See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/338150522

A Posterior-Neighborhood-Regularized Latent Factor Model for Highly

Accurate Web Service QoS Prediction

Article  in  IEEE Transactions on Services Computing · December 2019

DOI: 10.1109/TSC.2019.2961895

CITATIONS READS

105 119

6 authors, including:

Di Wu Qiang He
Chinese Academy of Sciences Swinburne University of Technology
46 PUBLICATIONS   1,882 CITATIONS    222 PUBLICATIONS   4,860 CITATIONS   

SEE PROFILE SEE PROFILE

Xin Luo Ming Sheng Shang

Chongqing University Chongqing Institute of Green and Intelligent Technology, Chinese Academy of Sci…
90 PUBLICATIONS   4,679 CITATIONS    134 PUBLICATIONS   5,031 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Service Computing View project

Edge Data Caching View project

All content following this page was uploaded by Qiang He on 23 July 2021.

The user has requested enhancement of the downloaded file.

 IEEE TRANSACTIONS ON SERVICES COMPUTING 1

1 A Posterior-Neighborhood-Regularized Latent
2 Factor Model for Highly Accurate Web
3 Service QoS Prediction
4 Di Wu, Member, IEEE, Qiang He , Senior Member, IEEE, Xin Luo , Senior Member, IEEE,
5 Mingsheng Shang , Yi He, and Guoyin Wang , Senior Member, IEEE

Abstract—Neighborhood regularization is highly important for a latent factor (LF)-based Quality-of-Service (QoS)-predictor since similar

of
6
7 users usually experience similar QoS when invoking similar services. Current neighborhood-regularized LF models rely prior information
8 on neighborhood obtained from common raw QoS data or geographical information. The former suffers from low prediction accuracy
9 due to the difficulty of constructing the neighborhood based on incomplete QoS data, while the latter requires additional geographical
10 information that is usually difficult to collect considering information security, identity privacy, and commercial interests in real-world
11 scenarios. To address the above issues, this work proposes a posterior-neighborhood-regularized LF (PLF) model for QoS prediction.
12 The main idea is to decompose the LF analysis process into three phases: a) primal LF extraction, where the LFs are extracted to
13 represent involved users/services based on known QoS data, b) posterior-neighborhood construction, where the neighborhood of each
14 user/service is achieved based on similarities between their primal LF vectors, and c) posterior-neighborhood-regularized LF analysis,
15
16

17

18 1
ro
where the objective function is regularized by both the posterior-neighborhood of users/services and L2 -norm of desired LFs. Experimental
results from large scale QoS datasets demonstrate that PLF outperforms state-of-the-art models in terms of both accuracy and efficiency.

Index Terms—Web service, quality-of-service, latent factor analysis, posterior-neighborhood, regularization, cloud computing, big data

INTRODUCTION
Ç
EP
Services are the fundamental components for cloud such circumstances, it is a challenge for potential users to
19
20
21
W EB
computing-based software applications [1]. They are
designed for easy exchange of data among software applica-
select appropriate web services, especially when many avail-
able candidate web services have highly similar or even
26
27
28
22 tions over the World Wide Web [2], [3], [4]. In the era of big equivalent functionality [6], [7], [8]. 29
23 data and cloud computing, many service providers deploy Quality-of-Service (QoS), which measures the nonfunc- 30
24 web services to serve their customers, which explosively tional characteristics of web services, e.g., response time and 31
25 increases the number of online web services [2], [3], [5]. Under throughput, plays an important role in service discovery and 32
selection [2], [3], [8], [9]. If the QoS data of candidate web 33

D. Wu is with the Chongqing Engineering Research Center of Big Data services are available, web services that fulfill potential 34
Application for Smart Cities and Chongqing Key Laboratory of Big Data users’ QoS requirements can be selected and recommended. 35
and Intelligent Computing, Chongqing Institute of Green and Intelligent Warming-up tests, which directly invoke web services, are 36
IEE

Technology, Chinese Academy of Sciences, Chongqing 400714, China also frequently performed to retrieve QoS data [10], [11]. How- 37
with the University of Chinese Academy of Sciences, Beijing 100049,
China. E-mail: [email protected]. ever, the number of candidate services is usually very large 38

X. Luo is with the Chongqing Engineering Research Center of Big Data in real-world applications. Besides, most web service invoca- 39
Application for Smart Cities and Chongqing Key Laboratory of Big Data tions are charged. Hence, it is very time-consuming, expen- 40
and Intelligent Computing, Chongqing Institute of Green and Intelligent
Technology, Chinese Academy of Sciences, Chongqing 400714, China also
sive and thus impractical to inspect all the candidate web 41
with the Hengrui (Chongqing) Artificial Intelligence Research Center and services for retrieving QoS data [6], [12], [13]. 42
Department of Big Data Analyses Techniques, Cloudwalk, Chongqing Alternatively, QoS prediction is another widely used 43
401331, China. E-mail: [email protected]. way to acquire QoS data without such deficiency [10], [12], 44

M. Shang, and G. Wang are with the Chongqing Engineering Research
Center of Big Data Application for Smart Cities and Chongqing Key [13], [14], [15], [16]. A QoS predictor aims at accurately pre- 45
Laboratory of Big Data and Intelligent Computing, Chongqing Institute of dicting unknown QoS data based on historical web service 46
Green and Intelligent Technology, Chinese Academy of Sciences, Chongqing invocations. Collaborative filtering (CF), as a very successful 47
400714, China. E-mail: [email protected], [email protected].
technique for implementing recommender systems in 48

Q. He is with the School of Software and Electrical Engineering, Swinburne
University of Technology, Melbourne 3122, Australia. e-commerce [17], [18], [19], [20], [21], [22], [23], [24], [25], 49
E-mail: [email protected]. [26], [27], [28], has been employed to implement the QoS 50

Y. He is with the University of Louisiana at Lafayette, Lafayette, Louisiana predictors in recent years [10], [11], [12], [13], [14], [15], [29], 51
70503. E-mail: [email protected].
[30], [31], [32], [33], [34]. A CF-based QoS predictor makes 52
Manuscript received 16 Dec. 2018; revised 11 Nov. 2019; accepted 18 Dec. predictions based on a user-service QoS matrix [10], [11], 53
2019. Date of publication 0 . 2019; date of current version 0 . 2019.
(Corresponding author: Xin Luo.) [12], [13], [14], [15], [29], [30], [31], [32], [33], [35], [36], where 54
Digital Object Identifier no. 10.1109/TSC.2019.2961895 each column denotes a web service, each row denotes a user, 55

1939-1374 ß 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See ht_tp://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 2 IEEE TRANSACTIONS ON SERVICES COMPUTING

users’/services’ neighborhood is identified based on the 105

similarity between their primal LF vectors; and c) posterior- 106
neighborhood-regularized LF analysis, where the objective 107
function is regularized by both the posterior-neighborhood 108
and l2 -norm of desired LFs. The main contributions of this 109
paper include: 110

A posterior-neighborhood-regularized latent factor 111

Fig. 1. Dilemma in building reliable neighborhood based on common (PLF) model with excellent abilities in achieving 112
sets defined on raw QoS data. highly accurate QoS prediction; 113

A suite of theoretical analyses and algorithms design 114

56 and each entry denotes the QoS record of a user’s innovation of PLF; 115

57 on a service. This user-service matrix is very sparse because
Extensive experiments on a widely-used real-world 116
QoS dataset regarding the QoS prediction compari- 117

of
58 a user usually experiences only a very small subset of all the
59 available services in real-world scenarios. Thus, the major sons between PLF and related state-of-the-art models. 118

60 problem of CF-based QoS prediction is how to make accurate To the best of our knowledge, PLF significantly advances 119

61 predictions based on a very sparse user-service QoS matrix. QoS prediction in accuracy for web services. Besides, PLF 120

62 Among various CF-based QoS predictors [12], [13], [37], differs from state-of-the-art models in two major ways. First, 121

63 latent factor (LF)-based QoS predictors are the most popular PLF is capable of constructing neighborhood accurately based 122

64 due to their high scalability and prediction accuracy [11], [14], on the full information in the user-service matrix, which makes 123

65 [15], [16], [29], [30], [31], [33], [38]. Neighborhood regulariza- PLF achieve highly accurate QoS predictions. Second, PLF 124

66
67
68
69
70
71
72
73
ro
tion plays a key role in an LF-based QoS predictor because
similar users usually experience similar QoS when invoking
similar services [11], [14], [15], [33]. Current neighborhood-
regularized LF-based QoS predictors mostly employ the com-
mon set defined on the raw partial QoS matrix or additional
geographical information to identify users’/services’ neigh-
bors. They rely on prior neighborhood that is identified before
the LF analysis process and suffer from the following major
relies solely on the user-service matrix and does not require
any additional information, which makes PLF easy to use and
generally applicable to various recommendation scenarios.
The rest of this paper is organized as follows. Section 2
states the preliminaries. Section 3 presents the theories and
algorithms that facilitate the PLF model. Section 4 provides
and discusses the experimental results. Section 5 analyzes
the related work. Finally, Section 6 gives conclusions and
125
126
127
128
129
130
131
132
EP
74 limitations: future works. 133

75
It is difficult to accurately identify users’/services’ 2 PRELIMINARIES 134

76 neighbors based merely on the common set defined
77 on the extremely sparse raw user-matrix matrix. The
2.1 Notations 135

78 major reason is that the common sets of users/serv- Table 1 summarizes the annotations used in this paper. 136

79 ices are often too small to model user/service similari-

80 ties precisely. For example, Fig. 1 shows that many 2.2 Latent Factor Analysis-Based QoS Prediction 137

