It is difficult to give the prediction to a specific item for the new user cold-start problem because the basic filtering methods in RSs, such as collaborative filtering and content-base
Trang 13
5
7
9
11
13
15
17
19
21
23
25
27
29
31
33
35
37
39
41
43
45
47
49
51
53
55
57
59
61
63
65
67
69
71
73
75
77
79
81
83
85
87
Dealing with the new user cold-start problem
in recommender systems: A comparative review
Q1
VNU University of Science, Vietnam National University, Vietnam
a r t i c l e i n f o
Article history:
Received 15 September 2014
Received in revised form
6 October 2014
Accepted 7 October 2014
Recommended by D Shasha
Keywords:
Collaborative filtering
NHSM
New user cold start
Recommender systems
a b s t r a c t
The Recommender System (RS) is an efficient tool for decision makers that assists in the selection of appropriate items according to their preferences and interests This system has been applied to various domains to personalize applications by recommending items such as books, movies, songs, restaurants, news articles and jokes, among others An important issue for the RS that has greatly captured the attention of researchers is the new user cold-start problem, which occurs when there is a new user that has been registered to the system and no prior rating of this user is found in the rating table In this paper, we first present a classification that divides the relevant studies addressing the new user cold-start problem into three major groups and summarize their advantages and disadvantages in a tabular format Next, some typical algorithms of these groups, such as MIPFGWC-CS, NHSM, FARAMS and HU–FCF, are described Finally, these algorithms are implemented and validated on some benchmark RS datasets under various settings of the new user cold start The experimental results indicate that NHSM achieves better accuracy and computational time than the relevant methods
& 2014 Elsevier Ltd All rights reserved
Contents
1 Introduction 2
2 Literature review 2
3 The analysis of existing methods 5
3.1 MIPFGWC-CS 5
3.2 NHSM 6
3.3 FARAMS 7
3.4 HU–FCF 8
4 Experiments 9
4.1 Environment setup 9
4.2 Results and discussion 9
5 Conclusions 15
Contents lists available atScienceDirect
journal homepage:www.elsevier.com/locate/infosys
Information Systems
http://dx.doi.org/10.1016/j.is.2014.10.001
0306-4379/& 2014 Elsevier Ltd All rights reserved.
n Tel.: þ84 904171284; fax: þ84 438623938.
E-mail addresses: sonlh@vnu.edu.vn , chinhson2002@gmail.com
1
Official address: 334 Nguyen Trai, Thanh Xuan, Hanoi, Vietnam.
Trang 23
5
7
9
11
13
15
17
19
21
23
25
27
29
31
33
35
37
39
41
43
45
47
49
51
53
55
57
59
61
63
65
67
69
71
73
75
77
79
81
83
85
87
89
91
93
95
97
99
101
103
105
107
109
111
113
115
117
119
121
123
Acknowledgments 17
Appendix A Supporting information 17
17
References 17
1 Introduction
The growing development of content-based systems
that provide a large amount of data, such as videos,
images, blogs, multimedia, and wikis, brings great
chal-lenges for analysts attempting to extract useful knowledge
and capture meaningful events from the massive data
Machine learning tools should indeed be oriented to what
users intend to do and how they want the results to be
returned in a given format An efficient tool that assists
decision makers to choose appropriate items according to
their preferences and interests and that is currently widely
used is the Recommender System (RS) Ricci et al [31]
defined the RS as a special type of information system that
(i) helps to make choices without sufficient personal
experience of the alternatives, (ii) suggests products to
customers, and (iii) provides consumers with information
to help them decide which products to purchase The RS is
based on a number of technologies, such as information
filtering, classification learning, user modeling and
adap-tive hypermedia, and it is applied to various domains to
personalize applications by recommending items such as
books, movies, songs, restaurants, news articles and jokes,
among others It has been applied to e-commerce to learn
from a customer and recommend products that he or she
will find most valuable from among the available products,
thus helping the customer find suitable products to
purchase Some e-commerce RSs are named as follows
[37,22] For example, Amazon.com is the most famous
e-commerce RS, structured with an information page for
each book while providing details of the text and purchase
information Two recommendations are found herein,
including books frequently purchased by customers who
purchased the selected book and authors whose books are
frequently purchased eBay.com is another example that
provides the Feedback Profile feature, which allows both
buyers and sellers to contribute to the feedback profiles of
other customers with whom they have done business The
feedback consists of a satisfaction rating and a specific
comment about the other customer On Moviefinder.com,
customers can locate movies with a similar“mood, theme,
genre or cast” through Match Maker or by their previously
indicated interests through WePredict We clearly
