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.dss.2013.09.008
 Y. Cai et al. / Decision Support Systems 63 (2014) 52–63
similar users. Hybrid recommendation methods represent a model-
based or heuristic-based combination of the aforementioned recom-
mendation methods. Although these recommendation methods have
been widely used in electronic commerce, they are faced with the fol-
lowing challenges for WoT recommendations:
• First, data sparsity is a big challenge given few ratings for a large num-
ber of items on the Web.
• Second, computational efficiency is another concern given the sheer
volume of information about user preferences and items on the
Web. Nevertheless, it is desirable for recommender systems to gener-
ate recommendations in real-time to facilitate the timely composi-
tions of WoT applications.
• Third, existing recommender systems may generate big-error predic-
tions. In other words, some items suggested by a recommender sys-
tem are very different from the actual preferences of a user. These
big-error predictions diminish users' trusts on WoT applications.
The main contribution of the research work presented in this paper
is that we exploit the principle of “object typicality” extensively studied
in the field of cognitive psychology [9] to develop a novel, cognitively
motivated recommendation method called Recommendation based
On Typicality (ROT) to address the aforementioned issues arising in
WoT recommendations, and hence to improve the situation awareness
of WoT applications. Since recommender systems deal with human per-
ceptions about objects, there is a distinct advantage of designing recom-
mender systems grounded in the principle of human cognition. To the
best of our knowledge, this is the first successful design of a cognitively
motivated recommender system to enhance WoT recommendations. In
particular, ROT exploits high-level item- and user-based similarities to
facilitate WoT recommendations. The basic intuition of the proposed
method is that a “typical” user of a specific user interest group should
be recommended “typical” items that the group is most interested in.
For example, a typical user of the “war movies” interest group tends to
prefer the typical movies (e.g., “The Longest Day”) characterizing the in-
terests of the group.
More specifically, the proposed ROT method first exploits the natural
partitions of items of a given recommendation space to establish item
groups. Then, the user interest group pertaining to each item group is
identified. Finally, items are recommended to a user according to a
novel typicality-based computational mechanism that exploits the
user's fuzzy memberships pertaining to various user interest groups
and the typical items characterizing each of these groups. Pragmatically,
items are recommended according to the match between user groups
and item groups at execution time. In other words, the ROT method
operates at a high level of data granularity (data generalization) [47].
This is a novel way on how ROT addresses the issues of data sparsity
and computational complexity arising in WoT recommendations.
The remainder of the paper is organized as follows. Section 2 dis-
cusses related research of recommender systems and compares existing
work with our proposed approach. The computational details about the
proposed object typicality based recommendation method are illustrat-
ed in Section 3. In Section 4, we discuss the results of our empirical ex-
periments based on the MovieLens data set1 and the Netflix data set.2
We then summarize the main characteristics of the proposed ROT
method in Section 5. Finally, we offer concluding remarks and describe
future directions of our research work.
2. Related work
2.1. Research on object typicality
Psychologists have found that people are more interested in typical
objects than atypical ones when a concept (i.e., a category of objects)
1
http://www.grouplens.org/.
2
http://archive.ics.uci.edu/ml/datasets.html.
53
is referred to [29]. According to the prototypical view of concepts,
each concept is represented by the best prototype capturing the salient
properties of objects belonging to that category [26]. Vanpaemel et al.
[42] extended the prototypical view of concepts by developing methods
to identify the prototypes of a concept based on typical objects. In par-
ticular, an object is considered to be an instantiation from the most sim-
ilar abstraction (prototype). Barsalou [3] proposed two quantitative
measures, namely central tendency and frequency of instantiation to esti-
mate the typicality of an object with respect to a given concept. Central
tendency refers to the degree of an object's “family resemblance”. An
object is considered to have a high central tendency if it is similar to
other members of the same category and it is different from the mem-
bers of other categories. On the other hand, frequency of instantiation
refers to the frequency of an object being referred to by people when
a specific concept is examined. If an object is often used as an exemplar
for a concept, it has a high frequency of instantiation, and therefore it is
considered a typical one with respect to that concept.
Rifqi [34] proposed a computational method to estimate object typ-
icality in large databases. In particular, the typicality of an object is esti-
mated according to its resemblance to other members of the same
category and its dissimilarities to the members of other categories.
Lesot et al. [21] developed a similar computational method in the con-
text of fuzzy systems. Desclés and Pascu [6] applied the notion of object
typicality to construct new quantifiers for natural language processing
and common sense reasoning. Cai and Leung [5] formalized object typ-
icality with reference to an ontology. Hua et al. [14] applied the principle
of object typicality to develop a typicality-based query operator that en-
hances the effectiveness of query processing in databases.
2.2. Recommender systems
The assumption of content-based recommender systems is that peo-
ple prefer items similar to those that they positively evaluate before. For
content-based recommender systems, the central issue is to examine
computational methods for learning user profiles and measuring item
similarity. For example, Pazzani and Billsus [33] applied the naive Bayes
classifier to construct a user profile that captured “relevant” and “non-
relevant” Web pages for the user. Mooney and Roy [28] developed the
LIBRA system for the recommendations of books. A detailed account of
content-based recommender systems is provided by Pazzani and Billsus
[32].
On the other hand, collaborative filtering-based recommender sys-
tems suggest items to a user based on the preferences of other like-
minded users. Since the collaborative filtering approach does not require
well-structured item descriptions, it has been widely used to recommend
a variety of items including images, videos, and music [1]. For example,
GroupLens [17] and PHOAKS [40] were developed based on the collabo-
rative filtering approach. The user-based CF approach first identifies the
nearest “neighbors” of a user by exploiting the user similarity relations.
Then, the system predicts the rating of an unrated item based on the rat-
ings given by these nearest “neighbors” [13]. Moreover, Zheng et al. [49]
developed a user-based CF approach for the discovery and selection of
Web services. In contrast, item-based CF approach recommends items
to a user based on other users' ratings of similar items. More specifically,
an item-based CF system first identifies the nearest “neighbors” of an
unrated item by examining the item similarity relations. Then, the system
predicts the rating of the unrated item based on other users' ratings
assigned to these nearest “neighbors” of items [7,36].
Hybrid recommender systems combine CB- and CF-based approaches
to address the limitations of individual recommendation method
[2,27,37]. Melville et al. [27] applied a content-based approach to aug-
ment a user-item rating matrix, and then used a collaborative filtering ap-
proach to generate the final recommendations. Xue et al. [46] developed a
cluster-based Pearson Correlation Coefficient method (SCBPCC) that
exploited both item similarity and user similarity in collaborative filtering.
Wang et al. [43] proposed the similarity fusion method to unify user- and
54 Y. Cai et al. / Decision Support Systems 63 (2014) 52–63
item-based CF methods. Ma et al. [24] proposed an effective missing
data prediction (EMDP) method that leveraged the advantages of both
item- and user-based CF approaches. Recently, Li et al. [22] developed
a transfer learning-based recommendation method by means of the ex-
traction of cross-domain contextual information. Kim et al. [16] pro-
posed a collaborative filtering approach to enrich user modeling for
more personalized recommendations. Xu et al. [45] combined social
network analysis and semantic concept analysis to facilitate personal-
ized researcher recommendations.
2.3. Differences between the proposed ROT method and existing methods
The proposed ROT method is grounded in the sound principle of
object typicality which has been verified based on numerous empirical
tests in the field of cognitive psychology [26,29,42]. In addition, the
principle of object typicality has been successfully applied to enhance
user query processing in database systems [14]. The basic intuition of
the proposed ROT method is that a typical user of a specific user interest
group should be recommended typical items that the user group is most
interested in. The proposed ROT method differs from existing content-
based recommendation approach in that it does not directly recom-
mend new items to a user based on the items s/he positively rated be-
fore. Moreover, it differs from CF-based recommendation approach in
that it does not need to compute the “nearest neighbors” of an individ-
ual user when a recommendation for that user is generated. In contrast,
the ROT method represents each user by a typicality vector which cap-
tures the user's typicality degrees with respect to various user interest
groups. The user typicality vectors and user groups are constructed by
means of an off-line system training process.
The ROT method is different from economic choice models in that it
generates recommendations based on the typicality of a user with re-
spect to various user interest groups and the typical items that these
groups prefer instead of the joint statistical distributions of item proper-
ties and user preferences. Pragmatically, ROT operates at item group and
user group level rather than individual user and item levels during rec-
ommendation time. This is a unique way on how the proposed ROT
method addresses the issues of data sparsity and computational efficien-
cy for WoT recommendations. Moreover, there is a good potential for
the ROT method to address the problem of big-error predictions due
to the sound principle of object typicality. It should be noted that the
proposed ROT method differs from clustering-based CF methods such
as SCBPCC [46] in that ROT applies data clustering to identify item
groups and user groups, whereas SCBPCC invokes clustering processes
to predict missing ratings. Though the proposed ROT method utilizes
item descriptions to construct item groups, it does not involve extra
data processing when compared to existing CB and hybrid methods
which also leverage item descriptions to generate recommendations.
3. The computational model for typicality-based recommendations
In this section, we illustrate the computational details of the pro-
posed typicality-based recommendation model which predicts users'
ratings on items based on both user typicality and item typicality. The
concept of typicality provides a sound mechanism to rank items in a
way that is close to human perception [9,29]. Since the main functional-
ity of recommender systems is to predict users' preferences about items,
it is desirable to design recommender systems that can imitate the way
how people evaluate real-world items.
3.1. The computational apparatus of object typicality
To illustrate the computational apparatus of the proposed typicality-
based recommendation method, we first formally define items, item
groups, users, and user groups, respectively. Let U and O denote a set
of users and a set of items (objects). Each item is represented by some
properties. For example, a movie is represented by directors, actors,
and country of production, whereas a Web service is represented by
the service provider, service quality, expected response time, and so
on. It should be noted that an item can be represented by any properties
rather than restricting to keywords. Formally, each item is represented
by an item vector as follows.
Definition 1. Item vector
An item Oi is represented by a vector of property–value pair such as
D


