Appl. Sci. 20 20 , 10 , x; doi: FOR PEER REVIEW www .mdpi.com/journal/ applsci [620398]

Appl. Sci.

20
20
,
10
,
x; doi: FOR PEER REVIEW

www
.mdpi.com/journal/
applsci

Article

1

Cooperation Based
Proactive Caching in Multi
–
tier
2

Cellular Networks

3

Fawad Ahmad

1
,
Ayaz

Ahmad
1
, Irshad

Hussain

2,
*
,
Peerapong

Uthansakul

3
,
*

and Suleman Khan
2

4

1

Department

of Electrical and computer Engineering,

COMSATS University

Islamabad,

Wah Campus, Wah

5

C
antt 47040,

Pakistan
;

6

2

Faculty of Electrical & Computer Engineering, University of Engineering and Technology Peshawar, 25000
7

Pakistan

(I.H.)
;

8

3

Sch
ool
of
Telecommunication Engineering, Suranaree University o
f Technology, Nakhon Ratchasima
9

30000,
Thailand

(P.U.);

10

11

*

Correspondence:

[anonimizat]
;
[anonimizat]
;

12

Received: date; Accepted: date; Published: date

13

Abstract:
The scarce caching capacity of the cache enabled local Base stations (BSs) mitigate the cache
14

hit ratio (CHR) and user satisfaction ratio (USR). However, Cache enable multier cellular networks
15

h
ave been presented as a promising candidate for fifth generation networks for much higher CHR
16

and USR through densification of networks by deploying smaller networks. In addition to this, the
17

cooperation among the BSs of various tiers for cached data trans
fer
intensify its significance many
18

folds .Therefore, in this paper we consider CHR and USR maximization of a multier cellular network.
19

We formulate a CHR and USR p
roblem for multitier cellular
networks.

W
e put major constraints on
20

caching space of BSs of
each tier. The unsupervised learning algorithms such as K
–
mean clustering
21

and col
laborative filtering are used
for clustering the similar BSs in each tier and estimating the
22

content popularity respectively.
A
novel scheme such as cluster average popularity

based
23

c
ollaborative filtering
(
CAP
–
CF
)

algorithm is used to cache the data and maximize the CHR in each
24

tier. Similarly, two novel method
s such as intra tier and Cross

tier cooperation (ITCT) and modified
25

ITCT algorithms are employed in order to optimize
the US
R. Simulations results
witnesses,

that the
26

proposed schemes yields significant performance in terms of average cache hit ratio and user
27

satisfaction ratio compared to conventional approaches
.

28

K
eyw
ords:

caching
;

cooperative network;

multi
–
tier cellular
network
;

content popularity
;
machine
29

learning
;

30

1.
Introduction

& Background

31

Recently the smart phone usage is enormously increased which result into exponential growth
32

of mobile data traf
fic. This Mobile data traffic
pos
s
es
s

an unbelievable challenge in the radio resource
33

demand [1], [2]. CISCO predicts that mobile traffic will increase 8 times from 2015
–
2020 [3] and would
34

reach to 30.6

Exabyte p
er month by the year 2020[4]. It is observed that a big chunk of increased data
35

traffic is due to duplicate downloads of famous videos files from the remote server [5]. In the recent
36

era, mobile media users are extensively sharing
information and their thoughts via
cross
net. The
37

mobile social media services such as Facebook and Twitte
rs have also increased

the data traffic
38

exponentially [6].
This unbelievable mobile data traffic has brought burden on the communication
39

networks as these are limited in the transmission rate and hardware capabilities. Also the rapid
40

growth of the contents

in the Web for example
cosse
tting

articles, games or other type of
41

entertaining contents such as videos, audios etc. have now started problems regarding quality and
42

downloading for the users [7
],

[8]. Thus, Proactive caching

performs a pivotal role in th
e upcoming
43

5G technology. Caching the popular files on nearby base station and then serving the users from the
44

Appl. Sci.
20
20
,
10
, x FOR PEER REVIEW

2

of
18

local cache are now becoming advantageous due to reducing latency and the data traffic load
45

especially on the back haul.

46

On the other hand, Machi
ne learning tool brings a new research area for the modeling and
47

prediction of the user behavior for proactive caching decisions [9]. Recently the stupendous success
48

of deep learning methods in pattern recognition, images and natural language have attracte
d the user
49

to use them for network optimization [10], [11]. Different machine learning tool can extract
50

information in the data traffic and cache the contents at the local base station [10]. For example,
51

extreme learning machine predicts the popularity of
the contents based on the
different meta features
52

of the contents [12]. Unsupervised learning such as K
–
mean clustering and Collaborative filtering
53

methods are employed to cluster the users with similar
cross
est for the contents and then cache the
54

popular
contents accordingly [11], [13]. Similarly ad
aptive caching scheme optimizes

the network
55

performance thereby decreasing the delay

and back haul load [12], [13].
In [14], the author suggested
56

the multi
–
armed bandit algorithm to estimate content popularity a
nd the authors in [15] proposed an
57

extreme
–
learning machine algorithm for estimation of the future content popularity.

58

Similarly, reinforcement learning methods are employed in cache enabled networks [16], [17]
59

and [18]. For example, in [16] a multi
–
step r
eturn actor
–
critic architecture which is reinforcement
60

learning agent is used for optimizing the cache hit ratio.

The information content can be extracted
61

from the data using a (LSTM) deep learning [19]. The contents are cached if the
cross
est of the data
is
62

known. This decreases the service latency. These contents are most likely requested by the users.