81 known data (red entries) are not taken into account in An LF-based QoS predictor takes a user-service matrix R as 138

82 the search of the common set among users. As a result, the input [15], [17], [18]. Since a user u from U usually 139

83 a lot of useful information is not properly utilized, invokes a very small subset of the services in S, R is very 140
IEE

84 which lowers the accuracy of the identified user’s/ sparse with jRK j << jRU j. According to prior researches 141

85 service’s neighbors. As shown in Fig. 1, although [10], [11], [15], [17], [18], [38], [39], an LF-based QoS predic- 142

86 target user u, user a, and user b are similar, user b is tor is defined as: 143

87 also considered as a dissimilar user to target user u Definition. Given U; S; R; RK and f, an LF-based QoS predic- 144
88 because they have no common QoS records. tor extracts an LF matrix P to represent U and another LF 145
89
It is often difficult to retrieve the required geographi- matrix Q to represent S based on RK to achieve R’s rank-f 146
90 cal information due to the issues of identity privacy, ^ P and Q are extracted by minimizing an objec-
approximation R. 147
91 information security, and commercial interest. More- tive function defined on RK while fulfilling the condition of 148
92 over, geographical similarities can be influenced by f << minðjUj; jSjÞ. R ^ ¼ PQ.
^ can be obtained by R 149
93 unexpected factors such as information facilities,
94 routing policies, network throughput, and time of To accurately represent RK with P and Q, an appropriate 150
95 invocation. objective function measuring the difference between R and R ^ 151
96 To overcome the above limitations, this paper proposes a is highly important [10], [11], [15], [17], [18], [38]. A Euclidean 152
97 posterior-neighborhood-regularized latent factor (PLF) model distance-based objective is a common choice [15], [17], [18]: 153
98 for QoS prediction, which neighborhood accurately for !2
1 X X f
99 regularization solely and fully based on the user-service arg min "ðP; QÞ ¼ ru;s  pu;n qs;n : (1)
100 matrix. The main idea of PLF is to decompose the LF analy- P;Q 2 ðu;sÞ2R n¼1 155
K
101 sis process into three phases: a) primal LF extraction, where 156
102 the LFs are extracted based on an objective function without Note that (1) is ill-posed [17], [38]. To address this issue, 157
103 any regularization for precisely fitting the given data; b) l2 -norm regularization is employed to enhance the generality 158
104 posterior-neighborhood construction, where the involved of a resultant LF model: 159
 WU ET AL.: POSTERIOR-NEIGHBORHOOD-REGULARIZED LATENT FACTOR MODEL FOR HIGHLY ACCURATE WEB SERVICE QOS PREDICTION 3

TABLE 1
Symbol Annotations

Symbol Explanation
U Target user set.
S Target web service set.
R Target user-service matrix with jUj rows and jSj columns.
ru;s R’s element at uth row and sth column denoting the QoS
record experienced by user u 2 U experience on service s 2 S.
RK Known entry set of R.
RU Unknown entry set of R.
f The number of dimensions in the LF space.
P jUjf LF matrix for U.
P 0jUjf LF matrix for U.
pu uth row-vector of P.
p0u uth row-vector of P’.

of
pu Mean of LFs in pu .
QfjSj LF matrix for S.
Q0fjSj LF matrix for S.
qs sth column-vector of Q.
q0s sth column-vector of Q0 .
qs Mean of LFs in qs .
R^ R’s rank-f approximation built on RK with
f << minðjUj; jSjÞ.
^ element at uth row and sth column denoting prediction Fig. 2. Flowchart of the proposed PLF model.
r^u;s R0s
for ru;s .

"ðP; QÞ
"u;s
Nmtr
h
PCCu ðu; kÞ

PCCs ðs; jÞ
K1
NK1 ðuÞ
Regualrization coefficient for l2-norm-based regularization.
Objective function with respect to P and Q.
Instant loss on ru;s .
Max-training-round count for an LF model.
Learning rate.
Pearson correlation coefficient (PCC) [16] between users u
and k.
PCC between services s and j.
The number of the K1-nearest-neighbors for a user.
K1-nearest-neighbor set for user u.
ro regularization may be not always useful to improve the pre-
diction accuracy because the prior neighborhood directly
identified on the raw partial QoS matrix are not reliable. To
address this issue, we propose the PLF model, which adopts a
posterior approach to identify users’/services’ neighborhood.
Fig. 2 shows the flowchart of a PLF model that consists of
three phases. Phase 1 is the primal LF extraction. By training
an LF model without any regularization, P and Q are
173
174
175
176
177
178
179
EP
180
Tu ðuÞ Regularization neighbor set for user u.
K2 The number of the K2-nearest-neighbors for a service. extracted from the full known entries in R for users and serv- 181
NK2 ðsÞ K2-nearest-neighbor set for service s. ices respectively. Phase 2 is the posterior neighbor identifica- 182
Ts ðsÞ Regularization neighbor set for service s. tion for target users/services. Specifically, each target user’s 183
Suu;k Normalized similarity weight between users u and k.
Sss;j Normalized similarity weight between services s and j.
similar users are identified from P and each target service’s 184
a1 Regularization coefficient for posterior-neighborhood-based similar services are identified from Q. Phase 3 is the posterior- 185
regularization on users. neighborhood-regularized LF analysis for predicting the 186
a2 Regularization coefficient for posterior-neighborhood-based missing QoS data in R, where the objective function is regular- 187
regularization on services.
G Testing set. ized by both the posterior-neighborhood and l2 -norm of the 188
G A synthetic dataset desired LFs. In this section, we discuss and analyze the three 189
gu;s G’s element at uth row and sth column. phases in detail. 190
jj The cardinality of an enclosed set.
IEE

j  jabs The absolute value of an enclosed number.

3.1 Primal LF Extraction 191
In this phase, the aim is to precisely extract primal LF matri- 192
arg min "ðP; QÞ ces P and Q for users and services respectively, which 193
P;Q means that the LF model should be trained without regular- 194
!2 !
1 X X  X X X ization to accurately fit RK . To do so, we apply SGD to (1).
f f f 195
¼ ru;s  pu;n qs;n þ p2u;n þ 2
qs;n : By considering the instant loss on a single instance ru;s , we 196
2 ðu;sÞ2R n¼1
2 ðu;sÞ2R n¼1 n¼1
K K
161 have 197
(2)
162 !2
163 With an optimization algorithm, such as stochastic gradi- 1 Xf
"u;s ¼ ru;s  pu;n qs;n : (3)
164 ent descent (SGD), P and Q can be extracted from (2). 2 n¼1
199
200
165 3 POSTERIOR-NEIGHBORHOOD-REGULARIZED With SGD, the LFs involved in (3) are trained by moving 201
166 LATENT FACTOR MODEL them along the opposite of the stochastic gradient of (3) 202

167 As analyzed in prior studies [11], [14], [15], [33], the predic- with respect to each single LF, i.e., we make 203

168 tion accuracy of an LF-based QoS predictor can be further 8  P 

169 improved by regularizing an objective function like (2) with < pu;n pu;n þ hqs;n ru;s  fn¼1 pu;n qs;n ;
On ru;s ; forn ¼ 1  f :  P 
170 users’/services’ neighborhood information. However, cur- : qs;n qs;n þ hpu;n ru;s  fn¼1 pu;n qs;n :
171 rent models rely on the prior neighborhood that is built 205
172 before the LF analysis to achieve the regularization. Such (4) 206
 4 IEEE TRANSACTIONS ON SERVICES COMPUTING

With the service similarity given by (8), we can obtain the 250
regularization neighbor set for each service by 251

 
Ts ðsÞ ¼ jj 2 NK2 ðsÞ; PCCs ðs; jÞ > 0; j 6¼ s : (9)
253
254
Given that the known data of a target QoS matrix are reli- 255
Fig. 3. An example of posterior-neighborhood construction.
able (i.e., they are true user-service invocation records), an 256

207 After all the ru;s in RK are used to train with (4), we can LF analysis-based model can assure that its resultant LFs 257

208 extract the primal LFs, which can be utilized to identify the well represent the highly valuable information hidden in 258

209 posterior-neighbors of the target users/services in Phase 2. them. From this point of view, we see the reason why we 259
choose to perform LF analysis on a sparse QoS matrix as a 260

210 3.2 Posterior-Neighborhood Construction pre-step before the neighborhood detection. Naturally, we 261
can achieve neighborhood based on a sparse QoS matrix 262

of
211 With the primal LF matrices P and Q extracted in Phase 1, we
212 can obtain the feature vectors for each user/service. Those fea- directly by computing PCC among its row/column entities, 263

213 ture vectors are dense. In addition, by accurately fitting RK , P i.e., the involved users/services. However, the quality of the 264

214 and Q well represent the information hidden in R. Thus, resultant neighborhood can be greatly impaired by the target 265

215 we can precisely identify target users’/services’ posterior- matrix’s sparsity. In comparison, a PLF model first extracts 266

216 neighbors based on P and Q. For example, as shown in Fig. 3, compact and dense LF matrices from the sparse QoS matrix, 267

217 through the primal LF extraction, we obtain the dense LF where each involved user/service is described by a dense LF 268

218 matrix P which includes the feature vectors for target user u, vector. Moreover, such an LF vector is acquired on premise 269

219
220
221
222
223
224
225
226
user a, and user b. Base on P, we can identify users a and b as
target user u’s similar users. This addresses the limitation
shown in Fig. 1 where user b cannot be identified as target
user u’s similar user because they have no common set.
As indicated by prior researches [11], [14], [15], [40], PCC
is a reliable similarity metric. Hence, we choose it in this
research as the similarity metric for posterior-neighbor iden-
tification. For two users u and k, we compute their similarity
ro of well representing observed interactions in the target sys-
tem. Thus, the neighborhood built on the LF matrices is
expected to be far more reliable on one built on the original
sparse QoS matrix.