E
!
p Oi ¼ pOi ;1 : lOi ;1 ; pOi ;2 : lOi ;2 ; ⋯; pOi ;n : lOi ;n where lOi ; j ∈½0; 1 in-
dicates the fuzzy degree of Oi possessing the property pOi ; j . If lOi ; j ¼ 0 is
established, it indicates that item Oi does not possess property pOi ; j at all,
whereas lOi ; j ¼ 1 suggests that Oi definitely possesses property pOi ; j .
For example, a computer Oi is represented by an item vector ! p Oi =
b (‘has intelCPU’: 1), …, (‘has wideScreen’: 0.8) N. For the proposed
computational method, a set of fuzzy clusters is first constructed
based on all the items under consideration. Each fuzzy cluster is
called an item group which consists of similar items. An item be-
longs to some fuzzy clusters with various degrees. For example,
!
p Oi = b (‘has intelCPU’: 1), …, (‘has wideScreen’: 0.8) N may belong
to the item group “desktop computer” and the item group “netbook”
with various degrees. From a cognitive perspective, each item group
represents a concept of a specific domain. Formally, an item group is
represented by a fuzzy cluster which is underpinned by the theory
of fuzzy sets [48]. Essentially, each cluster is represented by the
same set of properties (e.g., keywords). However, each cluster
owns these properties with various degrees (i.e., fuzzy member-
ships). An item group is defined as follows.
Definition 2. Item group
An item group kj is a fuzzy cluster of similar objects represented by:
n o
w w w
k j ¼ O1 j;1 ; O2 j;2 ; ⋯; Om j;m where m represents the number of items; Oi
is an item, and wj,i is the typicality degree of item Oi for the cluster kj.
Formally, the construction of fuzzy clusters is represented by a map-
ping function (i.e., a fuzzy membership function): m : O × K ↦ [0,1],
where O is a set of items and K = {k1,k2, ⋯,kn} is a set of clusters. The
term n denotes the number of clusters induced based on the natural pat-
terns of the set of items O. The term wj,i denotes the typicality degree
(i.e., fuzzy membership) of an item Oi for the item group kj. For the pro-
posed ROT method, different clustering methods can be applied to con-
struct the fuzzy clusters (i.e., item groups) [23,44]. However, the search
for the most effective clustering method given a specific domain will be
left as part of our future work. If a clustering method is effective, the se-
mantic distances among different fuzzy clusters tend to be large (i.e., the
overlapping among item groups is small).
According to the prototypical view of concepts [5,29], each item
group can also be represented by a single prototype induced based
on the set of object instances described by the corresponding con-
cept. For a computer-based implementation of concept prototype,
previous research work has proposed using a prototype vector to
represent a concept prototype which is an abstraction of a group of
similar objects [5,29,42]. A prototype vector is induced based on
the set of similar objects described by the concept. For example, a
prototype vector can be composed based on the mean, median, or
mode vector derived from the set of object (item) vectors of a specif-
ic category (concept) [21,29]. Formally, a prototype vector is defined
as follows.
Definition 3. Prototype vector
D



A prototype vector !
t kj ¼ pk j ;1 : r k j ;1 ; pk j ;2 : r k j ;2 ;⋯; pk j ;m : r k j ;m
〉
is a vector of property–value pairs which represent the prototype of an
item group kj. The term m denotes the number of properties of the
 Y. Cai et al. / Decision Support Systems 63 (2014) 52–63
prototype kj, and r k j ;i ∈½0; 1 represents the fuzzy degree of kj possessing
the property pk j ;i .
For the special case, if r k j ;i ¼ 0 holds, it means that the item group kj
does not possess the property pk j ;i at all. In contrast, if r k j ;i ¼ 1 holds, it
suggests that kj definitely possesses the property pk j ;i . For the set of
users U, each individual user may have specific favorites or preferences
for some kinds of items (i.e., item groups). For instance, a user Bob may
be interested in “Horror” movies, while his wife Amy may prefer
“Adventure” movies, and yet their son Tom likes “Thrillers” movies.
For each induced item group ki, the proposed recommendation method
constructs a fuzzy set [48] of users who like the corresponding item
group. Within a user interest group gi ⊂ U, each user Ux ∈ gi is interest-
ed in the group of items to various degrees. For example, Bob likes
“Horror” movies to a higher degree than Tom does. The set of users
pertaining to an item group ki is called a user group gi, and it is repre-
sented by a fuzzy set. The semantics of gi is that it describes the set of
users who like the particular item group ki. The membership of each
member Ux ∈ gi indicates the typicality degree of Ux with respect to gi.
The fuzzy membership of Ux for gi is called the user typicality. The ratio-
nale of using fuzzy sets to represent item groups and user groups is that
an item may belong to multiple item groups and a user may also have
multiple interests pertaining to different user interest groups in the
real-world.
Definition 4. User group
n o
v v v
A user group g i ¼ U 1i;1 ; U 2i;2 ; ⋯; U mi;m is a fuzzy set of users who like
the corresponding item group ki. The term m denotes the number of
users of the group gi, and Ux represents a user. The term vi,x ∈ [0,1]
(i.e., fuzzy membership) is the typicality degree of the user Ux for the
user group gi.
There are four important components of the proposed recom-
mender system, namely items, users, item groups, and user groups.
Fig. 1 shows the relationships between these four components. For
each item, it has different typicality degrees with respect to the
item groups. Similarly, each user has different typicality degrees
with respect to different user groups. It should be noted that item
groups are first constructed via a chosen clustering method, and
then the user group pertaining to each item group is composed. An
item Oi may have a fuzzy membership wj,i N 0 for multiple item
groups, whereas a user Ux may also have a fuzzy membership vi,
x N 0 for multiple user groups.
According to the principle of “object typicality”, a “typical” user of
a specific user interest group should be recommended “typical”
Fig. 1. The relationships among items, item groups, users, and user groups of the ROT
model.
55
items that the user group is most interested in. Therefore, the recom-
mendation strategy of the proposed ROT method is that if an item Oi
is a typical (atypical) item of the item group kj, and a user Uj is a typ-
ical (atypical) user of the user group gj that corresponds to the item
group kj, item Oi should be recommended to the user Uj with a high
(low) recommendation score. In this case, the pair (kj,gj) (i.e., the
matching item group and user group) represents a specific “recom-
mendation context” discovered via un-supervised item clustering
and user clustering. Formally, the recommendation score of Oi for
U
!j is!estimated
 via the typicality-based recommendation function RS
O i ; U j : O×U↦½0; 1, where O is a set of items represented by the cor-
responding item group vectors, and U is a set of users represented by
the corresponding user group vectors. More specifically, the pro-
posed recommendation function RS should observe the following
axioms.
 !
Axiom 1. For a user Uj and an item Oi, if ∀ x, vx,j = 0, then RS !
O i ; U j ¼ 0.
 !2. For a user Uj and an item Oi, if ∃ x, vx,j = 1 and wx,i = 1, then
Axiom
RS !
O i ; U j ¼ 1.
Axiom 3. For a user Uj and two items Ox and Oy,if ∃ ki : wi,x N wi,y, and
! 
∀ h ≠ i : wh,x = wi,y, then RS ! ! !
O x ; U j NRS O y ; U j .
Axiom 4. For two users Ux and Uy and one item Oi, if ∃ gj : vj,x N vj,y, and
! ! 
∀ h ≠ j : vh,x = vh,y, then RSð! !
O i ; U x ÞNRS O i ; U y .
Axioms 1 and 2 specify the boundary cases of generating the recom-
mendation scores. If a user does not like any item groups at all (i.e., the
user is not typical in any user group), it means that the user does not like
any items for all item groups, and so the recommendation score of any
item for the user should be zero. For a user Uj and an item Oi, if Uj likes
an item group kx with the highest degree (i.e., Uj is the most typical
user of the user group gx), and Oi is the most typical item of the item
group kx, the ROT system should recommend Oi to Uj with the highest
recommendation score. On the other hand, Axioms 3 and 4 specify the
influence of user typicality degrees from user groups and item typicality
degrees from item groups, respectively. If an item is more typical for
some item groups, and a user is more typical for the corresponding
user groups, the ROT system should recommend the item to the user
with a relatively high recommendation score. Accordingly, we propose
a novel typicality-based recommendation function RS that observes all
the aforementioned axioms.
Xn