63

Deep

CachNet is a Proactive caching framework in cellular network using deep learning
64

algorithm [20]. In this caching scheme, a huge

amount of data is coll
ected from the end mobile devices
65

of users connected to SBSs. In [21], the author solve the problem of context
–
aware data caching in the
66

heterogeneous small cell networks using deep learning algorithm. The author in [22], deals with the
67

problem of minimiza
tion average energy cost in cache enabled networks especially in limited cache
68

memory nodes using reinforced learning algorithm. Similarly multi
–
agent multi
–
armed bandit can be
69

applied to the caching problems regarding the device to device communication. S
uch design applied
70

the Q
–
learning in order to coordinate the caching decisions [23].

71

The author in paper [24], strives to jointly minimize the average transmission delay in cache
72

enabled networks by jointly considering scheduling and the caching problem us
ing reinforced
73

leaning and deep learning. In [25], the author optimizes multimedia service in 5G caching networks
74

based on semantic information of user. Content popularity is estimated by the singular value
75

decomposition (RSVD) which is based on collaborat
ive filtering (CF).

In [26], the author optimize the
76

cache hit ratio, and a decision matrix is formed from a huge available data. In paper [27], proposed
77

caching scheme for the ultra
–
dense networks. Backhaul load minimization problem is solved but k
–
78

Neares
t Neighbors (k
–
NN) classification.

79

In this article, we propose a

framework for
content

caching

and using satisfaction rate using
80

cross

tier and intra tier cooperation among the base stations.

O
ur

prominent

contributions
are as
81

under
:

82

•
Introducing a data
caching system; where data is cached

in every tier of th
e network, thereby,
83

Optimizing
the
CHR
and
USR
utilizing classical method of Collaborative filtering and
K
–
mean
84

clustering.

85

•

The requests
for different contents are considered
and

a popularity matri
x is formed in each
86

tier, which
is highly sparse in nature. Proactive caching is done by creating clusters of base stations
87

using the unsupervised K
–
mean clustering in each tier and fill half of the cache memory with the
88

most popular contents

of the cluste
r
and the rest of vacant cache is occupied by the contents that will
89

be requested in future predicted by collaborative filtering

C.F
.

90

•P
resent
ing

two novel methods for the user satisfaction rate. The first method deals with the
91

clusters made by K
–
mean clustering, based on contents
requests, in

each tier and content cashing is
92

done by C.F while user satisfaction is achieved thorough
cross

tier and i
ntra cooperation among the
93

different tiers.

94

•In the second method, initially contents are cached in all tier based on CF algori
thm. Then
95

clusters are created
using K
–
mean algorithm while taking all the base station into account,
96

Appl. Sci.
20
20
,
10
, x FOR PEER REVIEW

3

of
18

irrespective of which tier
they belong to .The cluster formed may have base station
s

of different tiers
97

.The second method is faster one as ones requests don’t require intra tier
and X
–
tier flow
for the
98

fulfillment of their content satisfaction.

99

• Extensive simulations are conducted

with different contents and their requests depends on the
100

mode of the users to elaborate the effectiveness of the proposed algorithms. It is shown that by using
101

learning algorithms, caching the data on base station in the different tier, considerably impr
ove the
102

cache ratio and the user satisfaction ration.

103

2.
System Model

104

Su
ppose a modern multitier cache enabled
cellul
ar

network consisting of
t tiers of M base
105

stations, denoted
as M
t
= {1,…, M
t
}, serve N
t

user terminals (UTs) from the set N
t
= {1,…,N
t
}

deployed
106

in a geographical area that consists of residential area,

factories,

colleges and schools

etc
. As the
107

network
is
multitier
so it comprises of Pico BSs, micro BSs, and a single Macro BS to serve greater
108

number of users. Each
BS
is equipped with it
s limited physical storages S
mt

and the storage
109

capacity of
the base stations is denoted by
,

S={S
11
,S
21
,…..Sm
1
,S
12
,S
22
,……S
13
,S
23
….,S
m3
}.

110

All the BSs

are connected through limited capacity optical fibre links.

When the users request

111

the contents and if, t
he contents are cached in
the associated base station the
n it is served otherwise
112

it is
ro
uted to the other base station in the same tier or
another tier

or the contents are fetched from
113

the content servers via
limited capacity back haul. The users request

a variety of contents such as
114

music, movie
s,

breaking news, face book, twitter
s, technical books, notes etc.
from a

content
library
115

F={1,…,…..,F}, where each content f has a size of L(f)Byte and bit rate requirement of B(f) M

byte/
s.
116

The
C
ontent serve
r

caches
F Bytes of data in its storage. In such a scenario, each base station
117

proactively caches selected contents from the library F during off
–
peak hours. Let P
t
(
t
)
∈

R
M×F

is the
118

content popularity matrix which describe the history of

requested contents

at time t, where each
119

coefficient p
m
t
,

f

(
t
)

represents the frequency of requesting the content f received by the
BS
mt

of tier t.

In
120

fact, the matrix P
t
(
t
) is the local content popularity distribution observed at base BS
mt
, whereas the
121

global popularity
distribution of the all the contents is described by the Zipf distribution PF(f),
∀
f
∈
F
122

i.e. lesser contents have greater popularity and greater contents are less popular. Moreover, we
123

consider that all SBSs

of all tiers
have stationary content popularity
matrix over T time slots, thus we
124

represent P(t) by P. Using tools from unsupervised machine learning and collaborative filtering
125

(CF)
,
we propose a proactive caching Procedure using P to determine which content should be cached.
126

To show this, let us defi
ne the cache decision matrix of BSs as X
t
(t)
∈
{0,1}
Mt×F
, where the entry x
mt,f

127

takes 0 if the f
th

content is not cached at m
th

BS

at tier t

in
time
t
, and 1 otherwise. We suppose that the
128

cache placement is done during off
–
peak hours, therefore X
t
(
t
) is
represented by a cache decision
129

matrix X
t

which remains static during peak hours .The second decision variable is Y which is user
130

satisfaction

ratio

parame
ter whose value is one

if request is served other wise 0.