3.3 Posterior-Neighborhood-Regularized
LF Analysis
In this phase, we adopt the posterior-neighbors identified in
270
271
272
273

274
275
276
EP
227 PCCu ðu; kÞ as follow phase 2 to help to extract two new LF matrices P 0 and Q0. 277

Pf    Considering similar users tend to experience similar QoS on 278

n¼1 pu;n  p u pk;n  pk similar services [11], [14], [15], we design the following loss 279
PCCu ðu; kÞ ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pf  2ffiqP ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 2ffi ;
u2U;k2U;u6¼k f function: 280
n¼1 pu;n  p u n¼1 pk;n  p k
229 (5) arg min "ðP 0 ; Q0 Þ
230 where pu and pk are respectively given by P 0 ;Q0
0 1
X
jU j X f 
X 2
1 X
f
1 X
f
¼ @ Suu;k p0u;n  p0k;n A
pu ¼ pu;n ; pk ¼ pk;n : (6) (10)
232 f n¼1 f n¼1 u¼1 k2Tu ðuÞ n¼1
0 1
233 X
jS j X f 
X 2
IEE

234 Note that PCCu ðu; kÞ lies in the interval of [-1,1], where a þ @ Sss;j q0s;n  q0j;n A;
235 large value denotes a high similarity between the corre- s¼1 j2Ts ðsÞ n¼1
236 sponding pair of users. Based on user similarity given by 282
237 (5), the K1 -nearest-neighbor set for each user can be identi- where Suu;k and Sss;j model the linear weight of user u’s kth 283
238 fied. Then, we can obtain the regularization neighbor set for neighbor and service s’s jth neighbor respectively based on 284
239 each user by the corresponding posterior similarities. To neutralize the 285
  impact of magnitudes, Suu;k and Sss;j are given by
Tu ðuÞ ¼ kk 2 NK1 ðuÞ; PCCu ðu; kÞ > 0; k 6¼ u :
286
241 (7)
242
PCCu ðu; kÞ PCCs ðs; jÞ
243 Note that the neighborhood relationship described by (7) Suu;k ¼ P ; Sss;j ¼ P :
244 is asymmetric. PCCu ðu; k0 Þ
k0 2Tu ðuÞ 0
0
j 2Ts ðsÞ PCCs ðs; j Þ
288
245 Similarly, given two services s and j, we compute their (11)
246 similarity PCCs ðs; jÞ with 289
From (10) and (11), we can see that a user’s/service’s 290
Pf   
n¼1 qs;n  q s qj;n  qj most similar neighbor has the highest impact on the gener- 291
PCCs ðs; jÞs2S;j2S;s6¼j ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pf  2ffiqffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pf  2 ; alized loss. Moreover, with (10) we attempt to minimize the 292
n¼1 qs;n  q s n¼1 qj;n  q j difference between LFs from the most similar neighbors, 293

1 X
f
1 X
f thereby keeping them similar as reflected in the posterior- 294
qs ¼ qs;n ; qj ¼ qj;n : neighborhood. 295
f n¼1 f n¼1
248 By integrating (10) into (2), we obtain the posterior- 296
249 (8) neighborhood-regularized objective function 297
 WU ET AL.: POSTERIOR-NEIGHBORHOOD-REGULARIZED LATENT FACTOR MODEL FOR HIGHLY ACCURATE WEB SERVICE QOS PREDICTION 5

!2 The computational complexity of Algorithm PLFA is

1 X X
f 328
0 0
"ðP ; Q Þ ¼ ru;s  p0u;n q0s;n 329
2 ðu;sÞ2R
K n¼1 Algorithm PNC 330
!
f  2 X f  2
 X X
0 0
Input: R 331
þ p þ qs;n Output: PCCu ðu; kÞ; Tu ðuÞ, PCCs ðs; jÞ, Ts ðsÞ Cost
2 ðu;sÞ2R n¼1 u;n n¼1
332

0
k
1 1 Initializing f; ; h; Nmtr Qð1Þ 333
(12)
f  2 2 while t  Nmtr && not converge Nmtr
a1 X @ X X
jU j 334
þ Suu;k p0u;n  p0k;n A 3 for each known entry ru;s inR == ru;s 2 RK  jRK j 335
2 u¼1 k2T ðuÞ
u n¼1 4 for n ¼ 1 to f f 336
0 1 5 computing pu;n according to formula (4) Qð1Þ 337
f  2
a2 X @ X X
j S j
þ Sss;j q0s;n  q0j;n A; 6 computing ps;n according to formula (4) Qð1Þ 338
2 s¼1 j2T ðsÞ n¼1 7 end for – 339
s
299 8 end for – 340
where the instant loss on a single instance ru;s is 9 t¼tþ1 Qð1Þ

of
300 341

!2 ! 10 end while – 342

X  X  0 2 X  0 2 11 for u ¼ 1 to jUj  jUj
f f f 343
1
"u;s ¼ ru;s  p0u;n q0s;n þ pu;n þ qs;n 12 computing pu according to formula (6) QðfÞ 344
2 n¼1
2 n¼1 n¼1 13 end for – 345
f  2
a1 X X 14 for u ¼ 1 to jUj  jUj 346
þ Suu;k p0u;n  p0k;n 15 for k ¼ u þ 1 to jUj  ðjUj  1Þ=2 347
2 k2T ðuÞ n¼1
u 16 computing PCCu ðu; kÞ according to formula (5) QðfÞ 348
X f 
X 2

302
303
304
þ
a2
2 j2T
s ðsÞ
Sss;j q0s;n  q0j;n :
n¼1

To train the involved LFs, we apply SGD to (13)

On ru;s ; forn ¼ 1  f :
8
< p0u;n
: q0s;n
p0u;n  h @p0u;s
q0s;n
@"
u;n
@"u;s
ro:
(13)

(14)
17 end for
18 end for
19 for u ¼ 1 to jUj
20 computing Tu ðuÞ according to formula (7)
21 end for
22 for s ¼ 1 to jSj
23 computing qs according to formula (8)
24 end for
25 for s ¼ 1 to jSj
–
–
 jUj
K1
–
 jSj
QðfÞ
–
 jSj
349
350
351
352
353
354
355
356
EP
306 h 0
@qs;n
357

307
26 for j ¼ s þ 1 to jSj  ðjSj  1Þ=2 358
27 computing PCCs ðs; jÞ according to formula (8) QðfÞ 359
308 Note that after phases 1 and 2, Suu;k and Sss;j are already
28 end for – 360
309 known, they should be treated as constant in (14). Hence,
29 end for – 361
310 we obtain the following training rules for a PLF model
30 for s ¼ 1 to jSj  jSj 362
8   31 computing Ts ðsÞ according to formula (9) K2 363
>
> p0u;n p0u;n þ hq0s;n ru;s  p0u;n q0s;n 32 end for –
>
>   364
>
> P 33 return: PCCu ðu; kÞ; Tu ðuÞ; PCCs ðs; jÞ; Ts ðsÞ Qð1Þ
< hp0u;n  ha1 k2T ðuÞ Suu;k p0u;n  p0k;n ; 365
On ru;s ; forn ¼ 1  f : u 
>
> q0s;n q0s;n þ hp0u;n ru;s  p0u;n q0s;n
>
>  
>
: hq0  ha P
>
Ss q0  q0 : C ¼ Qð1Þ þ jUj  ðQðK1 Þ þ jUj  Qð1ÞÞ
s;n 2 j2Ts ðsÞ s;j s;n j;n
IEE

312
313 (15) þ jSj  ðQðK2 Þ
314 3.4 Algorithm Design and Analysis
þ jSj  Qð1ÞÞ þ Nmtr  QðjUj  f  K1
315 In our design, a PLF model relies on two algorithms, i.e., Pos- þjS j  f  K2 þ jRK j  f  2Þ þ Qð1Þ
316 terior-Neighborhood Construction (PNC), and Posterior-
317 neighborhood-regularized Latent Factor Analysis (PLFA).  QðjUj2 þ jSj2 þ Nmtr  ðjRK j þ jUj  K1
318 Algorithm PNC constructs the posterior-neighborhood that þ jSj  K2 Þ  fÞ:
319 is adopted by Algorithm PLFA. Algorithm PLFA calls Algo-
320 rithm PNC to minimize (12) for predicting the missing QoS (17)
321 data. Their pseudocode and the time cost are given in Algo- In real-world scenarios, there are only a limited number 366
322 rithm PNC and Algorithm PLFA, respectively. Next, we ana- of users or services that are similar to the target user or ser- 367
323 lyze their computational cost. vice in their QoS experiences on co-invoked services. As a 368
324 The computational complexity of Algorithm PNA is result, K1 and K2 are much smaller than jUj and jSj for a 369
PLF model. Thus, the overall computational complexity of 370
C ¼ Qð1Þ þ Qð1Þ þ Nmtr  ðjRK j  f  2  Qð1Þ þ Qð1ÞÞ þ Qð1Þ Algorithm PLFA is 371
þ jUj  Qðf Þ þ jUj  ðjU j  1Þ=2  Qðf Þ þ jUj  K1  
þ jSj  Qðf Þ þ jSj  ðjS j  1Þ=2  Qðf Þ þ jSj  K2 þ Qð1Þ C  Q jUj2 þ jSj2 þ Nmtr  jRK j  f : (18) 373
   