 wx;i  vx; j
! !
O i; U j ¼ ð1Þ
x¼1
RS
n
The term n represents the number of item groups (also the number
of user groups). Moreover, wx,i is the typicality degree of the item Oi for
the item group kx, and vx,j is the typicality degree of the user Uj for the
user group gx that corresponds to the item group kx. As can be seen,
the proposed RS function is underpinned by the typicality degrees of
items and the typicality degrees of users. The computational details of
deriving these typical degrees will be illustrated in Sections 3.2 and
3.3, respectively. It should be noted that Eq. (1) does not rate an item
based on the average rating (i.e., simply recommending popular items
to a user). In contrast, it takes a user's specific interests into account
by evaluating their typicality with respect to various user interest
groups. Even though Eq. (1) produces a recommendation score that
falls in the unit interval, the proposed ROT system can simply invoke a
linear conversion function ϑ : [0,1] ↦ [Min, Max] that maps the unit
interval to any desirable range [Min, Max] of recommendation scores
for a variety of WoT recommendation applications.
3.2. Estimating typicality degrees of items in item groups
Previous research in the field of cognitive psychology shows that
central tendency (i.e., an object's similarity to a concept prototype) is
56 Y. Cai et al. / Decision Support Systems 63 (2014) 52–63

one of the main factors that largely influences an object's typicality the following aggregation function to estimate central tendency ct
with respect to a concept [4,5]. When the principle of object typical- of an item Oi.
ity is applied to design the proposed recommender system for WoT



recommendations, the central tendency of an item (i.e., an object) ! ! ! ! ! !
τk j ðOi Þ ≈ ct pO ; t c
i
¼ simint pO ; t c
i
 dissimext pO ; t c
i
ð5Þ
with respect to an item group (i.e., a concept) is estimated based
on the similarity between the corresponding item vector and proto-
type vector of the item group. More specifically, the central tendency For the proposed ROT method, each item has a typicality degree
of an item is estimated based on its internal similarity and external against each item group. Accordingly, an item typicality matrix is need-
dissimilarity [3,5]. Internal similarity refers to the similarity between ed to characterize item typicality. Formally, an item typicality matrix is
an object and the prototype of the corresponding concept, whereas defined as follows:
external dissimilarity is the distance between the object and the pro- Definition 5. Item typicality matrix
totypes of other concepts. As discussed in Section 3.1, objects and
prototypes of concepts are represented by the corresponding item An item typicality matrix MIK captures the typicality degree for each
vectors and prototype vectors of item groups. Conceptually, a simi- item Oi against each item group kj of a recommendation domain, and it
larity function sim is applied to map the product of the set of items has the form:
and the set of prototypes of item groups to the unit interval:
sim : P × T ↦ [0,1], where T is a set of prototype vectors of item 8


 9
< k1 : w1;1 ; k2 : w2;1 ; ⋯; kn : wn;1 >
> =
groups, and P is a set of property vectors of items. For ! each pair of
! MIK ¼

⋯ 

item vector ! !
p Oi ∈P and prototype vector t c ∈T , if sim t c ; p Oi ¼ 1 is >
: k :w >
;
1 1;m ; k2 : w2;m ; ⋯; kn : wn;m
established, it means that the item Oi is certainly typical with respect
to the prototype c of the item group kj. On the other hand, if sim
! !  where w j;i ¼ τk j ðOi Þ is the typicality degree of item Oi for the item group
t c ; p Oi ¼ 0 is true, it suggests that item Oi is totally atypical with re-
spect to the prototype c of the item group kj. kj.
For the proposed ROT method, any standard similarity functions
discussed in the literature can be used [35]. For the implementation 3.3. Estimating typicality degrees of users in user groups
of our current prototype system, we employ the following similarity
function: For the proposed ROT method, estimating the typicality degrees of
users with respect to various user groups is also an important proce-

 0 ∀k; ∃lOi ;k ¼ 0 dure. Similar to the item typicality matrix, a user typicality matrix cap-
!
sim !pO ; t c ¼ ð2Þ
i κ otherwise tures the typicality degrees of a user against various user groups.

0 0vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1θ 1 Definition 6. User typicality matrix

u n  2
uX 
κ ¼ exp@−@t r c;k −lOi ;k  A A ð3Þ
k¼1
A user typicality matrix MIG captures the typicality degree for each
user Ui against each user group gj of a recommendation domain, and it
where n represents the number of properties of the universe of dis- has the form:
course. The term lOi ;k represents the degree to which item Oi possesses 8


 9
>
< g 1 : v1;1 ; g 2 : v2;1 ; ⋯; g n : vn;1 > =
the property pOi ;k , while rc,k indicates the degree to which prototype c
MIG ¼

⋯ 

possesses the property pc,k. The parameter θ ∈ R+ is a positive real num- >
: g :v >
;
1 1;m ; g 2 : v2;m ; ⋯; g n : vn;m
ber which is applied to tune the rate of decay of the exponential function
so that the proposed similarity function is effective for different recom-
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 2ffi where v j;i ¼ τg j ðU i Þ is the typicality degree of user Ui for the user group
n  
mendation applications. The Euclidean distance ∑k¼1 r c;k −lOi ;k  is gj.
applied to estimate the similarity between an item vector ! p O and a pro-
In order to estimate the typicality degree of a user for each user
i
group, we need to interpret the principle of central tendency in
totype vector !t c . The proposed exponential function produces a high
the context of typical user group formation. In particular, a user in-
typicality degree if an item is extremely close to a prototype. On the
terest group gj is constructed with respect to a specific item group
other hand, the resulting typicality degree substantially decreases if
kj (i.e., the items generally preferred by the group of users). Under
the item is only moderately similar to the prototype [29].
such a circumstance, if a user Ui often rates the items of kj, s/he has
The internal similarity of an item Oi with respect
 to the prototype
! ! c of
! the tendency to prefer these items. Consequently, s/he is likely to
an item group is computed according to simint ! p Oi ; t c ¼ sim p Oi ; t c . In
 !
 be the typical member of the corresponding user interest group gj.
addition, the external dissimilarity dissimext ! p Oi ; t c is taken as the av-
In fact, it reveals Ui's central tendency in gj. Moreover, if the user Ui
erage of dissimilarities between the item and other prototypes exclud-
highly rates the items of kj, it also shows their tendency of preferring
ing the current one. More specifically, the external dissimilarity
those items. As a result, Ui has a high central tendency in gj. Accord-
between an item Oi and the prototype c of an item group is defined as
ingly, we estimate the central tendency, and hence the typicality de-
follows:
gree of a user Ui with respect to a user group gj based on the
X
 frequency of Ui rating each item Oi ∈ kj and the rating that Ui spec-
!