recog-nize that RSs are becoming important and with increasing
influence on various practical applications
An important issue for RSs that has greatly captured the
attention of researchers is the cold-start problem This
pro-blem has two variants: the new user cold-start propro-blem and
the new item cold-start problem The new item cold-start
problem occurs when there is a new item that has been
transferred to the system Because it is a new product, it has
no user ratings (or the number of ratings is less than a
threshold as defined in some equivalent papers) and is
therefore ranked at the bottom of the recommended items
list Moreover, this problem can be partially handled by staff
members of the system providing prior ratings to the new item Thus, the concentration of the cold-start problem is dedicated to the new user cold-start problem when no prior rating could be made due to the privacy and security of the system It is difficult to give the prediction to a specific item for the new user cold-start problem because the basic filtering methods in RSs, such as collaborative filtering and content-based filtering, require the historic rating of this user
to calculate the similarities for the determination of the neighborhood For this reason, the new user cold-start problem can negatively affect the recommender performance due to the inability of the system to produce meaningful recommendations [33] Addressing this problem has been the primary focus of various studies in recent years
The aim of this paper is to provide a comparative review
of those studies that could answer our research question
“which (group of) algorithm is the most effective among all?” For this purpose, we first provide a classification that divides the relevant studies into three groups: (i) makes use of additional data sources; (ii) selects the most prominent groups of analogous users; and (iii) enhances the prediction using hybrid methods A table that sum-marizes the advantages and disadvantages of all groups of methods is presented Second, some typical algorithms of the groups of methods, such as MIPFGWC-CS[46](the first group), NHSM[20](the second group), FARAMS[17]and
HU–FCF [42] (the third group), are described in detail Finally, these algorithms are implemented and validated
on some benchmark RS datasets, such as MovieLens[23]
and Jester[12], under various settings of the new user cold start The experimental results could reveal the answer for our research question stated above
The remainder of the paper is organized as follows In
Section 2, we present a literature review of the relevant studies according to the three aforementioned groups
Section 3elaborates on the four typical methods, namely, MIPFGWC-CS[46], NHSM[20], FARAMS[17]and HU–FCF
[42].Section 4 presents the comparative experiments of these algorithms involving benchmark RS datasets Finally,
research directions
2 Literature review
The beginning of this section starts with an example that clearly demonstrates the new user cold-start problem Example 1 We have a RS that includes three tables: the users' demographic data (Table 1), the movies' information (Table 2) and the rating (Table 3) This type of system is able
to predict the user rating of a movie, which is expressed in
Table 3 Nonetheless, the new user cold-start problem occurs with a new user, e.g., Kim (User ID: 6) inTable 1, who has no prior rating such that it is difficult to provide a prediction for the first movie, e.g., Titanic (ID: 1)
Trang 33
5
7
9
11
13
15
17
19
21
23
25
27
29
31
33
35
37
39
41
43
45
47
49
51
53
55
57
59
61
63
65
67
69
71
73
75
77
79
81
83
85
87
89
91
93
95
97
99
101
103
105
107
109
111
113
115
117
119
121
123
In the following, we briefly summarize the relevant
works in regards to the new user cold-start problem At the
milestone of 2014, there are various works aiming to
handle this problem Those studies could be divided into
three categories: (i) makes use of additional data sources,
(ii) chooses the most prominent groups of analogous users,
and (iii) enhances the prediction using hybrid methods
a) The principal idea of the first group is the use of some
additional sources, such as the demographic data (a.k.a
the users' profile), the users' opinions, and social tags,
for a better selection of the neighbors of the new user
Vozalis and Margaritis [49] demonstrated a modified
version of k-nearest neighborhood by adding a user
demographic vector to the user profile and embedding
it in the collaborative filtering algorithm for the
calcu-lation of similarity Poirier et al [27] proposed a
method that exploits blog textual data to reduce the
cold-start problem by labeling subjective texts
accord-ing to their expressed opinions to construct a user–
item-rating matrix and establishing recommendations
through collaborative filtering Zhang et al [53]
pre-sented a recommendation algorithm that makes use
of social tags, particularly user-tag-object tripartite
graphs, to provide more personalized
recommenda-tions when the assigned tags belong to diverse topics