374
 Q Nmtr  jRK j  f þ jUj2 þ jSj2  f :
326 Finally, the total computational complexity of building a 375
327 (16) PLF model is 376
 6 IEEE TRANSACTIONS ON SERVICES COMPUTING

    TABLE 2
C  Q Nmtr  jRK j  f þ jUj2 þ jSj2  f Properties of All the Designed Test Cases
 
þ Q jUj2 þ jSj2 þ Nmtr  jRK j  f Dataset No. Density Training data Testing data
 
 Q ðjUj2 þ jSj2 Þ  f þ QðNmtr  jRK j  f Þ: D1 5% 93,692 1,780,146
378 D2 10% 187,384 1,686,454
(19) Response Time
379 D3 15% 281,076 1,592,762
380 In Section 4, we will theoretically and experimentally D4 20% 374,768 1,499,070
381 demonstrate that the computational complexity of Algo-
382 rithm PLFA is comparable to the state-of-the-art LF-based properties of all the test cases, where the column “Density” 432
383 QoS predictors with neighborhood-regularization. indicates the density of the training data. Each test case is 433
384 repeated 10 times and the results are averaged. 434
385 Algorithm PLFA

of
386 Input: R; PCCu ðu; kÞ; Tu ðuÞ; PCCs ðs; jÞ; Ts ðsÞ
4.2 Evaluation Protocol 435

387 Output: R ^ Cost QoS prediction aims to predict the unknown QoS data 436

388 1 Calling Algorithm PNC Qð1Þ based on the known ones. Hence, this work mainly focuses 437

389 2 Initializing f; ; h; Nmtr Qð1Þ on the prediction accuracy, i.e., the closeness between the 438
390 3 for u ¼ 1 to jUjP  jUj prediction results and the actual values. In our experiments, 439
391 4 computing k0 2Tu ðuÞ PCCu ðu; k0 Þ QðK1 Þ we employ the mean absolute error (MAE) and the root 440

392 5 for k ¼ 1 to jUj  jUj mean squared error (RMSE), which are also widely used for 441

393 6 computing Suu;k according to formula (11) Qð1Þ researching QoS prediction [11], [14], [15], [29], [30], [31], 442
394
395
396
397

398
399
400
401
7 end for
8 end for
9 for s ¼ 1 to jSjP
10 computing j0 2Ts ðsÞ PCCs ðs; j0 Þ
11
12
13
14
for s ¼ 1 to jSj
computing Sss;j according to formula (11)
end for
end for
ro –
–
 jSj
QðK2 Þ
 jSj
Qð1Þ
–
–
[33], to evaluate the prediction accuracy of PLF:

MAE ¼ @

v0
u
0
X 

ðw;jÞ2G


1,

rw;j  r^w;j  A jGj;

abs
(20)

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u X 
u@
RMSE ¼ t rw;j  r^w;j A jGj;
2
1,
(21)
443

445
446
EP
402 15 while t  Nmtr && not converge Nmtr ðw;jÞ2G

403 16 for u ¼ 1 to jUj  jUj 448

404 17 for n ¼ 1 to f f where a lower MAE or RMSE value denotes a higher predic- 449
P
405 18 computing ha1 k2Tu ðuÞ Suu;k ðpu;n  pk;n Þ QðK1 Þ tion accuracy. All the experiments are run with 3.7 GHz i7 450
406 19 end for – central processing unit (CPU) and 64 GB random access 451
407 20 end for – memory. 452
408 21 for s ¼ 1 to jSj  jSj
409 22 for n ¼ 1 to f P f 4.3 Impacts of K1 and K2 453
410 23 computing ha2 j2Ts ðsÞ Sus;j ðqs;n  qj;n Þ. QðK1 Þ First, we analyze the impacts of K1 and K2 on the prediction 454
411 24 end for – accuracy of PLF. In this set of experiments, the parameters 455
412 25 end for – are set as a1 ¼ 0:2; a2 ¼ 0:2; f ¼ 20;  ¼ 0:01, and h ¼ 0:01, 456
for each known entry ru;s in R == ru;s 2 RK  jRK j
IEE

413 26 uniformly. Figs. 4 and 5 shows the experimental results on 457

414 27 for n ¼ 1 to f f MAE and RMSE respectively when K1 increases from 0 to 458
415 28 computing pd0u;n according to formula (15) Qð1Þ
50 and K2 from 0 to 100. From these figures, we have the fol- 459
416 29 computing p0s;n according to formula (15) Qð1Þ
lowing observations. 460
417 30 end for –
When K1 and K2 are both set as zero, PLF has the highest 461
418 31 end for –
MAE and RMSE in almost all the test cases because it is 462
419 32 t¼tþ1 Qð1Þ
420 33 end while – equivalent to an LF model. This observation confirms that 463

421 34 return P; Q Qð1Þ the posterior-neighborhood-based regularization is useful 464

in improving the prediction accuracy for an LF model. 465
When K1 or K2 is set as zero, PLF has a relatively high 466
MAE and RMSE in all the test cases except for the MAE in one 467
422 4 EXPERIMENTS AND RESULTS situation where K2 is set as zero on D1. The reason for this 468
423 4.1 Datasets exception is that the density of D1 is low (5 percent) and the 469
424 To evaluate the PLF model, we conduct extensive experi- number of web services (5,825) is much larger than that of the 470
425 ments on a benchmark dataset named Response Time. It users (339). In such a situation, more posterior-neighborhood 471
426 contains real-world web service QoS data and has been com- information is extracted from the users than the services. As a 472
427 monly used in prior research on QoS prediction [2], [6], [10], result, the users’ posterior-neighborhood-based regulariza- 473
428 [11], [14], [15], [29], [30], [31], [33]. It records the response tion is more useful than the services’ in improving the predic- 474
429 times of 5,825 web services experienced by 339 users’ tion accuracy of PLF. 475
430 1,873,838 invocations. Different test cases are designed to As K1 and K2 increase, the MAE and RMSE of PLF 476
431 validate the performance of PLF. Table 2 summarizes the decrease at first and then increase in general. For example, 477
 WU ET AL.: POSTERIOR-NEIGHBORHOOD-REGULARIZED LATENT FACTOR MODEL FOR HIGHLY ACCURATE WEB SERVICE QOS PREDICTION 7

Fig. 4. MAE of PLF when K1 increases from 0 to 50 and K2 increases from 0 to 100: (a) D1, (b) D2, (c) D3, (d) D4.

of
ro
Fig. 5. RMSE of PLF when K1 increases from 0 to 50 and K2 increases from 0 to 100: (a) D1, (b) D2, (c) D3, (d) D4.
EP
Fig. 6. MAE of PLF when a1 and a2 increase from 0 to 1: (a) D1, (b) D2, (c) D3, (d) D4.

478 on D4, since more similar users/services are included in 4.4 Impacts of a1 and a2 501
479 PLF as K1 and K2 increase, the MAE decreases from 0.4589 This set of experiments focuses on the impacts of a1 and a2 502
IEE

480 to 0.4238 at first. However, when K1 and K2 become too when their values increase from 0 to 1, as shown in Figs. 6 503
481 large, some dissimilar users/services are also introduced and 7. The other parameters are set as K1 ¼ 15; K2 ¼ 50; 504
482 into PLF and jeopardize its prediction accuracy. f ¼ 20;  ¼ 0:01, and h ¼ 0:01, uniformly. As demonstrated, 505
483 As the density of the training data increases, PLF tends to the MAE and RMSE of PLF decrease as a1 and a2 increase at 506
484 achieve the lowest MAE and RMSE on the smaller K1 and the beginning. After reaching a certain threshold, as a1 and 507
485 K2 . For example, on D1 (5 percent), PLF achieves the lowest a2 continue to increase, the MAE and RMSE of PLF increase. 508
486 RMSE when K1 is in the range of 40—50 and K2 is in the The reason is that when a1 and a2 are too large, PLF will 509
487 range of 50—100. On D4 (20 percent), the optimal ranges of heavily rely on the posterior-neighborhood-based regulariza- 510
488 K1 and K2 for PLF decrease to 10—20 and 40—60 respec- tion, which leads to underfitting. In real-world applications, 511
489 tively. The major reason is that more training data contains the datasets generated from different domains differ from 512
490 more posterior-neighborhood information, which is benefi- each other vastly in user count, web service count, instance 513
491 cial for PLF in finding highly similar users/services. In count, and data density. The optimal a1 and a2 depend 514
492 other words, some dissimilar users/services, which were heavily on such data characteristics. Hence, a1 and a2 are 515
493 included in the prediction model at the beginning, will be domain-specific, they should be experimentally inspected 516
494 discovered and discarded finally as the density of the train- and set for PLF to achieve high prediction accuracy. 517
495 ing data increases. According to this observation, we set
496 K1 ¼ 15 and K2 ¼ 50 as the default setting in the next 4.5 Impact of f 518
497 experiments. This set of experiments analyzes the impact of f with 519
498 In conclusion, these observations verify that the posterior- K1 ¼ 15; K2 ¼ 50; a1 ¼ 0:1; a2 ¼ 0:4;  ¼ 0:01, and h ¼ 0:01. 520
499 neighborhood-based regularization is significantly effective Fig. 8 shows the experimental results with f increasing from 521
500 in improving the prediction accuracy of PLF. 10, 20, 40, 80, 160 to 320. A higher-dimensional LF space 522
 8 IEEE TRANSACTIONS ON SERVICES COMPUTING

Fig. 7. RMSE of PLF when a1 and a2 increase from 0 to 1: (a) D1, (b) D2, (c) D3, (d) D4.

of
ro
Fig. 8. MAE and RMSE of PLF with different f: (a) D1, (b) D2, (c) D3, (d) D4.
EP
Fig. 9. MAE and RMSE of PLF as  increases: (a) D1, (b) D2, (c) D3, (d) D4.