 dissim !pO ; t x
dissimext ! !
pO ; t c ¼
x∈C=fcg i
ð4Þ ifies for Oi ∈ kj. Formally, these two factors are expressed by sig ; f
jC j−1
i j
and sig j ;r , respectively:

where C is the set of prototypes of the item groups for a given rec- i N j;i
ommendation domain. For the dissimilarity function dissim, it is de- sg j ; f ¼ Xn ð6Þ


 N
y¼1 y;i
! !
fined by dissim ! i
!
p O ; t c ¼ 1−sim p O ; t c
i
. Finally, the central
tendency of an item with respect to an item group is estimated XN
i y¼1
j;i
w j;y  Ri;y
based on an aggregation function that combines the internal simi- sg j ;r ¼ ð7Þ
larity and external dissimilarity [21]. More specifically, we propose N j;i  R max
 Y. Cai et al. / Decision Support Systems 63 (2014) 52–63 57

where n is the total number of item groups. The term Nj,i represents
the number of items that user Ui has rated for an item group kj. Ri,y is
the rating that Ui assigns to the item Oy ∈ kj. The term wj,y repre-
sents the typicality degree of Oy for the item group kj, and Rmax is
the maximal rating value for any items.
Finally, the typicality degree v j;i ¼ τg j ðU i Þ of a user Ui for a user
group gj is approximated by the aggregation of the above two
measures through a standard convex function. The weight factor
α = 0.5 is specified to assign equal weight for sig ; f and sig j ;r , respec-
j
tively. The proposed user typicality function has the property that
if the user Ui often assign high rating values for the items Oi ∈ kj,
Ui's typicality degree for the user group gj that corresponds to kj
will be high.
i i
τg j ðU i Þ ≈ α  sg j ; f þ ð1−α Þ  sg j ;r ð8Þ
4. System evaluation
In this section, we discuss the experiments that evaluate the perfor-
mance of the proposed typicality-based recommendation method when
compared to that of state-of-the-art recommendation methods. In par-
ticular, we try to answer the following research questions by running
two series of experiments:
1. How does item clustering affect the recommendation quality of the
proposed ROT method?
2. Does the proposed ROT method produce less big-error predictions
when compared to existing recommendation methods?
3. Does the proposed ROT method achieve good performance given
sparse training data?
4. Is the proposed ROT method more efficient than some classical rec-
ommendation methods?
5. Is the proposed ROT method effective when compared to state-of-
the-art recommendation methods?
4.1. Evaluation data sets
Since a benchmark data set specifically constructed for the eval-
uation of WoT recommender systems is not available, we applied
the common MovieLens [24,36] and Netflix [18] data sets to evalu-
ate the effectiveness of the proposed ROT prototype system. The
MovieLens data set we used contains 100,000 ratings generated
by 943 users for 1682 movies. Each user has rated at least 20
movies, and the ratings are in the range between 1 (bad) and 5 (ex-
100;000
cellent). The sparsity level of this data set is 1−943×1;682 ¼ 0:9369. We
downloaded the descriptions of movies from the IMDB movie site.3
Then, the properties of the movies (items) were identified by
extracting the corresponding keywords from these descriptions.
Since only the descriptions of 1334 movies were available, we
used this subset of movies in our experiments. In addition, we
also evaluated the ROT system based on a subset of the Netflix
data set which contained 100,480,507 ratings offered by 480,189
users for 17,770 movies. The sparsity level of our Netflix data set
100;480;507
is 1−480;189×17;770 ¼ 0:9882.
compare our experimental results with published results. Formally,
MAE is defined by:
Xn
j f i −hi j
MAE ¼ i¼1
n
where n is the number of recommendations. fi is the user-specified
rating, and hi is the prediction provided by a recommender system.
If the MAE score is low, it suggests that the recommender system
can predict users' ratings with a high accuracy.
Another common metric for the evaluation of recommender sys-
tems is root mean square error (RMSE). Since the errors are squared be-
fore they are averaged, the RMSE measure is more sensitive to big-error
predictions when compared to the MAE measure. In general, the RMSE
score achieved by a recommender system is always greater than or
equal to its MAE score. The greater difference between the MAE score
and the RMSE score achieved by a system, the larger variance of the in-
dividual errors produced by the system will be. Formally, the RMSE
measure is defined as follows.
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
sX
n 2
i¼1 i
j f −hi j
RMSE ¼
n
4.2. The experimental procedures
For all the experiments reported in this paper, we applied the topic
modeling based clustering method to generate item groups [39]. After
the clustering process, a prototype vector was extracted for each cluster
(i.e., item group). In particular, the top k most probable (typical) key-
words were applied to construct each prototype vector after the proba-
bilistic topic modeling process was invoked. For our experiments, the
parameter k = 20 was empirically established. Table 1 shows the sam-
ple prototype vector for the “crime movie” item group generated by
means of a topic modeling process [39]. According to our empirical ob-
servation, relevant keywords often appear in movie descriptions. As a
result, topic modeling processes tend to extract useful topics that char-
acterize different types of movies. However, we do not make a claim
that topic modeling is the best clustering method. Searching for the
most effective clustering method for a given domain will be left as
part of our future work. After the construction of item groups, the user
group corresponding to each item group was constructed according to
the computational method illustrated in Section 3.3. To answer the
aforementioned research questions, we performed two series of
experiments.
The objective of the first series of experiments was to examine the
impact of the number of item groups and the sparsity of training data
on the recommendation quality of the proposed ROT prototype system.
In addition, the computational efficiency of the ROT system was evalu-
ated when compared to several classical recommender systems. The
classical recommender systems involved in our experiments included
a content-based recommender system (CB) that used the cosine
Table 1
The Prototype of the Item Group “Crime Movies”.

Properties Membership Properties Membership

4.1. Performance metrics
Murder 0.9830 Police 0.8216
Prison 0.3689 Suicide 0.3662
To measure the accuracy of the experimental system, we adopted Revenge 0.3642 Serial-killer 0.3628
the Mean Absolute Error (MAE) which is defined as the average abso- Courtroom 0.3612 Obsession 0.3030
lute difference between the system predicted ratings and the actual Investigation 0.2564 Lawyer 0.2404
user ratings [1]. MAE is commonly used by researchers to evaluate Robbery 0.1830 Police-officer 0.1216
Fugitive 0.1089 Judge 0.1066
the effectiveness of recommender systems, and so it is easy to
Chase 0.1042 Confession 0.0982
Deception 0.0961 Smoking 0.0933
3 Gunfight 0.0864 Escape 0.0840
http://www.imdb.com/interfaces.
58 Y. Cai et al. / Decision Support Systems 63 (2014) 52–63

similarity function to compute item similarity, a user-based collabora- Table 2

tive filtering system (UBCF) that employed the Pearson Correlation Co- Sensitivity of MAE w.r.t. K and training/test ratios.