Almazro et al [3] introduced a hybrid demographic-based and collaborative filtering approach on the movie domain using demographic data to enhance the recom-mendation suggestion process Their method classified the genres of movies based on demographic attributes, e.g., user age (child, teenager or adult), student (yes or no), have children (yes or no) and gender (female or male) Preisach et al [28] argued that many user profiles contain untagged resources that could pro-vide valuable information, especially for the cold-start problem, and proposed a purely graph-based semi-supervised relational approach that uses untagged posts Said et al [34,36] modified the user similarity calculation method to employ the hybridization of demographic and collaborative approaches A modifi-cation to the k-nearest neighborhood that calculates the similarity scores between the target user and other users was introduced Wang et al [50] introduced Credible and co-clustering filterBot for cold-stArt reco-mmendations (COBA), which uses the rating confidence level to reduce the dimensionality of the item–user matrix The items and users were co-clustered, and the ratings within every user cluster were smoothed to overcome data sparsity The recommendations were fused from item and user clusters to predict user preference Zhang et al.[52] proposed the Cold-start Recommendations Using Collaborative Filtering (CRUC) scheme, which involves formulation, filtering and pre-diction steps They assumed that users are tracked by sensors such that each user has their own location, which is currently regarded as the item The item–user matrix was normalized and clustered to identify users who have a significant influence on the recommenda-tion The prediction steps were performed by taking the hybrid between the item-based and user-based filtering methods Chen et al.[8]employed additional informa-tion, such as the social sub-community and an ontology decision model, to assist the recommendation in the cold-start problem The social sub-community was divided according to the exiting users' history data and the mining relationship between each other An ontology decision model was then constructed on the basis of sub-community and users' static information, which makes recommendations for the new user based
on his static ontology information Guo[11]proposed three different approaches from the perspective of preference modeling First, the ratings of trusted neigh-bors were merged to form a new rating profile for the active users based on which better recommendations can be generated Second, a novel Bayesian similarity measure was introduced by taking both the direction and length of rating vectors into account Third, a new information source called prior ratings, based on virtual product experience in virtual reality environments, was proposed to inherently resolve the concerned pro-blems Chen et al.[7]proposed a cold start recommen-dation method for the new user that integrates a user model with trust and distrust networks to identify tru-stworthy users, which are then aggregated to provide useful recommendations for new users Demographic
Table 3
Rating data.
Table 1
Users' demographic data.
Table 2
Movies' information.
Trang 43
5
7
9
11
13
15
17
19
21
23
25
27
29
31
33
35
37
39
41
43
45
47
49
51
53
55
57
59
61
63
65
67
69
71
73
75
77
79
81
83
85
87
89
91
93
95
97
99
101
103
105
107
109
111
113
115
117
119
121
123
data or users' profiles are the most common additional
source for solving the cold-start problem Safoury and
Salah [33] presented a framework for evaluating the
influence of demographic attributes on the user ratings
This framework was examined using a movie dataset to
evaluate the accuracy and precision of the generated
recommendations Nazir and Yadav [24]introduced a
profile-based approach that consisted of three main
phases: Fetch, Process and Truncate The Fetch phase is
concerned with obtaining the required parameters,
such as Global Student Profile or Shared Profile (Profile
Token), for the Process phase Process is the main
engine where the actual recommendations are
gener-ated based on inputs from the Fetch phase Truncate is
involved with the discarding of the profile-tokens used
during the Fetch phase Formoso et al.[9]proposed a
novel profile-expansion approach that includes three
types of techniques, namely, item-global, item-local
and user-local, based on the query expansion
techni-ques in information retrieval The experimental
evalua-tion showed that both item-global and user-local offer
outstanding improvements in precision Son et al.[46]
presented a novel filtering method based on fuzzy
geographically clustering [45,38–44], the so-called
MIPFGWC-CS, that can handle the issues of selected
demographic attributes, the similarities between items
and missing ratings that existed in relevant
demogr-aphic-based algorithms Rosli et al.[32]designed a new
measure by combining similarity values obtained from
a movie “Facebook Page” First, the users' similarities
were computed according to the rating cast on the
Movie Rating System Then, the similarity values