523 usually provides PLF with better representation learning highly accurate prediction results. Since  is also dependent 547
524 ability [6], [17]. Fig. 8 shows that the MAE and RMSE of PLF on the data characteristics [38], its optimal value needs to be 548
525 decrease as f increases in general in all the test cases. More- experimentally inspected. 549
526 over, it shows that the MAE and RMSE of PLF tend to
IEE

527 decrease slightly or even increase after f exceeds a certain 4.7 Comparisons Between PLF 550
528 threshold. The reason is that when an appropriate value is and State-of-the-Art Models 551
529 assigned to f, PLF obtains the highest representation learning Finally, we theoretically and experimentally compare PLF 552
530 ability. An extra increase in f cannot bring significant with seven related state-of-the-art models in their prediction 553
531 improvement in the prediction accuracy. Instead, it increases accuracy and computational efficiency. These compared mod- 554
532 the probability of overfitting which lowers prediction accu- els are three LF-based models (BMF, NIMF, and RSNMF), 555
533 racy. Besides, the computational cost of PLF also increases three LF-based models with additional geographical informa- 556
534 linearly with f. Hence, we set f ¼ 20 to balance the computa- tion (NAMF, GeoMF, and LMF-PP), and one deep neural net- 557
535 tional cost and the prediction accuracy in all the experiments. work (DNN) based model (AutoRec). Note that NIMF, 558
NAMF, GeoMF, and LMF-PP are all prior-neighborhood- 559
536 4.6 Impact of  based, while PLF is posterior-neighborhood-based. These 560
537 In this set of experiments, we analyze the impact when  compared models have different characteristics as given in 561
538 increases. The other parameters are set as K1 ¼ 15; K2 ¼ 50; Table 3. Besides, their computational complexity is concluded 562
539 a1 ¼ 0:1; a2 ¼ 0:4; f ¼ 20, and h ¼ 0:01, uniformly. The in Table 4, where there are two major parts, one is for con- 563
540 results are shown in Fig. 9. Since the l2 -norm regularization structing the neighborhood regularization and the other one 564
541 can prevent PLF from overfitting, we find that the MAE and is for minimizing the objective function. Since AutoRec’s 565
542 RMSE of PLF decrease at first as  increases in all the test computational complexity is obviously much higher than its 566
543 cases. However, after  exceeds a certain point, the MAE and peers due to its DNN structure [43], we do not summarize its 567
544 RMSE of PLF increase. This indicates that PLF is largely computational complexity in Table 4. 568
545 impacted by the l2 -norm regularization. From the experimen- In prediction accuracy, the dimension of LF is set as 569
546 tal results, we conclude that  is crucial for PLF to achieve f ¼ 20 for all the models except for AutoRec—because it is 570
 WU ET AL.: POSTERIOR-NEIGHBORHOOD-REGULARIZED LATENT FACTOR MODEL FOR HIGHLY ACCURATE WEB SERVICE QOS PREDICTION 9

TABLE 3 TABLE 5
Descriptions of All the Compared Models Statistical Results of Prediction Accuracy by Conducting
Wilcoxon Signed-Ranks Test With A Significance Level of 0.05
Models Descriptions
The basic matrix factorization (MF) model [38] which consists Comparison Rþ R- p-value
BMF of two situations, i.e., with and without linear biases. We pick PLF vs. BMF 36 0 0.0039
the best one for comparisons on each test case. PLF vs. NIMF 36 0 0.0039
NIMF
The neighborhood-integrated MF model [16] extends BMF by PLF vs. RSNMF 36 0 0.0039
employing the information of similar users. PLF vs. NAMF 36 0 0.0039
The regularized single element dependent non-negative MF PLF vs. GeoMF 27 9 0.1250
RSNMF model [6] is designed for QoS prediction by incorporating PLF vs. LMF-PP 36 0 0.0039
regularization into a non-negative MF. PLF vs. AutoRec 36 0 0.0039
The network-aware MF model [15] employs the additional
NAMF
network distances information to construct the neighborhood.
The improved MF model [14] employs the additional

of
GeoMF
geographical relationships to construct the neighborhood.
The location-based MF model [11] employs the additional
LMF-PP
location information to construct the neighborhood.
The deep neural networks based model [41] employs an
AutoRec autoencoder [42] framework for CF and consists of I-AutoRec
and U-AutoRec. The best one is chosen to compare.
The neighborhood of latent-factor based MF model proposed
PLF
in this paper.

Model

BMF [38]
NIMF [16]
TABLE 4
The Computational Complexities of All the Models

Complexity of constructing
neighborhood regularization
/
QðjUj2  jSjÞ
ro
Complexity of minimizing
objective function
QðNmtr  jRK j  fÞ
QðNmtr  jRK j  f  K12 Þ
Fig. 11. The compared results of CPU running time on all the test cases.

To validate whether PLF achieves significantly higher 585

EP
RSNMF [6] / QðNmtr  jRK j  fÞ
NAMF [15] QðjUj2 Þ QðNmtr  jRK j  fÞ prediction accuracy than the other models, the Wilcoxon 586
GeoMF [14] QðjUj  jSj þ jSj2  jUjÞ
2
QðNmtr  jRK j  f 2  ðK1 þ K2 ÞÞ signed-ranks test [44], which is a nonparametric pairwise 587
LMF-PP [11] QðjUj2  jSj þ jSj2  jUjÞ QðNmtr  jRK j  fÞ comparison procedure, is applied to perform the statistical 588
PLF QððjUj2 þ jSj2 Þ  fÞ QðNmtr  jRK j  fÞ
analysis. Table 5 records the statistical results, where three 589
columns respectively show the achieved rankings Rþ and 590
R values and its associated p-value. A larger Rþ value 591
indicates that PLF has better prediction accuracy. The 592
accepted hypotheses that PLF outperforms the other mod- 593
els with a significance level of 0.05 are highlighted in bold. 594
Table 5 clearly validates that PLF has significantly higher 595
prediction accuracy than the other models except for 596
IEE

GeoMF. However, we can see that PLF also achieves much 597
higher Rþ rankings than GeoMF, which indicates that 598
PLF has slightly higher prediction accuracy than GeoMF. 599
More importantly, PLF does so without the need for 600
the additional geographical information that is required 601
Fig. 10. The compared results on all the test cases: (a) MAE, (b) RMSE. by GeoMF. 602
Considering the computational efficiency of involved 603
571 a deep neural networks based model—to facilitate fair com- models, we summarize their CPU running time in all testing 604
572 parisons. Besides, all other parameters involved in the com- cases in Fig. 11. From Fig. 11 and Table 4, we have several 605
573 pared models are all set as instructed in their corresponding observations: 606
574 papers. For PLF, the other parameters are set as K1 ¼ 15;
575 K2 ¼ 50; a1 ¼ 0:1; a2 ¼ 0:4, and h ¼ 0:01. Fig. 10 records a) BMF and RSNMF have no computational cost in con- 607
576 the compared results. It shows that PLF achieves the lowest structing neighborhood regularization because they 608
577 MAE in all the test cases, which are 4.8—6.62 percent, do not consider neighborhood regularization, making 609
578 2.62—4.69 percent, 0.66—3.57 percent, 3—6.03 percent, them consume less time than their peers do. How- 610
579 1.15—3.8 percent, 0.78—1.83 percent, and 4.08—7.5 percent ever, they are generally outperformed by their peers 611
580 lower than that achieved by BMF, NIMF, RSNMF, NAMF, in prediction accuracy for missing data. 612
581 GeoMF, LMF-PP, AutoRec respectively. In RMSE, PLF and b) PLF’s computational efficiency is higher than that of 613
582 GeoMF perform better than the other models in all the test GeoMF and LMF-PP because of f < < minfjUj; jSjg. 614
583 cases. Furthermore, the RMSE achieved by PLF is slightly c) For minimizing the objective function, NIMF and 615
584 worse than GeoMF on D1 and D2. GeoMF are slower than the other models. In other 616
 10 IEEE TRANSACTIONS ON SERVICES COMPUTING

TABLE 6 TABLE 8
Properties of All the Designed Test Cases on the Compared Results of CPU Running Time (sec)
Synthetic Large Scale Dataset
Test cases BMF RSNMF PLF
Dataset No. Density Training data Testing data DS 1 2348.91 11461.05 88.21þ9979.38
DS 1 5% 4,684,595 89,007,305 DS 2 2301.39 2690.64 76.88þ9507.93
Synthetic large DS 2 10% 9,369,190 84,322,710 DS 3 1205.52 2946.77 80.17þ6855.11
scale dataset DS 3 15% 14,053,785 79,638,115 DS 4 524.73 3341.05 75.16þ6328.33
DS 4 20% 18,738,380 74,953,520
The first part before þ is for constructing neighborhood regularization and
the second part after þ is for minimizing objective function.

TABLE 7
Compared Results on Prediction Accuracy

of
RMSE MAE
Test cases
BMF RSNMF PLF BMF RSNMF PLF
DS 1 1.0715 1.0589 1.0085 0.4053 0.3849 0.3758
DS 2 1.0181 0.9986 0.9804 0.3959 0.3611 0.3584
DS 3 1.0018 0.9666 0.9498 0.3964 0.3524 0.3511
DS 4 0.9948 0.9497 0.9458 0.3977 0.3494 0.3488

617
618
619
620
621
622
623
words, the PLF, BMF, RSNMF, NAMF and LMF-PP
models have the lowest computational complexity.
d) AutoRec has the most CPU running time among all
the models due to its DNN-based structure.
ro
e) PLF’s computational efficiency is comparable to that
of NIMF, NAMF, GeoMF, and LMF-PP.
Based on the experimental results and the analyses in
Fig. 12. PLF’s performance on DS1 as the core number of CPU
increases: (a) CPU running time, (b) Speedup.