efficient to compute user similarity, an item-based collaborative 0.1 0.3 0.5 0.7 0.9
filtering system (IBCF) that used Pearson Correlation Coefficient to com-
K = 10 0.8291 0.7804 0.7722 0.7664 0.7624
pute item similarity, a naive hybrid recommender system, and a collab- K = 15 0.8176 0.7774 0.7611 0.7587 0.7529
orative filtering system that supported missing data prediction (EMDP) K = 20 0.8231 0.7806 0.7689 0.7645 0.7582
[24]. For these experiments, we divided the whole data set into two K = 25 0.8262 0.7843 0.7688 0.7654 0.7607
K = 30 0.8308 0.7850 0.7686 0.7657 0.7624
parts namely, a training set and a test set. Each system under testing
K = 35 0.8414 0.7887 0.7754 0.7675 0.7645
was first trained using the training set, and then its performance was
evaluated based on the test set. We applied 5-fold cross-validation to
compute the average MAE and RMSE scores for each system. To assess clusters) has a high predictive power [41]. We tried different numbers
ROT's performance under the condition of sparse training data, various of topics and computed the corresponding perplexities. Fig. 2 shows
divisions of the training and the test sets were tried [36]. For instance, our experimental results; it seems that K ∈ [15,28] leads to lower per-
a splitting ratio x = 0.8 indicates that 80% of the evaluation data are plexity values. We then adopted the range of K ∈ [10,35] to establish
randomly selected to build the training set and the remaining 20% of the number of item groups for the proposed ROT system. For each cho-
data are used to create the test set. sen K value, we applied a different training/test splitting ratio
For the second series of experiments, we compared the performance x ∈ [0.1,0.9] to simulate different data sparsity conditions. Finally, 5-
of ROT with that of state-of-the-art recommender systems based on fold cross-validation was applied to compute the MAE and the RMSE
both the MovieLens [24,36] and the Netflix [18] benchmark data sets. scores achieved by the ROT system. Tables 2 and 3 report the MAE
The state-of-the-art recommender systems included cluster-based and the RMSE scores achieved by the ROT system with respect to vari-
Pearson Correlation Coefficient (SCBPCC) [46], Weighted Low-rank Ap- ous numbers of item groups and training/test splitting ratios.
proximation (WLR) [30], and Transfer Learning-based Collaborative Fil- Our experimental results show that both the number of item groups K
tering (CBT) [22]. Similar to previous research [22,24], we extracted a and the training/test splitting ratio have an impact on the system's recom-
subset of 500 users from the MovieLens data set. For instance, we ran- mendation performance. For example, both the MAE and the RMSE scores
domly selected the first 100, 200 and 300 users from the benchmark are the lowest when K = 15 is applied. The MAE and the RMSE scores
data set to build different training sets such as ML100, ML200, and slightly increase when K is deviated from 15 (e.g., K = 10 and K = 20).
ML300, respectively. The remaining 200 users were then applied to However, the absolute differences of the MAE and the RMSE scores are
build the test set. Moreover, similar to previous research, three different small after a reasonable range of K values are established. Accordingly,
sizes of the observed ratings such as 5 ratings (Given5), 10 ratings we applied the empirically established parameter value (e.g., K = 15)
(Given10), and 15 ratings (Given15) were applied to compute user sim- to the remaining experiments. On the other hand, more training data
ilarity for each user [22,24]. To better simulate a WoT recommendation (e.g., x = 0.9) lead to better recommendation performance, that is, a
environment, the larger and more sparse Netflix data set was applied to lower MAE or RMSE score. Although the proposed empirical method
further evaluate the performance of the ROT system. Based on this larg- may not be able to identify the optimal number of item groups, applying
er benchmark data set, we compared the performance of ROT with that more sophisticated methods such as genetic algorithms [20] for parame-
of other state-of-the-art recommender systems such as SVD [31], ter tuning will only further improve the performance of the proposed ROT
SVD++ [18], and socialMF [15] which were made available via the pub- system reported in this paper. However, more sophisticated parameter
licly accessible MyMediaLite library [10]. tuning will be left as part of our future work.

4.3. Experimental results 4.3.2. Recommendation quality

4.3.1. The impact of number of item groups on recommendation performance
For the proposed ROT system, determining the number of item
groups (clusters of items) is an important design issue. For topic model-
ing based conceptual clustering, perplexity is a common measure ap-
plied to estimate a reasonable number of clusters with respect to the
natural partitions of a given data set [41]. Accordingly, we empirically
established an appropriate number of item groups K based on the
movie descriptions pertaining to the MovieLens data set. In general, a
small perplexity value implies that the learned model (e.g., the derived
We performed a comparative evaluation of the performance of ROT
when compared to that of classical recommender systems under differ-
ent training/test splitting ratios. Figs. 3 and 4 show the comparative
MAE and RMSE performance of the experimental and the baseline sys-
tems. It is obvious that the proposed ROT system outperforms all the
baseline systems for both MAE and RMSE under various training/test
splitting ratios. For example, ROT achieves 0.7529 and 1.0327 for MAE
and RMSE, respectively for a training/test splitting ratio x = 0.9, while
EMDP (the best baseline system) achieves 0.8040 and 1.0826, respec-
tively. Under a testing/test splitting ratio of x = 0.3, ROT achieves
0.7774 and 1.0705 for MAE and RMSE, while the best performing base-
line system IBCF achieves 0.8803 and 1.1544, respectively. This experi-
mental result indicates that the recommendation quality of the ROT
system is higher than that of classical recommender systems.
Tables 4 and 5 depict the details of MAE and RMSE improvements
achieved by the proposed ROT system. When compared to the EMDP
baseline system, the performance improvement achieved by ROT

Table 3
Sensitivity of RMSE w.r.t. K and training/test ratios.

0.1 0.3 0.5 0.7 0.9

K = 10 1.1290 1.0703 1.0592 1.0526 1.0442

K = 15 1.1161 1.0705 1.0434 1.0451 1.0327
K = 20 1.1234 1.0705 1.0524 1.0463 1.0365
K = 25 1.1203 1.0719 1.0523 1.0503 1.0454
K = 30 1.1272 1.0781 1.0563 1.0491 1.0425
K = 35 1.1440 1.08174 1.0636 1.0528 1.0487
Fig. 2. Perplexity vs. number of item groups.
 Y. Cai et al. / Decision Support Systems 63 (2014) 52–63 59

Table 4
Comparative MAE improvements of ROT under different training/test ratios.

0.1 0.3 0.5 0.7 0.9

EMDP 17.56% 14.92% 9.64% 8% 6.35%

CB 14.17% 17.09% 18.54% 18.78% 19.34%
IBCF 9.89% 11.69% 12.08% 12.1% 12.03%
UBCF 11.77% 12.27% 11.55% 10.53% 9.63%
Naive hybrid 19.08% 20.62% 14.97% 12.91% 11.39%

Fig. 3. Comparative MAE performance of ROT.
reduces with the increment of the training/test splitting ratio. However,
when compared to the IBCF and the CB baseline systems, ROT achieves a
larger performance improvement with the increment of the training/
test splitting ratio. For the training/test splitting ratio x N = 0.4, ROT
outperforms the best baseline system EMDP by at least 6.35% in terms
of MAE, and at least 4.6% in terms of RMSE. On the other hand, for train-
ing/test splitting ratio x b 0.4, IBCF is the best performing baseline sys-
tem. ROT outperforms IBCF by at least 9.89% in terms of MAE and at
least 5.67% in terms of RMSE. According to paired one-tail t -tests, all
the performance improvements achieved by the ROT system are statis-
tically significant.
4.3.3. The impact of sparse training data
Figs. 3 and 4 also highlight an interesting property of the ROT system
in that it achieves the best MAE and RMSE scores when compared to
other baseline systems even with a low training/test splitting ratio
(i.e., sparse training data). For instance, ROT achieves a MAE score of
0.8176 and a RMSE score of 1.1161 for the lowest training/test splitting
ratio x = 0.1. The EMDP baseline system achieves similar results only
when sufficient training data is available (e.g., a training/test splitting
ratio x = 0.7). Such a characteristic of ROT is desirable for a WoT recom-
mendation environment where little user ratings (i.e., training data) are
available given a huge user-item recommendation space.
The reason of such a significant performance improvement of ROT
over other baseline systems is that it is difficult for collaborative filtering
based systems to accurately identify the nearest “neighbors” for each
user given sparse training data (e.g., commonly rated items). As a result,
the recommendation accuracy of these systems is generally low. In con-
trast, the ROT system relies on the item typicality of item groups and the
user typicality of user groups to generate a recommendation. The
construction of item groups does not rely on any rated training data at
all. Consequently, the proposed ROT system is less susceptible to sparse
training data. From a cognitive perspective, the proposed typicality-
based recommendation method operates based on the sound object
typicality principle in that the typical user of a specific user interest
group should be recommended the typical items that the user group is
most interested in. From a computational perspective, the proposed
typicality-based recommendation method mainly operates at the
group level rather than individual item or user level. More specifically,
item groups are matched with user groups in order to generate recom-
mendations during the recommendation time. This can be seen as a
kind of “generalization” approach to address the data sparsity issue of
WoT recommendations.
4.3.4. Big-error predictions
As mentioned before, big-error predictions produced by recom-
mender systems can seriously harm users' trusts on recommender sys-
tems. For this experiment, we evaluated the big-errors produced by the
ROT and other baseline systems. For the MovieLens data set, a rating
value between 1 and 5 is assigned to an item. So, the biggest prediction
error is 4. Fig. 5 shows the comparison of the prediction errors (PE)
committed by the ROT and other baseline systems. Our experimental re-
sult was obtained with 15 item groups (K = 15) and a training/test
splitting ratio of x = 0.9. The legend PE = 0 means that there is no pre-
diction error, and PE = 4 implies that the difference between a system
predicted rating and the actual user rating is 4 (the big-error). As shown
in Fig. 5, the proposed ROT system produces the most correct predic-
tions (e.g., 38.58% predictions with PE = 0) and the most small errors
(48.92% predictions with PE = 1) among all the systems. Besides, the
ROT system only generates 0.05% predictions with PE = 4 (the big-
errors). As a whole, ROT generates 12.5% prediction errors for PE ≥ 2,
while the EMDP and the CB baseline systems produce 16% and 29% pre-
diction errors at the same level. The IBCF, UBCF and Naive Hybrid base-
line systems all produce 17% or above prediction errors at the same
level. It is obvious that the proposed ROT system produces the least
big-error predictions when compared to other baseline systems. For
other training/test splitting ratios (i.e., x b 0.9), similar results were
observed. For the reason of brevity, we only report the typical result
(e.g., x = 0.9) in this paper.