131

2.1. Caching Model

132

In our model, every tie
r has
its

own users and they have their demands according to their
133

preferences. In particular, we consider such caching system where users download different types of
134

contents e.g
.

a university student will have to down load books, lecture notes and video
lecture
135

relevant to his fields. People
interest
ed

in current affairs will watch breaking news.

we consider
136

d
ifferent types of files and th
e BS that receive request for similar data will be virtually clustered,
137

which
will discussed i
n detail later on. In ou
r model.
The caching capacity of Pico BS is lesser
138

compared to micro BS and Macro BS. On the other hand, Macro base station will cache more files as
139

compared to micros and Picos. Similarly number of users for pico is lesser as compared to micro and
140

Macro .
In the multitier system, there will be more pico then micro and single macro. As the capacity
141

is limited, so only popular contents can be store that the associated users requested on its local base
142

station. Thus the base station will offload the traffic th
ereby reducing traffic on the back haul and
143

improve the quality of services.

144

Let us consider a base station BS
m

receives a set of requests R (m) for certain data during
145

time
inter
val T seconds .These requests are Zipf
–
like distribution in nature. Let X be the cache decision
146

Appl. Sci.
20
20
,
10
, x FOR PEER REVIEW

4

of
18

matrix that depends on the popularity matric P of the content and C (m) is the set of cached contents
147

according to then
the cache hit ratio is given in Equation

1 as under.

148

CHR=

∑

(

)
∩

(

)

∑

(

)

(
1
)

149

150

In t
–
tier wireless

communication system, each tier will require its cache to be filled by

the
151

requested contents.

The overall or the average cache hit ratio of the multi
–
tier system can be modelled,
152

as

153

α=

∑
∑

(

)
∩

(

)

∈


∈

∑
∑

(

)

∈


∈

(
2
)

154

Each tier has its own users and cache decision
matrix X
t
which depends on

the popularity matrix
155

P
t

(t)
∈

R
M×F

156

X
mt
=

1

if

content

f

is

cached

at

BS
m

of

tier

t

,
∀
m

ϵ

M
t
0

otherwise

157

158

2.2. User Satisfaction Ratio Model

159

In such a model, the user can get the
data from
locally associated BS

cache storage unit,
remote
160

BS cache storage unite but within the same
tier,
remote BS cache storage unit

but in

another tier and

161

content server.
The Remote content unit is the other BS that cache
s

the user’s required contents but
is
162

not associated to the user. It

means

there
is
cooperation among the BS not only with in the tier but
163

also across the tiers. Also, within the tier there will be intra cluster and
cross

cluster cooperation
164

which will be discuss in Section V
I

in detail.

165

The average user satisfaction can be mode
lled as

166

167

Ƞ
(y) =

∑

=
Average request satisfaction ratio

(
3
)

Where as

168

Y= {y
1
, y
2
, y
3
, y
4
… y
RT
}

169

R
T
=

∑
∑

(


)

= Total requests

170

y
r
=

1

if

request

r

is

served

o

otherwise

171

In order to serve the
user content request, two methods are proposed which will be discussed in
172

Section V.

173

3
.
Problem Formulation

174

Given this model, our prime goal is to develop a proficient caching approach to improve the cache
175

hit ratio and the user satisfaction ration based
on the user content requests and prediction of those
176

contents that will be requested in future. To achieve this goal, we formulate an average cache hit ratio
177

and user satisfaction optimizati
on problem whose objective is
to optimally cache
the popular
contents
178

on BS of each tier in a such way that

user find the intended data instantly
there by increasing the cache
179

hit ratio and user satisfaction ration in each tier of the cache enabled multi
–
tier network.

180

181

Max X
t,

Y
α

+

Ƞ
(y)

(
4
)

Subject to constraints

182

,

∈
{
0
,
1
}
,
∀

, t, f

(
5
)

Appl. Sci.
20
20
,
10
, x FOR PEER REVIEW

5

of
18

∈
{
0
,
1
}
,
∀

(
6
)

∑
L
(
f
)


,


∈

≤

,
∀

,

∈

(
7
)

The 1
st

term of the objective function accounts for the average cache hit ratio of the multitier
183

cache enabled network .The 2
nd

term of the equation
(4)

shows the average satisfaction ratio.
184

In other words, the system is trying to optimize the usage of the limite
d cache of the BS in
185

such a way that the user can easily find its requested content from its associated or remote BS
186

thereby increasing the downloading speed and decreasing the burden on the backhaul. The
187

caching capacity required for each
BS
is ind
icated
by the system constraint (7)

188

This means that the cache entity capacity is bounded by corresponding capacity limit.
189

Constraint(5) and Constraint(6) show

that X
m,t,f
and y
r
are binary variables respectively .Therefore, the
190

problem(1) is a constrained integer programming problem and is generally NP
–
hard which is very
191

challenging
task and difficult to solve
.