From them, we observe that: a) PLF has better prediction

accuracy while consumes more CPU running time than
BMF and RSNMF, b) PLF consumes much less CPU running
time for constructing neighborhood regularization than for
655
656
657
658
EP
624 this section, we conclude that PLF has better performance minimizing the objective function, and c) all the models con- 659
625 than the seven state-of-the-art models because a) it has the sume less CPU running time as density increases, which 660
626 best prediction accuracy, and b) when comparing with means that more training data can speed up their conver- 661
627 NAMF, GeoMF, and LMF-PP, it has comparable computa- gence. These results validate that on the large scale dataset, 662
628 tional complexity and does not require additional geo- PLF performs well on prediction accuracy while its compu- 663
629 graphical information. tational efficiency is not outstanding. 664
As analyzed in [45], [46], an LF-based model’s computa- 665
630 4.8 Performance on Large Scale Dataset tional efficiency can be greatly improved by implementing 666
631 As analyzed in Section 4.7, PLF consumes more CPU run- parallelization. Especially, for an LF-based model, its different 667
632 ning time than the ordinary LF-based QoS predictors with- users’/services’ LFs can be simultaneously trained by the 668
633 out neighborhood regularization (BMF and RSNMF). To alternating stochastic gradient descent (ASGD) without 669
IEE

634 evaluate PLF’s performance on a large scale dataset, we gen- impacting prediction accuracy [46]. On this basis, we develop 670
635 erate a synthetic large scale dataset by introducing Gaussian PLF to a parallel version that can parallel minimizing its objec- 671
636 noise to the dataset Response Time. The specific method of tive function (because PLF consumes most time to minimize 672
637 generating synthetic dataset is: a) for each known entry ru;s the objective function), please refer to [46] for details. Then, 673
638 in Response Time Rj339jj5825j , generating the corresponding we test PLF’s computational efficiency on DS1 as the core 674
639 synthetic entry gu;s by the formula gu;s ¼ ru;s þ ru;s  0:05  number of CPU increases, the results are recorded in Fig. 12. 675
640 randðÞ, where randðÞ indicates a random real number that From it, we can see that PLF’s computational efficiency has 676
641 obeys the standard normal distribution, then we have a syn- been improved as the core number of CPU increases. Spe- 677
642 thetic dataset Gj339jj5825j , b) repeating the step a for 50 times cially, we find that the CPU running time with 8 cores 678
643 to obtain 50 different datasets Gj339jj5825j , and c) merging (1860.91) is less than BMF (2348.91 seconds), which means 679
644 the 50 different datasets by rows to form a large scale that PLF has higher computational efficiency than BMF when 680
645 dataset Gj16950jj5825j . After that, the original Response Time is it employs more than 8 cores to develop a parallel version. 681
646 expanded from 339 users to 16950 users, and the total num- Hence, based on the above analyses, we conclude that on 682
647 ber of invocations is increased from 1,873,838 to 93,691,900. the large scale dataset, PLF not only performs well on pre- 683
648 Then, different test cases as designed in Table 2 are also used diction accuracy but also has high computational efficiency 684
649 to validate PLF’s performance on the synthetic large scale by implementing parallelization. 685
650 dataset, as shown in Table 6.
651 Then, we compare PLF with BMF and RSNMF (without
652 neighborhood regularization) on prediction accuracy and 5 RELATED WORK 686

653 computational efficiency. The parameters are set the same The proposed PLF is an LF-based model. An LF model orig- 687
654 as in Section 4.7. Tables 7 and 8 record the compared results. inates from matrix factorization [6], [17], [47] and has been 688
 WU ET AL.: POSTERIOR-NEIGHBORHOOD-REGULARIZED LATENT FACTOR MODEL FOR HIGHLY ACCURATE WEB SERVICE QOS PREDICTION 11

689 widely used to develop the QoS predictor due to high scal- accurate Quality-of-Service (QoS) prediction for web services. 748
690 ability and prediction accuracy [16], [29], [30], [31], [33], PLF goes through three phases: primal LF extraction, poste- 749
691 [38]. An LF-based QoS predictor is built on a low-rank rior neighbor identification, and posterior-neighborhood-reg- 750
692 approximation to a user-service matrix based only on the ularized LF analysis. Extensive experiments are conducted on 751
693 known entries in the matrix. It maps both users/services a real-world benchmark dataset to evaluate PLF against the 752
694 into the same low-dimensional LF space, trains desired LFs state-of-the-art models. The experimental results validate 753
695 on the known entries in the user-service matrix, and then that: a) users’/services’ posterior-neighborhood information 754
696 predicts missing entries (i.e., QoS data) in the matrix heavily is significantly effective for PLF to improve its prediction 755
697 based on resultant LFs [6], [17], [30], [38]. accuracy, b) PLF has significantly better performance than the 756
698 Since similar users usually experience similar QoS when state-of-the-art models in both prediction accuracy and pre- 757
699 invoking similar services, neighborhood information can be diction efficiency, and c) PLF relies only on the user-service 758
700 used to improve an LF-based QoS predictor’s prediction matrix and does not require any additional information. 759
701 accuracy. Zheng et al. first propose a neighborhood- Besides, PLF has also been demonstrated that it performs well 760

of
702 integrated LF-based QoS predictor based only on the user- in terms of both prediction accuracy and computational effi- 761
703 service matrix [16]. Next, they develop their model by inte- ciency on the large scale dataset. 762
704 grating the network map of users [15]. After that, Chen et al. Although PLF has shown promising prospects, several 763
705 propose an LF-based QoS predictor that incorporates the issues remain unveiled: 1) how to make the parameters of 764
706 knowledge of geographical neighborhoods [14], Ryu et al. PLF self-adaptive based on particle swarm optimization [55] 765
707 propose an LF-based QoS predictor by using the location so that they do not need to be inspected experimentally, 2) 766
708 information of users/service [11], and Feng and Huang pro- how to extend PLF to iteratively and dynamically identify 767
709 pose a neighborhood enhanced LF-based QoS predictor by users’/services’ neighborhood based on a deep-random- 768
710
711
712
713
714
715
716
717
ro
systematically considering geographical information, sample
set diversity computation and platform context [33]. These
proposed QoS predictors have the common point that the
neighborhood information, which is incorporated into an LF
model as the regularization term, is defined on sparse QoS
data and/or geographical information.
In this study, PLF is significantly different from existing
QoS predictors. First, existing predictors without neighbor-
forest structure [43], 3) as discussed in Section 3, the first two
steps of a PLF model focus on identifying neighborhood
based on an QoS matrix. Given that a QoS matrix can be
extremely sparse, it is necessary to investigate other efficient
neighborhood detectors like a random walk-based one [56],
[57] for a PLF model, and 4) as discussed in prior research
[48], [49], [50], [51], [52], [53], [54], modeling and acquisition
of temporal dynamics in QoS data is a vital issue in services
769
770
771
772
773
774
775
776
EP
718 hood regularization [6], [17], [30], [38] commonly suffer low computing society. It appears highly important to further 777
719 prediction accuracy because they ignore relationships among develop the temporal-aware extension of a PLF model. We 778
720 involved users/services. In comparison, PLF models such plan to address them in our future work. 779
721 relationships via constructing the neighborhood among them
722 accurately based on the dense LF matrices well representing
723 the known data of a sparse QoS matrix. Second, existing pre-
ACKNOWLEDGMENTS 780

724 dictors with neighborhood regularization [11], [14], [15], [33] This work was supported in part by the National Natural Sci- 781
725 require additional geographical information to construct ence Foundation of China under Grants 61702475, 61772493, 782
726 users’/services’ neighborhood, while PLF relies solely on the 61902370, 91646114, and 61602434, in part by the Natural 783
727 user-service matrix and does not require any additional infor- Science Foundation of Chongqing (China) under Grants 784
728 mation. It is expected to achieve even higher prediction accu- cstc2019jcyj-msxmX0578 and cstc2019jcyjjqX0013, and in part 785
IEE

729 racy for missing QoS data by combining heterogeneous by the Pioneer Hundred Talents Program of Chinese Acad- 786
730 neighborhoods originated from different data source. emy of Sciences. Di Wu and Qiang He are co-first authors of 787
731 Generally, the performance of a web service changes over this article.. 788
732 time because of variations of network conditions and service
733 status. Hence, it is necessary to make the QoS predictors
734 time-aware. To do so, Zhang et al. propose a time-aware per-
REFERENCES 789