4.3.5. Evaluation of computational efficiency

Since a WoT recommendation environment involves a large number
of items and users, computational efficiency is a major concern. Fig. 6
shows the recommendation time consumed by all systems under the

Table 5
Comparative RMSE improvements of ROT under different training/test ratios.

0.1 0.3 0.5 0.7 0.9

EMDP 10.1% 9.32% 7.12% 5.26% 4.6%

CB 8.91% 11.08% 12.72% 12.53% 13.5%
IBCF 5.67% 7.26% 8.29% 8.13% 8.84%
UBCF 7.33% 8.02% 7.89% 6.94% 7.07%
Naive hybrid 11.97% 13.34% 10.06% 8.09% 7.39%
Fig. 4. Comparative RMSE performance of ROT.
60 Y. Cai et al. / Decision Support Systems 63 (2014) 52–63

Table 6
Comparison with state-of-the-art methods on MAE.

Training set Methods Given5 Given10 Given15

ML100 SCBPCC 0.874 0.845 0.839

WLR 0.915 0.875 0.890
CBT 0.840 0.802 0.786
ROT 0.807 0.797 0.781
ML200 SCBPCC 0.871 0.833 0.828
WLR 0.941 0.903 0.883
CBT 0.839 0.800 0.784
ROT 0.813 0.771 0.764
ML300 SCBPCC 0.870 0.834 0.819
WLR 1.018 0.962 0.938
CBT 0.840 0.801 0.785
ROT 0.785 0.755 0.753

Fig. 5. Comparative performance of big-error predictions.
training/test splitting ratio of x = 0.9. For the proposed ROT system,
item groups and user groups are computed in advance during the sys-
tem training time. During the recommendation time, the ROT system
only needs to aggregate the typicality degrees of items and users in
order to generate the final recommendation scores. As a result, the
ROT system can efficiently generate the recommendation scores on
the fly. The CB baseline system is the second best system in terms of rec-
ommendation time because it only needs to compute the similarities
between a target item and all the items rated by the current user. Both
the UBCF and the IBCF baseline systems need to identify the “nearest
neighbors” of an item or a user in the entire recommendation space.
As a result, they are not as efficient as the ROT system and the CB sys-
tem. The EMDP baseline is the least efficient system since it needs to
predict the missing data first, and then it combines the ratings obtained
through a user-based CF method and an item-based CF method. For
other training/test splitting ratios, we obtained similar experimental re-
sults. For the reason of brevity, we only report the result under the con-
dition x = 0.9.
4.3.6. Comparison with the state-of-the-art methods
To further evaluate the performance of the proposed ROT system, we
adopted a similar experimentation procedure and data set as reported
in the literature [22]. Table 6 shows the comparative performance of
ROT and other state-of-the-art recommender systems reported in the
literature [22]. For the reason of limited space, we only show the exper-
imental result based on MAE. According to Table 6, ROT outperforms all
state-of-the-art systems under different configurations. In particular,
given both sparse training data (e.g., ML100 and Given5) and relatively
rich training data (e.g., ML300 and Given15), ROT achieves the lowest
MAE scores when compared to state-of-the-art systems such as
SCBPCC, WLR, and CBT.
We also compared the performance of ROT with that of three state-
of-the-art recommender systems, SVD, SVD++, and SocialMF because
these systems achieved very good performance under the larger Netflix
benchmark data set [18]. For this experiment, we utilized the Netflix
benchmark data set and adopted a training/test splitting ratio of x =
0.8. The experimental results that were generated based on 5-fold
cross-validation are plotted in Figs. 7 and 8, respectively. According to
Fig. 7, the proposed ROT system significantly outperforms SVD,
SocialMF, and SVD++ by 5.27% (t(4) = 30.0, p b. 01), 4.45% (t(4) =
73.4, p b. 01), and 3.16% (t(4) = 61.6, p b. 01) in terms of MAE, re-
spectively. Moreover, the ROT system achieves a comparable RMSE per-
formance with that of state-of-the-art systems. Our post-experimental
analysis found that some users were not assigned to any user groups.
Accordingly, the ROT system only generated a default rating for an
item (i.e., the average rating of that item) with respect to these users.
Nevertheless, these recommendations may lead to big-error predic-
tions. Since the metric of RMSE is sensitive to big-error predictions,
the RMSE score achieved by the ROT system is not as good as those
achieved by other state-of-the-art systems. We will explore other
methods (e.g., default user profiling) to specifically deal with the cold-
start problem so that new users can also be classified to some user inter-
est groups. However, this line of research will be conducted as part of
our future work. As a whole, we compare the performance of the pro-
posed ROT system with quite a number of state-of-the-art recommend-
er systems based on two different benchmark data sets. The ROT system
significantly outperforms state-of-the-art systems based on the
MovieLens data set. As for the Netflix data set, the results are mixed.
The ROT system outperforms state-of-the-art systems in terms of