Also storing the most popular data in the cache can increase the
192

n
umbe
r

of users which are recipient of the c
ache
d

data
.

so caching the data that favours the users in
193

limited cache is a challenging job. For this such an intelligent system is require that predict the
194

popularity of the data and cache them properly. For this ma
chine

learning plays a pivotal role. With
in
195

this frame work, based on past requests we will first use K means clustering for clustering the BSs that
196

got similar requests and for predicting the future data collaborative filtering is used. This will enhance

197

the cache hit ratio. Similarly, for maximizing user
satisfaction ratio, two types methods are employed
198

which are based on mutual cooperation among the tier will be studied in
section VI
.

199

4.
Collaborative Filtering for Content Prediction

200

The optimization pr
oblem in (1) is a challenging issue to resolve because the cache hit ratio and
201

user satisfaction depends on the proactive caching which in turn depends of the prediction of the
202

contents which is not requested now but hopefully requested in future .Since th
e BS need to aware
203

of the content popularity in advance. The optimization problem is very difficult to solve using
204

conventional optimization algorithm that are unable to predicts the content requests in advance for
205

the BS especially in dense environment wh
ere there are a lot number of users and limited cache size
206

of the cache entity. This reason renders the optimization problem in (1) very challenging as the BS
207

got very limited information about the users preferences.to address this challenge, we propose a
K
208

means clustering for cluster the BS that receive the similar request for the contents. After this, the
209

contents whose request are not received on BS abut it is expected that for those Contents requests
210

will be received in future then they are predicted b
y the CF as this is used for the prediction of the
211

contents on the similar BS.

212

4.1. K
–
mean clustering

213

In this subsection, the k
–
mean algorithm
is
applied to the past requests re
ceived on the BS in
214

each tier. K
–
mean clustering is an algorithm that tries to partition data set into K predefined distinct
215

non
–
overlapping subgroups clusters where each data point belong to only one group As the content
216

popularity matrix gives the frequency of the request on the BS. Hen
ce, those BSs that got similar
217

requests for the contents on the BSs, are grouped together to form a
cluster. As Micro BSs comprise
218

of greater number of users as compared to Pico therefore the request on the former will greater as
219

compared to later. Also, t
here will be more Pico BS as compared to micro therefore the cluster of Pico
220

cells will be more crowded as compared with micro BS. So clusters will be formed in each tier. The
221

algorithm has following steps

222

1)1
st

all of specify the number of the cluster in each tier

223

2) Initialize centroids by first shuffling the data set which is the past requests on the base stations and
224

then randomly selecting K data points for the centroids without replacement

225

3) Keep iterati
ng until there is no change of data points to cluster isn’t changing

226

4) Compute the sum of the squared distance between data points and all centroids

227

5) Assign each data points to the closest cluster (centroid)

228

Appl. Sci.
20
20
,
10
, x FOR PEER REVIEW

6

of
18

6) Compute the centroids for the cluster by
taking the average of the all data points that belong to
229

each cluster.

230

4.2. Content Prediction

231

In this subsection, we formulate the CF based content request distribution prediction algorithm.
232

The prediction is based on the idea that the people who have agr
eed on something in the past may
233

also agree in the future. It means that the people who have similar taste may get best
234

recommendations from one another. We can apply this idea to our multitier

cellular caching

network.
235

If we find the similarity between th
e two BSs then, we can predict the popularity of the contents on
236

the BSs that they have not requested by now but will most probably be requesting in the future.

237

Let us consider A
t
(mi) be the set of the contents requested by the users of the BS
t
,

mi
and A(m
j) be
238

the set contents requested by the B
S
t,
mj
.
A
t
(f) be the set of the BSs on which file f is requested .Then to
239

find the similarity between two BSs we have to find the weighted matrix W
i,j
which is given as

240

In matrix form W

241

W
T
=

1

⋯




⋮
⋱
⋮

⋯
1

242

The matrix is obtained from the following equation

243

,


=

.
1
log

(
1
+
|


(

)
|

|


(

)
|
|

(

)
|


∈


(

)
∩


(

)

(
8
)

Now the probability that the content f is requested by the BS
mi
is obtained as

244

P
t
(m
i,f)

=
∑
1
.



∈

(

,

)
∩


(

)


,


∗


,

(
9
)

Where Z (mi, G) contains the most similar G SBSs with mi. P
mj,f
is the entry of the matrix P. After
245

calculating p (mi,f) of each
content for the SBS mi, we can cache the requested contents on

the BS
246

based on p(mi,f) in decreasing order.

247

5
.
Proposed Framework

248

The prediction of the contents requested in future must now leverage to determine which
249

content is to be cache at the BS in ea
ch ti
er.
Hence, Cache hit ratio is maximized by not only placing
250

the required contents, requested by the users now but cache the future contents to be requested in
251

future
.

If the respective base station of a user get the requested file in the local base st
ation then the
252

cache is hit or otherwise miss. But due to limited capacity of the base station, cache hit ratio cannot
253

be achieved its maximum value 1 but if the requested contents and off course, the future requested
254

contents are placed then optimum value

of the cache hit ratio can be obtained. We have here different
255

tiers compris
es of Pico BSs, micro BSs and a Macro. The average cache hit ratio is maximize by taking
256

all tier into account.