735 sonalized QoS predictor by employing a tensor factorization [1] B. Bai, Y. Fan, W. Tan, and J. Zhang, “DLTSR: A deep learning frame- 790
work for recommendation of long-tail Web services,” IEEE Trans. 791
736 model [48]. Zhang et al. extend the model proposed in [48] by Services Comput., to be published, doi: 10.1109/TSC.2017.2681666. 792
737 considering the non-negative constraint [49]. Next, some [2] Z. Zheng, Y. Zhang, and M. R. Lyu, “Distributed qos evaluation 793
738 other time-aware QoS predictors are proposed, including for real-world web services,” in Proc. 2010 IEEE Int. Conf. Web 794
739 spatial-temporal model [50], time-aware collaborative filter- Services, 2010, pp. 83–90. 795
[3] Z. Zheng, H. Ma, M. R. Lyu, and I. King, “Wsrec: A collaborative 796
740 ing model [51], time-aware and sparsity-tolerant model [52], filtering based web service recommender system,” in Proc. IEEE 797
741 and time-aware and privacy-preserving model [53]. Besides, Int. Conf. Web Services, 2009, pp. 437–444. 798
742 Syu et al survey the recent progress in time-aware QoS pre- [4] L. Purohit and S. Kumar, “A classification based Web service 799
selection approach,” IEEE Trans. Services Comput., to be published, 800
743 dictors [54]. For PLF, it can also be extended to a time-aware doi: 10.1109/TSC.2018.2805352. 801
744 QoS predictor by considering the time information. [5] R. Lu, X. Jin, S. Zhang, M. Qiu, and X. Wu, “A study on big knowl- 802
edge and its engineering issues,” IEEE Trans. Knowl. Data Eng., 803
vol. 31, no. 9, pp. 1630–1644, Sep. 2019. 804
745 6 CONCLUSIONS [6] X. Luo, M. Zhou, Y. Xia, Q. Zhu, A. C. Ammari, and A. Alabdulwahab, 805
“Generating highly accurate predictions for missing QoS data via 806
746 In this paper, we propose a posterior-neighborhood- aggregating nonnegative latent factor models,” IEEE Trans. Neural 807
747 regularized latent factor (PLF) model for achieving highly Netw. Learn. Syst., vol. 27, no. 3, pp. 524–537, Mar. 2016. 808
 12 IEEE TRANSACTIONS ON SERVICES COMPUTING

809 [7] M. Alrifai and T. Risse, “Combining global optimization with [30] J. Zhu, P. He, Z. Zheng, and M. R. Lyu, “Online qos prediction for 885
810 local selection for efficient QoS-aware service composition,” in runtime service adaptation via adaptive matrix factorization,” 886
811 Proc. 18th Int. Conf. World Wide Web, 2009, pp. 881–890. IEEE Trans. Parallel Distrib. Syst., vol. 28, no. 10, pp. 2911–2924, 887
812 [8] C. Jatoth, G. Gangadharan, and R. Buyya, “Computational intelli- Oct. 2017. 888
813 gence based QoS-aware web service composition: a systematic [31] C. Wu, W. Qiu, Z. Zheng, X. Wang, and X. Yang, “QoS prediction 889
814 literature review,” IEEE Trans. Services Comput., vol. 10, no. 3, of web services based on two-phase k-means clustering,” in Proc. 890
815 pp. 475–492, May-Jun. 2017. IEEE Int. Conf. Web Services, 2015, pp. 161–168. 891
816 [9] D. Geebelen et al., “QoS prediction for web service compositions [32] A. Liu et al., “Differential private collaborative Web services QoS 892
817 using kernel-based quantile estimation with online adaptation of prediction,” World Wide Web, pp. 1–24, 2018, doi: 10.1007/s11280- 893
818 the constant offset,” Inf. Sci., vol. 268, pp. 397–424, 2014. 018-0544-7. 894
819 [10] K. Lee, J. Park, and J. Baik, “Location-based Web service QoS predic- [33] Y. Feng and B. Huang, “Cloud manufacturing service QoS predic- 895
820 tion via preference propagation for improving cold start problem,” tion based on neighbourhood enhanced matrix factorization,” J. 896
821 in Proc. IEEE Int. Conf. Web Services, 2015, pp. 177–184. Intell. Manuf., pp. 1–12, 2018, doi: 10.1007/s10845-018-1409-8. 897
822 [11] D. Ryu, K. Lee, and J. Baik, “Location-based Web Service QoS [34] X. Luo et al., “Incorporation of efficient second-order solvers into 898
823 prediction via preference propagation to address cold start latent factor models for accurate prediction of missing QoS data,” 899
824 problem,” IEEE Trans. Services Comput., to be published, IEEE Trans. Cybern., vol. 48, no. 4, pp. 1216–1228, Apr. 2018. 900
825 doi: 10.1109/TSC.2018.2821686. [35] M. Wang, X. Zheng, Y. Yang, and K. Zhang, “Collaborative filter- 901

of
826 [12] X. Chen, X. Liu, Z. Huang, and H. Sun, “Regionknn: A scalable ing with social exposure: A modular approach to social recom- 902
827 hybrid collaborative filtering algorithm for personalized web ser- mendation,” in Proc. 33rd Nat. Conf. Artif. Intell., 2017, pp. 1–8. 903
828 vice recommendation,” in Proc. IEEE Int. Conf. Web Services, 2010, [36] Y. Yang, Z. Zheng, X. Niu, M. Tang, Y. Lu, and X. Liao, “A location- 904
829 pp. 9–16. based factorization machine model for Web service QoS prediction,” 905
830 [13] Z. Zheng, H. Ma, M. R. Lyu, and I. King, “Qos-aware web service IEEE Trans. Services Comput., to be published, doi: 10.1109/ 906
831 recommendation by collaborative filtering,” IEEE Trans. Services TSC.2018.2876532. 907
832 Comput., vol. 4, no. 2, pp. 140–152, Apr.-Jun. 2011. [37] J. Wu, L. Chen, Y. Feng, Z. Zheng, M. C. Zhou, and Z. Wu, 908
833 [14] Z. Chen, L. Shen, F. Li, and D. You, “Your neighbors alleviate cold- “Predicting quality of service for selection by neighborhood-based 909
834 start: On geographical neighborhood influence to collaborative web collaborative filtering,” IEEE Trans. Syst., Man, Cybern., Syst., vol. 43, 910
835 service QoS prediction,” Knowl.-Based Syst., vol. 138, pp. 188–201, no. 2, pp. 428–439, Mar. 2013. 911
836
837
838
839
840
841
842
843
844
845
[15]

[16]

[17]
2017.

ro
M. Tang, Z. Zheng, G. Kang, J. Liu, Y. Yang, and T. Zhang,
“Collaborative web service quality prediction via exploiting
matrix factorization and network map,” IEEE Trans. Netw. Service
Manage., vol. 13, no. 1, pp. 126–137, Mar. 2016.
Z. Zheng, H. Ma, M. R. Lyu, and I. King, “Collaborative web service
qos prediction via neighborhood integrated matrix factorization,”
IEEE Trans. Services Comput., vol. 6, no. 3, pp. 289–299, Jul.-Sep. 2013.
X. Luo, M. Zhou, S. Li, Z. You, Y. Xia, and Q. Zhu, “A nonnegative
latent factor model for large-scale sparse matrices in recom-
[38] Y. Koren, R. Bell, and C. Volinsky, “Matrix factorization techniques
for recommender systems,” Comput., vol. 42, no. 8, pp. 30–37, 2009.
[39] D. Wu, X. Luo, M. Shang, Y. He, G. Wang, and M. Zhou, “A deep
latent factor model for high-dimensional and sparse matrices in
recommender systems,” IEEE Trans. Syst., Man, Cybern., Syst.,
to be published, doi: 10.1109/TSMC.2019.2931393.
[40] Z. Zheng, H. Ma, M. R. Lyu, and I. King, “Qos-aware web service
recommendation by collaborative filtering,” IEEE Trans. Services
Comp., vol. 4, no. 2, pp. 140–152, Apr.-Jun. 2015.
[41] S. Sedhain, A. K. Menon, S. Sanner, and L. Xie, “AutoRec:Autoen-
912
913
914
915
916
917
918
919
920
921
EP
846 mender systems via alternating direction method,” Trans. Neural coders meet collaborative filtering,” in Proc. 24th Int. Conf. World 922
847 Netw. Learn. Syst., vol. 27, no. 3, pp. 579–592, 2016. Wide Web, 2015, pp. 111–112. 923
848 [18] Y. Shi, M. Larson, and A. Hanjalic, “Collaborative filtering beyond [42] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, 924
849 the user-item matrix:A survey of the state of the art and future vol. 521, no. 7553, pp. 436–444, 2015. 925
850 challenges,” ACM Comput. Surveys, vol. 47, no. 1, pp. 1–45, 2014. [43] Z.-H. Zhou and J. Feng, “Deep forest: Towards an alternative to 926
851 [19] H. Chen and J. Li, “Learning multiple similarities of users and deep neural networks,” in Proc. 26th Int. Joint Conf. Artif. Intell., 927
852 items in recommender systems,” in Proc. IEEE Int. Conf. Data Min- 2017, arXiv:1702.08835. 928
853 ing, 2017, pp. 811–816. [44] D. Wu, X. Luo, G. Wang, M. Shang, Y. Yuan, and H. Yan, 929
854 [20] H. Zhao, Q. Yao, J. T. Kwok, and D. L. Lee, “Collaborative filtering “A highly accurate framework for self-labeled semisupervised 930
855 with social local models,” in Proc. IEEE Int. Conf. Data Mining, classification in industrial applications,” IEEE Trans. Ind. Informat., 931
856 2017, pp. 645–654. vol. 14, no. 3, pp. 909–920, Mar. 2018. 932
857 [21] S. Zhang, W. Wang, J. Ford, and F. Makedon, “Learning from [45] H. Li, K. Li, J. An, and K. Li, “MSGD: A novel matrix factorization 933
858 incomplete ratings using non-negative matrix factorization,” in approach for large-scale collaborative filtering recommender sys- 934
859 Proc. SIAM Int. Conf. Data Mining, 2006, pp. 549–553. tems on GPUs,” IEEE Trans. Parallel Distrib. Syst., vol. 29, no. 7, 935
IEE