Fig. 6. Comparison of recommendation time. Fig. 7. Comparative MAE performance of ROT based on the Netflix data set.
 Y. Cai et al. / Decision Support Systems 63 (2014) 52–63
Fig. 8. Comparative RMSE performance of ROT based on the Netflix data set.
MAE, but it only achieves a comparable performance with state-of-the-
art systems in terms of RMSE.
5. Discussions
The proposed typicality-based recommendation method is novel
since it can take into account the way how people evaluate objects
with reference to a concept (e.g., a user's specific interest). More specif-
ically, the proposed recommendation method is underpinned by the
principle of object typicality empirically evaluated in the field of cogni-
tive psychology. The proposed method estimates the specific interest
of a user in terms of the typicality of the user within a particular user in-
terest group and the typicality of items that the user group is most inter-
ested in. It addresses the issues of data sparsity and computational
efficiency pertaining to the WoT recommendation environment by
recommending things at the level of item groups and user groups in-
stead of directly evaluating individual items and users on the fly.
These are the reasons why the proposed ROT system outperforms
(e.g., in terms of MAE and computational time) all the baseline systems
that directly operate at individual item and user level under the condi-
tion of sparse training data.
For example, for the MovieLens-based experiments, the ROT system
predicted a rating of 5 for the movie “Usual Suspects, The (1995)”
(movie id 12) and the movie “Pulp Fiction (1994)” (movie id 56) with
respect to the user 99 (a user with benchmark id 99). The user 99 is
very typical (with a typicality degree of 0.9) for the user interest
group “Crime Movies”. Besides, the typicality degrees of “Usual Sus-
pects, The (1995)” and “Pulp Fiction (1994)” for the item group
“Crime Movies” are 1.0 and 0.9517, respectively. Since the user 99 is typ-
ical for the user interest group “crime movie”, and the two movies are
the typical items that the user group likes, the ROT system generates
high ratings for these movies according to the principle of object typical-
ity. In fact, the predictions provided by the ROT system are exactly the
same as user 99's actual ratings for the respective movies.
Basically, all the research questions have been answered through the
two series of benchmark experiments. The number of item groups K has
an impact on the performance of the ROT prototype system. The empir-
ical parameter tuning method that we adopted can identify a reasonable
range of K values for the ROT system. Our experiments also show that
the ROT system is more effective (e.g., producing the least big-errors)
than other baseline systems given sparse training data. Moreover, it is
more efficient than other baseline systems because it mainly operates
at the item group and user group level during the recommendation
time. Finally, the ROT system significantly outperforms other state-of-
the-art recommender systems in terms of MAE based on both the
MovieLens and the Netflix benchmark data sets.
61
6. Conclusions
With the rapid growth of “Web of Things” applications, there is a
pressing need to develop an effective and efficient method for the dis-
covery and selection of smart things on the Web. This paper proposes
a recommender system-based solution to address WoT discovery and
selection, and hence to improve the situation awareness of WoT applica-
tions. Since the recommendation space of WoT is typically sparse, and
computational efficiency as well as users' trusts on recommendations
are also the primary concerns, the main contribution of our research
work is the design of a novel typicality-based recommendation method
that addresses the aforementioned critical issues for WoT recommenda-
tions. As the proposed recommendation method is underpinned by the
“object typicality” principle empirically verified in the field of cognitive
psychology, it tends to produce less big-error recommendations, and
hence it promotes users' trusts on WoT applications. In addition, since
the proposed method exploits data generalization by operating at the
item group and the user group level during recommendation time, it al-
leviates the problems of data sparsity and computational efficiency for
large-scale WoT recommendations.
Based on the benchmark MovieLens data set, our experimental re-
sults confirm that the proposed ROT prototype system significantly out-
performs all the baseline recommender systems even under the
condition of sparse training data. Moreover, the ROT system is more ef-
ficient than other baseline recommender systems and it produces the
least big-error recommendations. Based on the Netflix data set that sim-
ulates a large WoT recommendation space, the ROT system significantly
outperforms state-of-the-art recommender systems in terms of MAE,
and it achieves comparable performance with that of state-of-the-art
recommender systems in terms of RMSE. As a whole, our experimental
results reveal that the proposed ROT prototype system has great poten-
tial to support large-scale WoT discovery and recommendation. To the
best of our knowledge, it is the first successful research work which in-
volves the design and development of a cognitively motivated recom-
mender system for WoT recommendations. The business implication
of our research work is that the proposed recommendation method
can enhance the situation awareness of WoT applications which facili-
tate the reuse of enterprise resources and the interoperability among
enterprises.
For our current experiments, only keyword-based properties (e.g., the
textual descriptions of movies) have been explored to construct item
groups. Future work will explore a richer model of item properties to en-
hance the effectiveness of the proposed ROT system. Another limitation of
our current work is that the number of item groups identified by the ROT
system may not be optimal. More sophisticated methods such as genetic
algorithms will be explored to identify optimal or near optimal clusters of
items. In addition, alternative clustering algorithms and similarity func-
tions will be applied to bootstrap the performance of item group and
user group construction. A hybrid method will be examined to tackle
the cold-start problem such that new users of the system can also be
assigned to some user groups. A comparative study about the recommen-
dation performance of ROT and that of economic choice models will be
performed. Finally, we will explore Big Data Analytics (e.g., MapReduce)
to support parallel incremental clustering of item groups and user groups
to further improve the computational efficiency of the ROT system for
real-time WoT recommendations.
Acknowledgments
The work described in this paper was supported by a grant from
the Research Grants Council of the Hong Kong Special Administra-
tive Region, China (project no. CityU 115910), City University of
Hong Kong (project no. 7008138), the Shenzhen Municipal Science
and Technology R&D Funding — Basic Research Program (project no.
JCYJ20130401145617281), the Shenzhen Research Institute, City University
of Hong Kong, National Natural Science Foundation of China (project no.
62 Y. Cai et al. / Decision Support Systems 63 (2014) 52–63
61300137), the Guangdong Natural Science Foundation of China
(project no. S2011040002222 and S2013010013836), the Funda-
mental Research Funds for the Central Universities, SCUT (project
no. 2012ZM0077).
References
[1] G. Adomavicius, A. Tuzhilin, 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) (2005) 734–749.
[2] Marko Balabanović, Yoav Shoham, Fab: content-based, collaborative recommenda-
tion, Communications of the ACM 40 (3) (1997) 66–72.
[3] Lawrence W. Barsalou, Ideals, central tendency, and frequency of instantiation as
determinants of graded structure in categories, Journal of Experimental Psychology:
Learning, Memory, and Cognition 11 (4) (October 1985) 629–654.
[4] Lawrence W. Barsalou, Cognitive Psychology: An Overview for Cognitive Scientists,
Lawrence Erlbaum As. Pu, 1992.
[5] Yi Cai, H.F. Leung, Multi-prototype concept and object typicality in ontology, Pro-
ceedings of the 21st International Florida Artificial Intelligence Research Society
Conference, AAAI Press, 2008, pp. 470–475.
[6] Jean-Pierre Desclés, Anca Pascu, Logic of determination of objects: the meaning of
variable in quantification, International Journal of Artificial Intelligence Tools 15
(6) (2006) 1041–1052.
[7] Mukund Deshpande, George Karypis, Item-based top-n recommendation algo-
rithms, ACM Transactions on Information Systems 22 (1) (2004) 143–177.
[8] Carolina Fortuna, Matevz Vucnik, Blaz Fortuna, Klemen Kenda, Alexandra Moraru,
Dunja Mladenic, Towards building a global oracle: a physical mashup using artificial
intelligence technology, in: Simon Mayer, Dominique Guinard, Erik Wilde (Eds.),
Third International Workshop on the Web of Things, WoT'12, Newcastle, United
Kingdom, June 19, 2012, ACM, 2012.
[9] K.M. Galotti, K.M. Galotti, Cognitive Psychology In and Out of the Laboratory, third
edition, Wadsworth, Belmont, CA, 2004.
[10] Zeno Gantner, Steffen Rendle, Christoph Freudenthaler, Lars Schmidt-Thieme,
Mymedialite: a free recommender system library, Proceedings of the Fifth ACM
Conference on Recommender Systems, 2011, pp. 305–308.
[11] Dominique Guinard, Mathias Fischer, Vlad Trifa, Sharing using social networks in a
composable Web of Things, Proceedings of the 8th IEEE International Conference
on Pervasive Computing and Communications, IEEE, 2010, pp. 702–707.
[12] Jing He, Yanchun Zhang, Guangyan Huang, Jinli Cao, A smart Web service based
on the context of things, ACM Transactions on Internet Technology (TOIT) 11
(3) (January 2012)(Article 13).
[13] Jonathan L. Herlocker, Joseph A. Konstan, Al Borchers, John Riedl, An algorithmic
framework for performing collaborative filtering, Proceedings of the 22nd Annual
International ACM SIGIR Conference on Research and Development in Information