257

In this
section we formulate the proposed algorithm to the
problem
in (3
). First of all we clustered
258

the BSs in each tier based on past requests and initialize the capacity matrix S
mt

and cache decision
259

matrix
X
t

Then in cluster each, we find the average request of the files and sort them in descending
260

order. Now fill hal
f of the cache of each BSs with the contents whose average requests is greater and
261

set the entry in the X
t

matrix equal to 1 and update the cache size and repeat it till size of the BS
262

reaches 50% the maximum size. Rest of the cache is filled by the conte
nts that will be requested

in

263

future predicted by CF. Wights

similarity
are find out using
equation (8
) and

P
t
(m
i,f)

using
equation
264

9

and are sorted in descending order then if file f is not in the SBS m and the BS current cache plus
265

the file size is lower than maximum capacity than cache it and set the entry Xm,j of the cache
266

decision Xt

and update the current cache of the SBS m. Repe
at this until reaching the SBS caching
267

capacity. Finally return the cache decision matrix X
t
.

268

269

Appl. Sci.
20
20
,
10
, x FOR PEER REVIEW

7

of
18

Algorithm 1 Contents Caching

270

Input: P
1
, P
2
, P
3
, S1, S2, S3, L
1
, L
2
, L
3
, F
1
, F
2
, F
3

271

Output: X
mt

272

1.

Cluster the based stations k=1, 2, 3….Hk based on past requests
in each tier except Macro
273

cell

274

2.

For Pico BS do

275

3.

Initialize the cache size S
1

and caching matrix X1

276

4.

[a,b] = SORT

AR
.

where AR is the average popularity of the most requested file. SORT AR
277

by descending order

278

5.

For BSm in H
K
do

279

6.

For i=1………..F do

280

7.

Gets index of
i
th

highest AR in cluster k,

281

8.

j
←
bi

282

9.

if L
1
j +
Sm1 ≤ 0.5
∗
Sm then

283

10.

X m,j = 1

284

11.

Sm1 = L
1
j + ˆ Sm1

285

12.

else

286

13.

break

287

14.

Calculate the similarity matrix i.e. Wi,j

288

15.

Find the P(mit,f) and make TM matrix

289

16.

[a,b]
←

SORT(Tm)
. Sorts TM by descending order, returns a, b as ordered values and
290

indices.

291

17.

Repeat 6
–
9

292

18.

if L
1
f + ˆ Sm1 ≤ Sm1 then

293

19.

if Xm,f =1then

294

20.

Make Xm,f = 1

295

21.

Sm1 = L
1
f + ˆ Sm1

296

22.

else

297

23.

continue

298

24.

else

299

25.

break

300

26.

return X1

301

27.

For Micro BS

302

28.

Repeat 3
–
26 and rep
lacing 1 subscript with 2 and
return X
2

303

29.

For Macro BS

304

30.

Initialize S
3

305

31.

[a,b]=SORT CP by descending order, where CP content popularity.

306

32.

For i=1
––––
F
3

do

307

33.

get index of the highest CP

308

34.

j
←
bi

309

35.

if Lj + S3 ≤ S3 then

310

36.

X m,j = 1

311

37.

S3 = L
3
j + S
3

312

Appl. Sci.
20
20
,
10
, x FOR PEER REVIEW

8

of
18

38.

else, Break

313

39.

Return X
3

314

40.

Form X
t
from X
1
,X
2
,X
3

315

41.

Calculate average
cache hit ratio

316

6. User Satisfaction Ratio

317

The second goal of our research work is to optimize the user satisfaction ratio.

If a user
request for a
318

particular content then it is served by the
local base station. If a base station does not have that content,
319

then it is fetched fro
m neighbouring

base station within the same cluster but if it is still not in the
320

same cluster, then it is brought from another cluster in the same tier. Also if it i
s not in the same tier
321

then it is served from higher tier if that content exist there otherwise it is fetched from the main server.
322

The above process is a time consuming and require surplus handover etc. so besides this method a
323

more robust
method is also
presented.

324

325

MNO
Core
Internet
SP
1
SP
2
SP
3
SP
4
twitter
facebook
netflix
youtube
Macro
Micro
Pico
Femto
Femto
Femto
Femto
Femto
Femto
Femto
Femto
Femto
Femto
Femto
Femto
Femto
Femto
Femto
Femto
Pico
Pico
Pico
Pico
Pico
Micro
k
_
means C lustering based on the passed request on base station Femto
,
pico
,
micro and macro
Micro
Pico
Femto
Femto
Femto
Femto
Femto
Femto
Femto
Femto
Femto
Femto
Femto
Femto
Femto
Femto
Femto
Femto
Pico
Pico
Pico
Pico
Pico
Micro
MNO
:
Mobile network Operator
SP
:
Service Provider

326

Figure 1.

Cluster formation in each tier using K mean clustering
.

327

6
.1.
Intra

tier and
Cross

tier based cooperation
(ITCTC)

328

The objective

is to opti
mize the user demand instantly.F
or this,there is cooperation among the BS
329

regarding the content sharing.T
his coper
ation is not limited to the intra tier.The assocuated BS

of
330

one tier can demand the requested content from the
other
tiers.

331

6.1.1 intra tier cooperation

332

This cooperation can be further diveded in to two types

333

(1)

Intra cluster cooperation (2)
cross

cluster cooperation

334

Appl. Sci.
20
20
,
10
, x FOR PEER REVIEW

9

of
18

6.1.1.1

Intra
–
Cluster cooperation
(ICC)

335

Clusters are formed in each
tier and contents are cached
accord
ingly. Fig
ure.
1 shows clusters
336

formation in

Tier 1 and Tier 2.

The

user request may be fulfilled by the nearby
associated base

station
337

o
r the contents may

be

fetch

be from

the other base station with in
the
cluster .The cluster comprises
338

of similar
contents on the base station but probably some base station have different contents within
339

the cluster. The user will send request to the local base station and if it is not satisfied then the request
340

will be forwarded to the other base station within the
cluster.