860 [22] J. M. Hern andez-Lobato, N. Houlsby, and Z. Ghahramani, pp. 1530–1544, Jul. 2018. 936
861 “Probabilistic matrix factorization with non-random missing [46] X. Luo, H. Liu, G. Gou, Y. Xia, and Q. Zhu, “A parallel matrix fac- 937
862 data,” in Proc. 31st Int. Conf. Mach. Learn., 2014, pp. 1512–1520. torization based recommender by alternating stochastic gradient 938
863 [23] Y. Cai, H. Leung, Q. Li, J. Tang, and J. Li, “TyCo: Towards typical- decent,” Eng. Appl. Artif. Intell., vol. 25, no. 7, pp. 1403–1412, 2012. 939
864 ity-based collaborative filtering recommendation,” IEEE Trans. [47] L. Yao, X. Wang, Q. Z. Sheng, B. Benatallah, and C. Huang, 940
865 Knowl. Data Eng., vol. 26, no. 3, pp. 766–779, Mar. 2014. “Mashup recommendation by regularizing matrix factorization 941
866 [24] Y. Koren, and R. Bell, “Advances in collaborative filtering,” with API Co-invocations,” IEEE Trans. Services Comput., to be 942
867 Recommender Systems Handbook. Berlin, Germany: Springer, 2015, published, doi: 10.1109/TSC.2018.2803171. 943
868 pp. 77–118. [48] Y. Zhang, Z. Zheng, and M. R. Lyu, “WSPred: A time-aware per- 944
869 [25] M. Wang, X. Zheng, Y. Yang, and K. Zhang, “Collaborative filter- sonalized QoS prediction framework for Web services,” in Proc. 945
870 ing with social exposure: A modular approach to social recom- IEEE 22nd Int. Symp. Softw. Rel. Eng., 2011, pp. 210–219. 946
871 mendation,” in Proc. 33rd Nat. Conf. Artif. Intell., 2018, pp. 1–8. [49] W. Zhang, H. Sun, X. Liu, and X. Guo, “Temporal QoS-aware web 947
872 [26] X. Luo, Z. Liu, S. Li, M. Shang, and Z. Wang, “A fast service recommendation via non-negative tensor factorization,” in 948
873 non-negative latent factor model based on generalized Proc. ACM 23rd Int. Conf. World Wide Web, 2014, pp. 585–596. 949
874 momentum method,” IEEE Trans. Syst., Man., Cybern., Syst., to [50] X. Wang, J. Zhu, Z. Zheng, W. Song, Y. Shen, and M. R. Lyu, 950
875 be published, doi: 10.1109/TSMC.2018.2875452. “A spatial-temporal QoS prediction approach for time-aware 951
876 [27] X. He, L. Liao, H. Zhang, L. Nie, X. Hu, and T.-S. Chua, “Neural col- Web service recommendation,” ACM Trans. Web, vol. 10, no. 1, 952
877 laborative filtering,” in Proc. 26th Int. World Wide Web Conf., to be pp. 1–25, 2016. 953
878 published, doi: 10.1109/TSMC.2018.2872842. [51] C. Yu and L. Huang, “Time-aware collaborative filtering for QoS- 954
879 [28] L. Wu, P. Sun, R. Hong, Y. Ge, and M. Wang, “Collaborative based service recommendation,” in Proc. IEEE Int. Conf. Web Services, 955
880 neural social recommendation,” IEEE Trans. Syst., Man, Cybern., 2014, pp. 265–272. 956
881 Syst., pp. 1–13, 2018. [52] W. Chen, W. Qiu, X. Wang, Z. Zheng, and X. Yang, “Time-aware 957
882 [29] H. Wu, K. Yue, B. Li, B. Zhang, and C.-H. Hsu, “Collaborative QoS and sparsity-tolerant QoS prediction based on collaborative filter- 958
883 prediction with context-sensitive matrix factorization,” Future ing,” in Proc. IEEE Int. Conf. Web Services, 2016, pp. 637–640. 959
884 Gener. Comput. Syst., vol. 82, pp. 669–678, 2018.
 WU ET AL.: POSTERIOR-NEIGHBORHOOD-REGULARIZED LATENT FACTOR MODEL FOR HIGHLY ACCURATE WEB SERVICE QOS PREDICTION 13

960 [53] L. Qi, R. Wang, C. Hu, S. Li, Q. He, and X. Xu, “Time-aware Mingsheng Shang received the PhD degree 1023
961 distributed service recommendation with privacy-preservation,” in computer science from the University of 1024
962 Inf. Sci., vol. 480, pp. 354–364, 2019. Electronic Science and Technology of China 1025
963 [54] S. Yang, C. M. Wang, and Y. Y. Fanjiang, “A survey of time-aware (UESTC). He joined the Chongqing Institute of 1026
964 dynamic QoS forecasting research, its future challenges and research Green and Intelligent Technology, Chinese Acad- 1027
965 directions,” in Proc. Int. Conf. Services Comput., 2018, pp. 36–50. emy of Sciences, Chongqing, China, in 2015, 1028
966 [55] B. Chopard and M. Tomassini, “Particle swarm optimization,” An and is currently a professor of computer science 1029
967 Introduction to Metaheuristics for Optimization. Berlin, Germany: and engineering. His research interests include 1030
968 Springer, 2018, pp. 97–102. data mining, complex networks, and cloud com- 1031
969 [56] M. Tang, X. Dai, B. Cao, and J. Liu, “Wswalker: A random walk puting and their applications. 1032
970 method for QoS-aware Web service recommendation,” in Proc.
971 IEEE Int. Conf. Web Services, 2015, pp. 591–598
972 [57] M. Jamali and M. Ester, “Trustwalker: A random walk model for Yi He received the BE degree in transportation 1033
973 combining trust-based and item-based recommendation,” in Proc. engineering from the Harbin Institute of Technol- 1034
974 15th Int. Conf. Knowl. Discovery Data Mining, 2009, pp. 397–406. ogy, China,and the MS degree in civil engineering 1035
from the University of Louisiana at Lafayette, USA, 1036
in 2013 and 2017, respectively. He is currently 1037

of
975 Di Wu (SM’18–M’19) received the BS degree in working toward the PhD degree in computer sci- 1038
976 applied physics from the Nanjing University of ence from the University of Louisiana at Lafayette, 1039
977 Science and Technology, Jiangsu, China, in 2009, USA. His research interests include optimization, 1040
978 the MS degree in optical engineering from the machine learning, and data mining. 1041
979 Chongqing University, Chongqing, China, in 2012, 1042
980 and the PhD degree in computer science from
981 Chongqing Institute of Green and Intelligent Tech-
982 nology, Chinese Academy of Sciences, Chongq-
Guoyin Wang (M’98–SM’03) received the BE deg- 1043
983 ing, China, in 2019. He was a visiting scholar
ree in computer software, in 1992, the MS degree in 1044
984 from April 2018 to April 2019 at the University of
985
986
987

988
989
990
991
992
is a member of the IEEE.

ro
Louisiana at Lafayette, USA. His research inter-
ests include data mining and machine learning. He

Qiang He received the first PhD degree from the

Swinburne University of Technology, Australia, in
2009, and the second PhD degree in computer
science and engineering from the Huazhong
University of Science and Technology, China,
computer software, in 1994, and the PhD degree in
computer organization and architecture, in 1996,
from Xi’an Jiaotong University, Xi’an, China. He
worked at the University of North Texas, USA, and
the University of Regina, Canada, as a visiting
scholar during 1998-1999. Since 1996, he has been
working at the Chongqing University of Postsand
Telecommunications, where he is currently a profes-
sor and PhD supervisor, the director of the Chongq-
ing Key Laboratory of Computational Intelligence, and the dean of the
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
EP
College of Computer Science and Technology. His research interests 1055
993 in 2010. He is currently a senior lecturer with include data mining, machine learning, rough set, granular computing, cogni- 1056
994 Swinburne. His research interests include service tive computing, etc. He is the president of International Rough Set Society 1057
995 computing, software engineering, cloud computing (IRSS), a vice-president of the Chinese Association for Artificial Intelligence 1058
996 and edge computing. He is a senior member of the (CAAI), a fellow of the China Computer Federation (CCF) and a senior 1059
997 IEEE. For more information, please visit https:// member of the IEEE. 1060
998 sites.google.com/site/heqiang/.

" For more information on this or any other computing topic, 1061
999 Xin Luo (M’14–SM’17) received the BS degree in please visit our Digital Library at www.computer.org/csdl. 1062
1000 computer science from the University of Electronic
1001 Science and Technology of China, Chengdu,
1002 China, in 2005, and the PhD degree in computer
1003 science from Beihang University, Beijing, China, in
1004 2011. In 2016, he joined the Chongqing Institute of
IEE

1005 Green and Intelligent Technology, Chinese Acad-

1006 emy of Sciences, Chongqing, China, as a professor
1007 of computer science and engineering. He is cur-
1008 rently also a distinguished professor of computer
1009 science with the Dongguan University of Technol-
1010 ogy, Dongguan, China. His current research interests include big data anal-
1011 ysis and intelligent control. He has published more than 100 papers
1012 (including more than 30 Ieee Transactions papers) in the above areas. He
1013 was a recipient of the Hong Kong Scholar Program jointly by the Society of
1014 Hong Kong Scholars and China Post-Doctoral Science Foundation, in
1015 2014, the Pioneer Hundred Talents Program of Chinese Academy of Sci-
1016 ences, in 2016, and the Advanced Support of the Pioneer Hundred Talents
1017 Program of Chinese Academy of Sciences, in 2018. He is currently serving
1018 as an associate editor for the IEEE/CAA Journal of Automatica Sinica,
1019 IEEE Access, and Neurocomputing. He has received the Outstanding
1020 associate editor reward of IEEE ACCESS, in 2018. He has also served
1021 as the Program Committee Member for more than 20 international
1022 conferences. He is a senior member of the IEEE.

View publication stats