obta-ined from a user's genre interest in“Like” information
extracted from “Facebook Pages” were combined
Finally, all of the similarity values were integrated to
produce a new user's similarity value Lika et al [18]
proposed a model that incorporates classification
met-hods with demographic data for the identification of
other users with similar behaviors
Limitations of the first group: although the additional
data sources are necessary, we sometimes do not have
these types of data for the selection, e.g., in some
e-shopping systems when users do not record their
profiles and associated Facebook/Twitter accounts
b The idea of the second group is to improve the methods
that determine the analogous users without the aid of
additional data sources Ahn[2]addressed the limitations
of the existing methods for the new user cold-start
problem by primarily focusing on the similarity measures,
such as the Pearson coefficient and the cosine measure,
and proposed a heuristic similarity measure, i.e., the
so-called PIP (Proximity–Impact–Popularity) measure The
Proximity factor is based on the arithmetic difference
between two ratings, the Impact factor considers how
strongly an item is preferred or disliked by buyers, and
the Popularity factor provides greater value to a similarity
for ratings that are further from the average rating of a
co-rated item Lam et al.[15]discussed a hybrid model
based on the analysis of two probabilistic aspect models
using pure collaborative filtering to combine with users'
information Sun et al.[47]clustered users based on the
user–item rating matrix and then utilized the clustering results and users' demographic information to construct a decision tree to achieve the associations between the existing users and the new users The predictions for new users were made by combining the decision tree with the collaborative filtering algorithm Zhou et al [54] pre-sented functional matrix factorization (fMF), a novel cold-start recommendation method that constructs a decision tree with each node being a question fMF enables the recommender to query a user adaptively according to her prior responses and associates latent profiles for each node of the tree to gradually refine the profiles It also consists of an iterative optimization scheme that alter-nates between decision tree construction and latent profile extraction Qiu et al [29] introduced an item-oriented function and incorporated it with a hybrid algorithm between heat conduction and the probability spreading process so that the proposed algorithm does not require any additional information, such as tag Liu
et al [21] noted that the existing recommendation methods lacked a principled model for guiding how to select the most useful ratings and that ratings on the selected representatives are considerably more useful for making recommendations; thus, they proposed a princi-ple approach to identify representative users and items using representative-based matrix factorization Bobadilla
et al [5] presented a new similarity measure using optimization based on neural learning, which exceeds the best results obtained with current metrics, and described the mathematical formalization that shows how to obtain the main quality measures of a recom-mender system using leave-one-out cross validation Said
et al.[35]performed a set of tests to identify whether the weighting schemes on three common similarity mea-sures using two different movie datasets can be beneficial for the purpose of overcoming problems related to cold-start, as well as profiling users to generate more accurate profiles not based on the most popular items They claimed that the weighting schemes appear to have little effect on datasets with a wide rating scale and high concentration of ratings on popular items Moreover, the cosine measure is very insignificantly affected by any weighting measure and produces identical results regard-less of whether weighting is applied Sun et al [48]
proposed a novel algorithm that learns to conduct the interview process guided by a decision tree with multiple questions at each split The splits, represented as sparse weight vectors, are learned through an L_1-constrained optimization framework The users are directed to child nodes according to the inner product of their responses and the corresponding weight vector A linear regressor is learned within each node, using all previously obtained answers as inputs to predict item ratings Liu et al.[20]
presented a new user similarity model– NHSM – that takes into account the global preference of user behaviors
in addition to the local context information of user ratings
to improve the recommendation performance in the cold-start situation
Limitations of the second group: how to choose the optimal number of groups and the splitting criteria
is worth considering
Trang 53
5
7
9
11
13
15
17
19
21
23
25
27
29
31
33
35
37
39
41
43
45
47
49
51
53
55
57
59
61
63
65
67
69
71
73
75
77
79
81
83
85
87
89
91
93
95
97
99
101
103
105
107
109
111
113
115
117
119
121
123
c After determining the most analogous users to the new