341

6.1.1
.
2
Cross
–
Cluster Cooperation

(CCC)

342

If a user does not find its requested contents in
associated cluster
then request is forwarded to the
343

other cluster of the same tier. This is
cross

cluster cooperation

344

6.1.
2
.
Cross
–
Tier Cooperation

(CTC)

345

The
cross

tier cooperation means that if the requested contents is not found in the local tier then
346

the contents should be routed from the other tier for example if a user does not get the contents from
347

the
Pico

cell then it
will be request the data to the

M
ic
r
o cell and so on, also t
he M
ic
r
o cell base station
348

are clustered and request is forwarded from one cluster to another cluster.

349

Algorithm 2
:

I
ntra
T
ier and
C
ross
T
ier
C
ooperation

350

Input:

Xt

351

Output: Yr

352

1.

Cluster the based stations into k=1, 2, 3….Hk
based on past requests in each tier except
353

Macro cell

354

2.

Cache the data using CF in each cluster of each tier

355

3.

For Pico BS
do

356

4.

If

request arrives at Pico BS then

357

5.

Serve the request at associated Pico BS

358

6.

Else if

359

7.

Send the request
to
neighboring

BS but within the
same cluster

360

8.

Else if

361

9.

Send the request to Pico BS which is in any of the
neighboring

clusters
(
Cross
–

362

cluster cooperation)

363

10.

Else if

364

11.

Send the request to 1
st

Cluster of micro BS

365

12.

Else if

366

13.

send it to
neighboring

clusters of micro BS

367

14.

Else if

368

15.

send it Macro BS

369

16.

Else

370

17.

send it content server

371

18.

End

372

19.

For Micro BS do

373

20.

If

request is covered in Micro BS then

374

21.

Serve the req
uests at associated Micro BS

375

Appl. Sci.
20
20
,
10
, x FOR PEER REVIEW

10

of
18

22.

Repeat steps 5
–
9

376

23.

Else if

377

24.

send it Macro BS

378

25.

Else

379

26.

send it content server

380

27.

End

381

28.

Repeat for all requests

382

29.

Return Y

383

384

6.2.

Modifed

Intra tier and
Cross

tier cooperation

(MITCTC)

385

The above method
is a time consuming and requires extensive signalling and hand off
386

because the request has not only to be passed through all cluster of each t
ier but also from one tier to
387

other tier. The proposed method is that initially the popular contents are cached
on the base station
388

utilizing
CF
.
T
he
n

cluster
s are formed based on the similar
file cached
in all tiers. Hence

a cluster may

389

have BS
s

of Pico cell and micro cell

only
if th
ey

have similar c
ontents and

if request
is received by a
390

base station the it is sent to local BS if not fulfilled then it is moved to the base station within
cluster.
.

391

There may be fair chances that contents may be found on the micro BS or Pico Base station
of
the
392

cluster as the cluster contents micro Base stati
ons
, P
ic
o

base stations etc.

These clusters are shown in
393

Figure 2
.

T
he

proposed

algorithm is as

under

394

Algorithm 3
:
M
odified
I
ntra
T
ier and
C
ross
T
ier
C
ooperation

395

Input: X
t

396

Output: Y
r

397

1.

Cache the data using CF in each tier based

on past requests.

398

2.

Cluster all BSs from all the tires(including Pico, Micro and Macro) based on similar cached
399

data files

400

3.

For Pico BS do

401

4.

If

request arrives at Pico BS then

402

5.

Serve the request at associated Pico BS

403

6.

Else if

404

7.

Send the request to neighboring BS

(that may be of Pico, Micro
or Macro ) but within
405

the same
cluster

406

8.

Else if

407

9.

Send the request to another cluster

408

10.

Else

409

11.

Send it to content server

410

12.

End

411

13.

For Micro BS do

412

14.

If

request arrives at Micro BS then

413

15.

Serve the request at associated Micro BS

414

16.

Repeat steps 6
–
12

415

17.

For Macro BS do

416

18.

If

request arrives at Macro BS then

417

Appl. Sci.
20
20
,
10
, x FOR PEER REVIEW

11

of
18

19.

Serve the request at associated Macro BS

418

20.

Repeat steps 6
–
12

419

21.

Repeat for
all requests

420

22.

Return Y

421

422

Figure 2.

Cluster formation considering all tiers
.

423

4
. Simulations & Results

424

In this section, we simulate the cache hit ratio and the user satisfaction ration of the collaborative
425

filtering based multitier caching system
s. The popularity matrix for
tier

1

and the list of parameter
426

used in this simulation is
given in
T
ables I and
Table
II

respectively
. The

requests for the contents on
427

BS on each tier
is different e.g. some people like one contents while others
are interest
ed
in some
428

different contents. Similarly, the number of BS in each tier is
different
.

The size of each content is
429

assumed to be same and is
1
. Each

tier has own data set. The data set compromisers of training
data
430

and testing data
.

First the collaborative
filter is train
by using training data which is
the
content
431

popularity
and then testing data is utilize to predict the missing data in the data set which are the
432

contents request for the data to be requested in future. The popularity matrix of each tier is

highly
433

sparse matrix because users have different modes of
likeness
. Based on the similarities
among the
434

data
,

clusters are
formed

in each tier. Let the cluster
s

in
tier 1
and cluster
s

in tier2 is

2. As we have
435

one macro, so no need to form cluster in
tier 3
.

436

Table I
.

Popularity Matrix
.