Retrieval, ACM Press, New York, NY, USA, 1999, pp. 230–237.
[14] Ming Hua, Jian Pei, Ada W. Fu, Xuemin Lin, Ho-Fung Leung, Top-k typicality queries
and efficient query answering methods on large databases, The VLDB Journal 18 (3)
(2009) 809–835.
[15] Mohsen Jamali, Martin Ester, A matrix factorization technique with trust propagation
for recommendation in social networks, Proceedings of the Fourth ACM Conference
on Recommender Systems, RecSys'10, ACM, New York, NY, USA, 2010, pp. 135–142.
[16] Heung-Nam Kim, Inay Ha, Kee-Sung Lee, GeunSik Jo, Abdulmotaleb El-Saddik, Col-
laborative user modeling for enhanced content filtering in recommender systems,
Decision Support Systems 51 (4) (2011) 772–781.
[17] Joseph A. Konstan, Bradley N. Miller, David Maltz, Jonathan L. Herlocker, Lee R.
Gordon, John Riedl, GroupLens: applying collaborative filtering to Usenet news,
Communications of the ACM 40 (3) (1997) 77–87.
[18] Yehuda Koren, Robert M. Bell, Chris Volinsky, Matrix factorization techniques for
recommender systems, IEEE Computer 42 (8) (2009) 30–37.
[19] Raymond Y.K. Lau, Peter D. Bruza, Dawei Song, Towards a belief-revision-based
adaptive and context-sensitive information retrieval system, ACM Transactions on
Information Systems 26 (2) (2008)(Article 8).
[20] Raymond Y.K. Lau, M. Tang, O. Wong, S. Milliner, Y. Chen, An evolutionary learning
approach for adaptive negotiation agents, International Journal of Intelligent Sys-
tems 21 (1) (2006) 41–72.
[21] Marie-Jeanne Lesot, Laure Mouillet, Bernadette Bouchon-Meunier, Fuzzy prototypes
based on typicality degrees, Proceedings of the 8th Fuzzy Days'04, volume 33, Spring-
er, 2004, pp. 125–138.
[22] Bin Li, Qiang Yang, Xiangyang Xue, Can movies and books collaborate? Cross-domain
collaborative filtering for sparsity reduction, in: Craig Boutilier (Ed.), Proceedings of
the Twenty-first International Joint Conference on Artificial Intelligence, 2009,
pp. 2052–2057.
[23] Cen Li, Gautam Biswas, Unsupervised learning with mixed numeric and nomi-
nal data, IEEE Transactions on Knowledge and Data Engineering 14 (2002)
673–690.
[24] Hao Ma, Irwin King, Michael R. Lyu, Effective missing data prediction for collabora-
tive filtering, Proceedings of the 30th Annual International ACM SIGIR Conference
on Research and Development in Information Retrieval, ACM, New York, NY, USA,
2007, pp. 39–46.
[25] Simon Mayer, Dominique Guinard, An extensible discovery service for smart things,
in: Dominique Guinard, Vlad Trifa, Erik Wilde (Eds.), Second International Work-
shop on the Web of Things (WoT 2011), San Francisco, California, June 2011.
[26] D.L. Medin, E.E. Smith, Concepts and concept formation, Annual Review of Psychol-
ogy 35 (1984) 113–138.
[27] Prem Melville, Raymod J. Mooney, Ramadass Nagarajan, Content-boosted collabora-
tive filtering for improved recommendations, Proceedings of the Eighteenth National
Conference on Artificial Intelligence, American Association for Artificial Intelligence,
Menlo Park, CA, USA, 2002, pp. 187–192.
[28] Raymond J. Mooney, Loriene Roy, Content-based book recommending using learn-
ing for text categorization, DL'00: Proceedings of the Fifth ACM Conference on Dig-
ital Libraries, ACM, New York, NY, USA, 2000, pp. 195–204.
[29] Gregory L. Murphy, Gregory L. Murphy, The Big Book of Concepts, MIT Press,
2002.
[30] Nathan Srebro Nati, Tommi Jaakkola, Weighted low-rank approximations, In 20th
International Conference on Machine Learning, AAAI Press, 2003, pp. 720–727.
[31] A. Paterek, Improving regularized singular value decomposition for collabora-
tive filtering, Proceedings of the ACM SIGKDD Cup and Workshop, ACM, 2007,
pp. 39–42.
[32] M.J. Pazzani, D. Billsus, Content-based recommendation systems, The Adaptive
Web: Methods and Strategies of Web Personalization, 2007. 325–341.
[33] Michael Pazzani, Daniel Billsus, Learning and revising user profiles: the identifica-
tion of interesting web sites, Machine Learning 27 (3) (1997) 313–331.
[34] M. Rifqi, Constructing prototypes from large databases, Proceedings of the
Information Processing and Management of Uncertainty Conference, 1996,
pp. 301–306.
[35] Simone Santini, Ramesh Jain, Similarity matching, Proceedings of the Second Asian
Conference on Computer Vision, 1995, pp. 571–580.
[36] Badrul Sarwar, George Karypis, Joseph Konstan, John Reidl, Item-based collaborative
filtering recommendation algorithms, Proceedings of the 10th International World
Wide Web Conference, ACM, New York, NY, USA, 2001, pp. 285–295.
[37] Ian M. Soboroff, Charles K. Nicholas, Combining content and collaboration in text fil-
tering, In Proceedings of the IJCAI'99 Workshop on Machine Learning for Informa-
tion Filtering, 1999, pp. 86–91.
[38] D. Song, R.Y.K. Lau, P.D. Bruza, K.F. Wong, D.Y. Chen, An adaptive information agent
for document title classification and filtering in document-intensive domains, Deci-
sion Support Systems 44 (1) (2008) 251–265.
[39] Jie Tang, Jing Zhang, Limin Yao, Juanzi Li, Li Zhang, Zhong Su, Arnetminer: extraction
and mining of academic social networks, Proceedings of the 14th ACM SIGKDD Inter-
national Conference on Knowledge Discovery and Data Mining, ACM, New York, NY,
USA, 2008, pp. 990–998.
[40] Loren Terveen, Will Hill, Brian Amento, David McDonald, Josh Creter, Phoaks: a
system for sharing recommendations, Communications of the ACM 40 (3) (1997)
59–62.
[41] J.P. Ueberla, An extended clustering algorithm for statistical language models, IEEE
Transactions on Speech and Audio Processing 4 (4) (1996) 313–316.
[42] W. Vanpaemel, G. Storms, B. Ons, A varying abstraction model for categorization,
Proceedings of the 27th Annual Cognitive Science Conference, Lawrence Erlbaum,
Mahwah, NJ, 2005, pp. 2277–2282.
[43] Jun Wang, Arjen P. de Vries, Marcel J.T. Reinders, Unifying user-based and
item-based collaborative filtering approaches by similarity fusion, Proceedings of
the 29th Annual International ACM SIGIR Conference on Research and Development
in Information Retrieval, ACM, New York, NY, USA, 2006, pp. 501–508.
[44] Rui Xu, D. Wunsch, Survey of clustering algorithms, IEEE Transactions on Neural
Networks 16 (3) (2005) 645–678.
[45] Xu. Yunhong, Xitong Guo, Jinxing Hao, Jian Ma, Raymond Y.K. Lau, Wei Xu,
Combining social network and semantic concept analysis for personalized aca-
demic researcher recommendation, Decision Support Systems 54 (1) (2012)
564–573.
[46] Gui-Rong Xue, Chenxi Lin, Qiang Yang, WenSi Xi, Hua-Jun Zeng, Yong Yu, Zheng
Chen, Scalable collaborative filtering using cluster-based smoothing, Proceed-
ings of the 28th Annual International ACM SIGIR Conference on Research and
Development in Information Retrieval, ACM, New York, NY, USA, 2005,
pp. 114–121.
[47] Xin Yan, Raymond Y.K. Lau, Dawei Song, Xue Li, Jian Ma, Toward a semantic granu-
larity model for domain-specific information retrieval, ACM Transactions on Infor-
mation Systems 29 (3) (2011)(Article 15).
[48] L.A. Zadeh, Fuzzy sets, Journal of Information and Control 8 (1965) 338–353.
[49] Zibin Zheng, Hao Ma, Michael R. Lyu, Irwin King, WSRec: a collaborative filtering
based web service recommender system, Proceedings of the 2009 IEEE International
Conference on Web Services, IEEE, 2009, pp. 437–444.
Yi Cai received his Ph.D. degree in computer science from the Chinese University
of Hong Kong. He is currently an associate professor of the School of Software Engi-
neering at the South China University of Technology, Guangzhou, China. His research
interests are recommendation system, personalized search, semantic web and data
mining.
Raymond Y. K. Lau is an associate professor in the Department of Information Systems at
the City University of Hong Kong. He received his Ph.D. degree from the Queensland
University of Technology in 2003. He is the author of over 100 refereed internation-
al journals and conference papers. His research work has been published in re-
nowned journals such as the MIS Quarterly, ACM Transactions on Information
Systems, IEEE Transactions on Knowledge and Data Engineering, INFORMS Journal
on Computing, Journal of MIS, etc. His research interests include information re-
trieval, WoT, and social media analytics. He is a senior member of IEEE and ACM,
respectively.
 Y. Cai et al. / Decision Support Systems 63 (2014) 52–63
Stephen S.Y. Liao is a professor and the director of the Advanced Transportation Information
Systems (ATIS) Research Center at the City University of Hong Kong. He received a bachelor's
degree from Beijing University and a Ph.D. from the University of Aix-Marseille III and Insti-
tute of France Telecom. His research work has been published in journals including the MIS
Quarterly, Decision Support Systems, Communications of the ACM, Information Science, etc.
His current research interests include the use of information technology in mobile commerce
applications and intelligent business systems, especially intelligent transportation systems.
Chunping Li received his Ph.D. degree from Darmstadt University of Technology,
Germany in 1999. He is currently an associate professor in Tsinghua University, China.
His research interests include machine learning, data mining, text classification and topic
modeling. He has published more than eighty research papers in related fields.
63
Ho-fung Leung is currently a professor and the chairman of the Department of Com-
puter Science and Engineering, at the Chinese University of Hong Kong, China. He has
been active in research on intelligent agents, multi-agent systems, game theory, and
semantic web.
Louis C. K. Ma is the acting director of the School of Continuing and Professional
Education at the City University of Hong Kong. He obtained his MBA from the
University of Technology, Sydney, and his Ph.D. from the University of
Warwick, UK. His areas of research interest are in IS policy and strategy, pro-
ject management, and e-markets. His publications have appeared in major IS
journals such as the Journal of MIS, Information & Management, and Decision Sup-
port Systems.