one, some authors used hybrid methods for the
calcula-tion of similarity and/or the prediccalcula-tion of rating This is
the basic idea of the third group Leung et al [16,17]
introduced a collaborative filtering framework based on
Fuzzy Association Rules and Multiple-level Similarity
(FARAMS), which extends existing techniques by using
fuzzy association rule mining and takes advantage of
product similarities in taxonomies to address data
spar-seness and non-transitive associations Basiri et al [4]
proposed a hybrid recommender system using the
opti-mistic exponential type of an ordered weighted averaging
operator to fuse the output of recommender system
strat-egies for the new user cold-start problem Kim et al.[13]
presented a method for the cold-start problem that
includes the prediction of actual ratings, the identification
of prediction errors for each user, and the construction of
an error-reflected model for the prediction of new users
or items Ge and Ge[10]claimed that a lower-rank
appro-ximation could remove data noise resulting from
uns-table user behaviors and thus lead to better
recommen-dation quality; based upon this idea, they proposed
Singular Value Decomposition-based Collaborative
Filter-ing Kim et al.[14]proposed three hybrid recommenders
based on user similarity and two content-boosted
recom-menders used in conjunction with interaction-based
col-laborative filtering; they experimentally showed that the
best hybrid and content-boosted recommenders improve
on the interaction-based collaborative filtering method
Quijano-Sánchez et al.[30]extended a group
recommen-der system with a case based on previous group
recom-mendation events Carrer-Neto et al.[6]presented a
hyb-rid recommender system based on knowledge and social
networks Negre et al.[25]introduced a process for
solv-ing the cold-start problem in cases of data warehouses
composed of four steps: patternizing OLAP queries,
pre-dicting candidate operations, computing candidate
reco-mmendations and ranking these recoreco-mmendations Xie
et al [51] proposed Elver, which employs an iterative matrix completion technology and a nonnegative factor-ization procedure to work with meager content inklings
to recommend and optimize page-interest targeting on Facebook Aharon et al.[1]introduced a recommendation algorithm based on Latent Factor analysis called One-pass Factorization of Feature Sets (OFF-Set), which is able to model non-linear interactions between pairs of features and updates its model per each recommendation-reward observation in a pure online fashion Lin et al.[19] des-cribed a method that considers the nascent information culled from Twitter to provide relevant recommendations
in cold-start situations Nilashi et al.[26]proposed new recommendation methods using ANFIS and SOM clust-ering A hybrid user-based fuzzy collaborative filtering method has been proposed by Son[42]
Limitations of the third group: irrelevant users are still included in the computation of similarities
d Table 4summarizes the relevant works by groups along with their advantages and disadvantages
3 The analysis of existing methods
In this section, we provide more details about the typical algorithms of the groups of methods inTable 4to address the new user cold-start problem These algorithms are MIPFGWC-CS[46], NHSM[20], FARAMS[17]and HU– FCF[42], and these methods are presented in the subse-quent sections
3.1 MIPFGWC-CS
The basic concept of the MIPFGWC-CS algorithm[46]is
to use fuzzy geographically clustering[45,38–44], particu-larly MIPFGWC, for the determination of similar users with respect to all attributes in the demographic data Because
Table 4
Groups of methods for the new user cold-start problem.
algorithms
Makes use of
additional data
sources
Use some data (users' profile, opinions, social tags) to support the selection of the neighbors of the new user
MIPFGWC-CS [46]
Determination of analogous users is more accurate
Additional data are sometimes not available
Chooses the most
prominent groups
of analogous users
Improve the methods determining the analogous users by clustering algorithms, decision trees
NHSM [20]
The similarity degrees between users are enhanced
Additional data are not required
How to choose the optimal number
of groups and the splitting criteria
is worth considering
Enhances the
prediction using
hybrid methods
Use hybrid methods for the calculation of similarity and/or the prediction of ratings
FARAMS [17]
HU–FCF [42]
Utilizes the results
of existing methods for prediction
Controls the final results by parameters
Specification of values of parameters is hard
Irrelevant users are still included in the computation of similarities
Trang 63
5
7
9
11
13
15
17
19
21
23
25
27
29
31
33
35
37
39
41
43
45
47
49
51
53
55
57
59
61
63
65
67
69
71
73
75
77
79
81
83
85
87
89
91
93
95
97
99
101
103
105
107
109
111
113
115
117
119
121
123
the new user has no prior rating, the demographic data are
the only medium to calculate the similarities between
users After finding users similar to the new one,