437

S
BS

Number
of requests

B1

F1

F2

F3

F4

B2

3

6

6

0

B3

3

0

6

8

B4

0

6

0

9

B5

5

5

5

5

438

Appl. Sci.
20
20
,
10
, x FOR PEER REVIEW

12

of
18

Fig
ure
3 (a), Fig
ure
3 (b) and Fig
ure
3
(c) shows how the cache hit ratios varies as the caching
439

capacity varies in the tier1,
tier2, tier3 r
espectively. In these figures, we show
ed

that the
cluster
440

average popularity with
CF

(CAP
–
CF)

is far b
etter than other method such
as ground truth method
441

and pure CF. In ground truth methods only popul
ar contents are cached, based
on past req
uests but
442

there is no fu
ture prediction of the contents
.

As no clustering is formed in ground truth method
so
443

no similarity between the BSs is observed, which results into poor
utilization of the cache memory
.

444

On the other hand in Pure CF method the conten
ts which are not yet
requested but
hopefully
445

requested in future,
are predicated and are cached in descending based on future popularity
446

predication, without taking into account the pre
sent popularity of the contents
.

The simulation
447

reveals that
our
propos
ed algorithm, which

is based on the highest average popularity of the contents
448

in the cluster and the popularity of the con
tents to be requested in future
,

perform well by optimizing
449

the cache hit ratio
for each tier and also the average cache hit ratio fo
r the
whole multitier caching
450

network
.

The main achievement in the proactive caching network is to cache the frequent requested
451

content in the
limited cache
.

452

Fig
ure
4
(
a
)

and
Fig
ure
4
(
b
)

show the proposed method

1

for t
he user satisfaction ratio, USR
453

based on intra tier and
cross

cooperation for two cluster of tier 1
.

Similarly, Fig
ure
4
(
c
)

and Fig
ure
454

4
(
d
)

shows the same method for tier 2.

This

novel method
is compared with other methods such as
455

CF,
int
ra cluster cooperation
and
cross

cluster cooperati
on
methods. In CF there is

no cooperation

456

among the BSs
for

the data sharing while intra cluster cooperation the data sharing take place

only

457

among
the
BS
s

with in cluster and in
cross

cluster cooperation

the content exchange take place

only
458

among the clusters with in the tier.
O
n the
hand, the

proposed s
c
heme

perform both intra tier and
459

cross

cooperation in order to maximize the USR.

The simulation suggests that the proposed method
460

perform well as compare to all other algorithms due to coope
ration among all the
tier. Figure 5 show

461

2
nd

method

for
MITCTC
.

Simulations

shows that novel method utilizing CF is better

than ground
462

truth

463

464

F
igure

3
(a):

Cache hit ratio versus Cache capacity of Pico Cells

465

Appl. Sci.
20
20
,
10
, x FOR PEER REVIEW

13

of
18

466

F
igure 3
(b):

Cache hit ratio versus Cache
capacity of Micro Cells

467

468

F
igure 3
(c)
:

Cache hit ratio versus Cache capacity of Macro Cells

469

Appl. Sci.
20
20
,
10
, x FOR PEER REVIEW

14

of
18

470

Figure
4
(
a
):
User Satisfaction Rate versus Cache C
apacity in Pico Cell (cluster1)

471

472

Figure 4
(b): User Satisfaction Rate versus Cache Capacity in Pico Cell
(cluster2)

473

Appl. Sci.
20
20
,
10
, x FOR PEER REVIEW

15

of
18

474

Figure 4
(
c
): User Satisfaction Rate versus Cache Capacity in
M
ic
r
o Cell (cluster
1
)

475

476

477

478

Figure 4
(
d
): User Satisfaction Rate versus Cache Capacity in Micro Cell (cluster
2
)

479

Appl. Sci.
20
20
,
10
, x FOR PEER REVIEW

16

of
18

480

Figure
5
: User Satisfaction Rate versus Cache Capacity
in Method 2

481

Table 2.

List of parameters

482

Parameters

Tier1

Tier 2

Tier3

No.

of BS

30

20

1

No. of past request

5563

7636

700

No.

of future request

2000

2300

500

Caching capacity

30

40

50

483

5. Conclusions

484

In this paper, we
presented
popularity predicting caching policy for multitier cellular networks
485

and two novel methods for user satisfaction rate based on intra cluster,
cross

cluster,
cross

tier and
486

intra tier cooperation among the BSs. We have formulated an optimization problem tha
t seek to
487

maximize the average cache hit ratio and the user satisfaction rate. To solve this problem algorithms
488

are developed which use machine learning tools such as unsupervised clustering and collaborative
489

filtering and a heuristic approach for user sat
isfaction rate. Initially, clustered are formed in each tier
490

based on past requests, received by each base station. Similarly, the contents which are not requested
491

now but will be requested in future are also predicated using the collaborative filtering an
d the
492

average cache hit ratio is maximized .For maximization of user satisfaction rate, two methods are
493

suggested. In first method, we have
ITCT
among the base stations. Intra tier cooperation comprises
494

of intra cluster cooperation and
cross

clu
ster cooper
ation in each tier.
. Simulations results have
495

shown that the proposed schemes yields significant performance in terms of average cache hit ratio
496

and user satisfaction ratio compared to conventional approaches.