MIPFGWC-CS checks whether they rated the considered
item or not If ratings are found, then they are considered
to be the representative ratings of users Otherwise, a
similar item to the considered one is found by the Pearson
coefficient, and the rating on the similar item is assumed
to be the representative rating Finally, the rating of the
new user to the considered item is approximated by the
weighted average operator of the representative ratings
Fig 1andTable 5illustrate the idea in detail
As described above, the MIPFGWC-CS algorithm
con-tains several disadvantages, as follows:
a) Determining the optimal number of clusters for
MIPFGWC is required before running the clustering
algorithm Although other parameters of MIPFGWC
were suggested by Son et al.[39], how to determine
the optimal number of clusters is still an on-going topic
of research The exact number of clusters would lead to
more accurate results for finding the similar users to a
new user and thus enhance the prediction accuracy
b) InFig 1, finding a similar item to the considered one by
the Pearson coefficient could somehow not achieve
good results because the Pearson metric has some
limitations where there is a poor signal-to-noise ratio
and negative spikes In other words, if the relationship
between two variables is non-linear, the Pearson
coef-ficient cannot accurately measure the correlation
c) The MIPFGWC-CS relies solely on the demographic
data (Fig 1) If this type of data is not available, the
algorithm cannot be performed
3.2 NHSM
Liu et al.[20]introduced a new similarity metric called
NHSM to replace the traditional Pearson coefficient or the
cosine similarity measure This heuristic similarity
mea-sure is composed of three factors of similarity, which are
Proximity, Significance and Singularity Proximity
consid-ers the distance between two ratings Significance shows
that the ratings are more significant if the two ratings are
more distant from the median rating Singularity
repre-sents how the two ratings are different from other ratings
Furthermore, NHSM integrates the modified Jaccard and
the user rating preference in the design The definition of
NHSM is stated below
sim uð ; vÞNHSM¼ sim u; vð ÞJPSSsim uð ; vÞURP; ð8Þ
sim uð ; vÞURP
þexp jμ uμvjjσuσvj; ð9Þ sim uð ; vÞJPSS¼ sim u; vð ÞPSSsim uð ; vÞJaccard; ð10Þ
sim uð ; vÞJaccard
¼jIu\ Ivj
sim uð ; vÞPSS¼XpA IProximity ru;p; rv;p
Signif icance ru;p; rv;p
Singularity ru;p; rv;p
; ð12Þ Proximity ru ;p; rv ;p
¼ 11þexp jr1
u ;prv ;pj
Signif icance ru ;p; rv ;p
¼1þexp jr 1
u ;prmedjjrv ;prmedj
ð14Þ
Fig 1 The MIPFGWC-CS algorithm.
Trang 73
5
7
9
11
13
15
17
19
21
23
25
27
29
31
33
35
37
39
41
43
45
47
49
51
53
55
57
59
61
63
65
67
69
71
73
75
77
79
81
83
85
87
89
91
93
95
97
99
101
103
105
107
109
111
113
115
117
119
121
123
Singularity ru ;p; rv ;p
1þexp ru;pþ r v;p
2 μp
where μu and σu are the mean rating and the standard
variance of user u, respectively Iu represents the set of
ratings of user u The operator means the common
ratings between two users ru;pis the rating of user u on
item p rmed is the median value in the rating scale.Fig 2
describes the filtering method using the NHSM metric The
limitations of the NHSM-based filtering algorithm as
follows:
a) The algorithm is based solely on the rating data and
makes no use of additional data, such as demographic
data; thus, it somehow leads inaccurate calculations of
the similarity
b) The algorithm must assume that the new user has rated
some prior rating in the rating data
3.3 FARAMS
Leung et al [17] integrated fuzzy sets theory into
association rule mining techniques and applied the
pro-posed work to the collaborative filtering of recommender
systems First, the rating data are converted to the
transac-tional database of association rule mining, fuzzified by
fuzzy memberships of linguistic variables and transformed into the type of transaction ID (TID)– items where each TID is in the form of {Item, linguistic variable}, and each item is a list of users with equivalent fuzzy memberships that opted for the {Item, linguistic variable} Then, an Apriori-like algorithm is used to define candidate item sets and possible rules with the support of MinSupp and MinConf thresholds The difference between this algo-rithm and the original Apriori algoalgo-rithm is the use of Fuzzy Support – FChh iA;X ; B;Y h ii and Fuzzy Confidence
FChh iA ;X ; B;Y h ii between two items A and B equipped by their memberships X and Y respectively (Eqs (16–21)) After defining the fuzzy rules, the predicting score of a recom-mendable item is calculated and used to provide the final rating of the new user Fig 3 highlights the concept in detail
FSh i ¼A;X
P
tiA TΠajA Aμðti½ajÞ
FCh iA;X ; B;Y h i¼FShFSA[ B;X [ Yi
A ;X
CORRh iA;X ; B;Y h i¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiCovh iA;X; B;Yh i
Varh iA;X nVarhB;Yi
Table 5
The pseudo-code of the MIPFGWC procedure.
Input Geo-demographic data X The number of elements (clusters) – NðCÞ The dimension of dataset r Threshold ε and other parameters m; η; τ, a i
ði ¼ 1; 3Þ, γ j ðj ¼ 1; C Þ Geographic parameters α; β; γ; a; b; c; d.
Output Final membership values u 0
k and centers Vðtþ 1Þ MIPFGWC
1: Set the number of clusters C, thresholdε40 and other parameters such as m; η; τ41, a i 40 ði ¼ 1; 3Þ, γ j ðj ¼ 1; C Þ as in [39]
2: Initialize centers of clusters Vj, j¼ 1; C at t ¼ 0
3: Set geographic parameters α; β; γ; a; b; c; d satisfying condition (1)
4: Use the formulas to calculate the membership values, the hesitation level and the typicality values, respectively
P C
i ¼ 1
‖X k V j ‖
‖X k V i ‖
! 2 m 1 ; k ¼ 1; N; j ¼ 1; C ;
ð2Þ
P C i¼ 1
‖X k V j ‖
‖X k V i ‖
τ 1 ; k ¼ 1; N; j ¼ 1; C ;
ð3Þ
1þ a 2 ‖X k V j ‖ 2
γ j
η 1 ; k ¼ 1; N; j ¼ 1; C:
ð4Þ 5: Perform geographic modifications through Eqs (5 – 6 )
u 0
k ¼ α u k þβ Xk 1
j ¼ 1
wkj u 0
j þγ 1
A XC
j ¼ k
wkj u j ;
ð5Þ
w kj ¼
pop k pop j
p c
kj IM d kj
d a
kj kaj
: 8
k
is a completely monotone increasing sequence or ukZu 0
k for most k ¼ 1; C , then conclude that there is no suitable solution for the given geographic parameters Otherwise, go to Step 7.
7: Calculate the centers of clusters at tþ1 by Eq (7)
V j ¼
P N k¼ 1 a 1 u m þa 2 tηkjþa 3 hτkj
X k
P N
k ¼ 1 a 1 u m þa 2 tηkjþa 3 hτkj ; j ¼ 1; C :
ð7Þ 8: If the differencejjV ð tþ 1 Þ V ð Þ t jjrε, then stop the algorithm Otherwise, assign V ð Þ t ¼ V ð t þ 1 Þ and return to Step 4.
Trang 83
5
7
9
11
13
15
17
19
21
23
25
27
29
31
33
35
37
39
41
43
45
47
49
51
53
55
57
59
61
63
65
67
69
71
73
75
77
79
81
83
85
87
89
91
93
95
97
99
101
103
105
107
109
111
113
115
117
119
121
123
Covh iA;X; B;Y h i¼ FShA [ B;X [ Yi FSh iA ;X nFSh B ;Y i; ð19Þ
Varh i ¼ FSA;X h iA;X 2FSh iA;X 2; ð20Þ
FSh iA;X2¼
P
t i A TnΠajA Aμðti½ajÞo2 T
In Eq.(16), A; X represents an〈Itemset, FuzzySet〉 ti½aj is
the value of aj in the ith record of the transactional
database T μðti½ajÞ is the membership value of ti½aj
Eqs.(16–18) provide the formulas of Fuzzy Supports, Fuzzy
Confidence and Correlation, respectively The limitations of
the FARAMS algorithm are as follows:
a) The fuzzification in FARAMS could lead to inaccurate
prediction results The FARAMS algorithm was designed
for movie applications, e.g., MovieLens, Jester and
Each-Movie, where the linguistic variables are“Like”, “Dislike”
and “Neutral” with pre-defined membership functions
When applied to other applications, knowing how to set
up the membership functions is a matter of concern
Wrong membership values would result in the activities
of the algorithm In fact, not all recommender system
applications require fuzzy parameters; thus, for the sake of
stability and processing time, the fuzzification step should
be reduced
b) The limitation of rating data in the NHSM-based filter-ing algorithm is available
c) The FARAMS algorithm could be regarded as an effi-cient method for calculating the similarity betw-een items
3.4 HU–FCF The basic concept of the HU–FCF method [42] is to integrate the fuzzy similarity degrees between users based
on the demographic data, with the hard user-based degrees calculated from the rating histories integrated into the final
Fig 2 The NHSM-based filtering algorithm.
Fig 3 The FARAMS algorithm.
Trang 93
5
7
9
11
13
15
17
19
21
23
25
27
29
31
33
35
37
39
41
43
45
47
49
51
53
55
57
59
61
63
65
67
69
71
73
75
77
79
81
83
85
87
89
91
93
95
97
99
101
103
105
107
109
111
113
115
117
119
121
123
similarity degrees As such, those degrees would reflect more
exactly the correlation between users in terms of the internal
(attributes of users) and external information (interactions
between users) Each similarity degree (fuzzy/hard) is
accom-panied by weights automatically calculated according to the
numbers of analogous users After the final similarity degrees
are calculated, the final rating will be constructed based on
the rating values of neighbors of the considered user
Depend-ing on the domain of a specific problem, the final ratDepend-ing will
be approximated to its nearest value in that domain
accom-panied by an error threshold, which is normally less than 5%
A list of nearest values with equivalent error thresholds is also
given as the prediction ratings of a user for an item Fig 4
illustrates the concept in detail
The limitations of the HU–FCF algorithm are as follows:
a) If the demographic data are not provided in the data list,
the HU–FCF algorithm does not work because the rating
data have no prior ratings of the new user Thus, the
similarities between the new user and others cannot be
calculated, and the final rating cannot be found
b) Similar to the deficiencies of the MIPFGWC-CS
algo-rithm, the Pearson coefficient cannot accurately
mea-sure the correlation Thus, a better similarity metric
should be used instead of the Pearson coefficient
c) In the branch of demographic data, the GFD matrix is
calculated from all users in the system Indeed, irrelevant
users may be included in the computation of similarities,
thus degrading the performance of the prediction
4 Experiments
4.1 Environment setup
Experimental tools: we have implemented MIPFGWC-CS
[46], NHSM[20], FARAMS[17]and HU–FCF[42]in the C
programming language and executed them on a PC with an Intel Pentium 4 CPU 2.66 GHz, 1 GB RAM, and
80 GB HDD
Experimental datasets: we use the following benchmark
RS datasets
MovieLens 1 M[23]: contains 1,000,209 anonymous ratings of approximately 3900 movies provided by
6040 MovieLens users Ratings are discrete values from 1 to 5 Demographic data are provided in the following form: “Gender: Age: Occupation: Zip-code”
Jester[12]: contains ratings of 100 jokes from 73,421 users Ratings are real values ranging from 10 to
10 The value “99” corresponds to “null”¼“not rated” Demographic data are no longer supported for this dataset
Generating cold-start users: we adopt the Hold-out and the k-fold cross validation methods, where the users in the testing set are the start users For each cold-start user, we use those algorithms to predict the ratings for items that have been rated, except three rated items selected to be the basis for the calculation
of similarity Each trial is measured by the evaluation indices The final results are computed as the average value of those according to users and trials
Evaluation indices: we use the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) for the validation of accuracy
MAE¼N1X
u ;i
pu;iru ;i
RMSE¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1
N X
u ;i
pu ;iru;i
s
where pu;iðru ;iÞ is the predicted (real) rating of user u for item i
Experimental objectives: we compare the accuracy and the computational time of the algorithms to determine the most effective algorithm
4.2 Results and discussion
First, we present the comparative results of the algo-rithms using the Hold-out cross validation method described
Fig 4 The HU–FCF algorithm.
Table 6 The experimental results using the Hold-out cross validation method.
MAE values
0.697
RMSE values
Computational time (s)
a
Smallest value for a given dataset.
Trang 103
5
7
9
11
13
15
17
19
21
23
25
27
29
31
33
35
37
39
41
43
45
47
49
51
53
55
57
59
61
63
65
67
69
71
73
75
77
79
81
83
85
87
89
91
93
95
97
99
101
103
105
107
109
111
113
115
117
119
121
123
in Table 6 The experimental results are evaluated by the
evaluation indices In this table, the results of MIPFGWC-CS
on the Jester dataset are null because this dataset does not
support the demographic data, which are essential for the
calculation in MIPFGWC-CS The experimental results have
clearly shown that the MAE, RMSE values and the
computa-tional time of NHSM are mostly smaller than those of the
other algorithms To visualize the experimental results, the
MAE and RMSE values of the algorithms on the MovieLens
and Jester datasets are presented inFig 5and6, respectively
It is clear that the MAE and RMSE values of NHSM are
the smallest among all of the algorithms For instance, the
MAE value of NHSM on the Jester dataset is 0.821, which approximates to 91.3% and 91.7% of those of FARAMS and
HU–FCF, respectively Similarly, the RMSE value of NHSM
on the Jester dataset is 1.04, which approximates to 95.3% and 94.7% of those of FARAMS and HU–FCF, respectively The only case in which the MAE value of NHSM is larger than those of other algorithms occurs on the MovieLens dataset with MAE values of NHSM, MIPFGWC-CS, FARAMS and HU–FCF being 0.641, 0.701, 0.636 and 0.697, respec-tively Despite this fact, the MAE value of NHSM is only larger than that of FARAMS and is smaller than those
of other algorithms Thus, the experimental results have
Fig 5 The MAE and RMSE values of algorithms on the MovieLens dataset.
Fig 6 The MAE and RMSE values of algorithms on the Jester dataset.