497

Author Contributions:
Conceptualization
,
F
.
A
.

and
A
.
A
.; methodology,
F
.
A
.; software,
S
.
K
.; validation,
A
.
A.
;
498

formal analysis,
F.A
.; investigation,
A.A
.

and F.A.
; resources,
I.H
.; data curation,
I.H
.; writing
—
original draft
499

preparation,
F.A
.; writing
—
review and editing,
I
.
H
.
; F.A. and A.A.
; visualizat
ion,
I
.
H
.; supervision,
P.U. and
500

I.H
.; project administration,
I.H. and P.U
.; funding acquisition,
I.H
.

and P.U.

All authors
participated equally,
501

have read and agreed to the published version of the manuscript.

502

Funding:
This researc
h is supported by SUT,
Research and Development fund
.

503

Acknowledgments:

The authors are
extremely
thankful to

Govt. of Pakistan and Thai
.

504

Conflicts of
Cross
est:

The authors declare no conflict of
cross
est
.

505

Appl. Sci.
20
20
,
10
, x FOR PEER REVIEW

17

of
18

Abbreviation
:
The following abbreviation are used in this manuscript.

506

CHR:

cache hit ratio

507

USR: user satisfaction ratio

508

BSs: Base stations

509

CAP
–
CF: cluster average popularity based collaborative filtering

510

ITCL: intra tier and Cross tier cooperation

511

CF: collaborative filtering

512

UTs:
user terminals

513

ITCTC:
Intra tier and Cross tier b
ased cooperation

514

ICC:
Intra
–
Cluster cooperation

515

CCC:
Cross
–
Cluster Cooperation

516

CAP
–
CF:
cluster average popularity with CF

517

CTC:
Cross
–
Tier Cooperation

518

MITCTC:
Modif
i
ed Intra tier and Cross tier cooperation

519

References

520

1.

M. Agiwal, A. Roy, and N. Saxena, “Next
generation 5g wireless networks: A comprehensive survey,” IEEE
521

Trans. Inform. Theory, vol. 18, no. 3, pp. 1617
–
1655, Feb. 2016.

522

2.

Optimal joint session admission control in integrated WLAN and CDMA cellular networks with vertical
523

handoff,” IEEE Trans. Mobile

Computing, vol. 6, no. 1, pp. 126
–
139, Jan. 2007.

524

3.

http://www.cisco.com/c/en/us/solutions/collateral/service
–
provider/visualnetworking
–
index
–
525

vni/complete
–
white
–
paper
–
c11
–
481360.html.

526

4.

Cisco Visual Networking Index: ‘‘Global Mobile Data Traffic Forecast Updat
e 2015
–
2020,’’ 2016.

527

5.

L. Breslau, P. Cao, L. Fan, et al. Web caching and Zipf
–
like distributions: evidence and implications. IEEE
528

INFOCOM, 1999, pp. 126
–
134.

529

6.

Z. Han, M. Hong, and D. Wang, Signal Processing and Networking for Big Data Applications. Cambridge

530

University Press, 2017.

531

7.

Z.Chen,J.Xu,J.Tang,K.Kwiat,C.Kamhoua,andC.Wang,“Gpu accelerated high
–
throughput online stream
532

data processing,” IEEE Transactions on Big Data, vol. PP, no. 99, pp. 1
–
1, 2017.

533

8.

E. Dinc, M. Ozger, A. F. Ates, I. Delibalta, and O. B. A
kan, “Crowdsourcing
–
based mobile network
534

tomography for xg wireless systems,” in 2016 IEEE Symposium on Computers and Communication (ISCC),
535

Messina, Italy, June 2016, pp. 346
–
351.

536

9.

M. CenkGursoy, and Senem Velipasalar “A Deep Reinforcement Learning
–
Based Fr
amework for Content
537

CachingChen Zhong,” 2018 52nd Annual Conference on Information Sciences and Systems (CISS)

538

10.

Tianmu Gao
∗
, Mingzhe Chen
∗
, Hanzhou Gu
†
, and Changchuan Yin
”

Reinforcement Learning based
539

Resource Allocation in Cache
–
Enabled Small Cell Networks with Mobile User
”
s 2018 52nd Annual
540

Conference on Information Sciences and Systems (CISS)

541

11.

Yanxiang Jiang†
∗
, Miaoli Ma
†
, Mehdi Bennis‡, Fuchun Zheng†§, and Xiaohu You†
“A Novel Caching
542

Policy with Content Popularity Prediction and User Preference Learning in Fog
–
RAN”IEEE conference
543

2017.

544

12.

Yu Ye
∗
, Zhengquan Zhang
†
, Guang Yang
∗

and Ming Xiao
”

Minimum Cost Based Clustering Scheme for
545

Cooperative Wireless Caching Network with

Heterogeneous File Preference” IEEE ICC 2017
546

Communications QoS, Reliability, and Modeling Symposium

547

13.

Preference Binqiang Chen and Chenyang Yan” Caching Policy Optimization for D2D Communications by
548

Learning User “

549

14.

J. Song, M. Sheng, T. Q. Quek, C. Xu, and

X. Wang, “Learning based content caching and sharing for
550

wireless networks,” IEEE Transactions on Communications, 2017.

551

15.

S. S. Tanzil, W. Hoiles, and V. Krishnamurthy, “Adaptive scheme for caching youtube content in a cellular
552

network: Machine learning app
roach,” IEEE Access, vol. 5, pp. 5870
–
5881, 2017

553

Appl. Sci.
20
20
,
10
, x FOR PEER REVIEW

18

of
18

16.

A. Gruslys, M. G. Azar, M. G. Bellemare, and R. Munos, “The reactor: A sample
–
efficient actor
–
critic
554

architecture,” arXiv preprint arXiv: 1704
–
04651, 2017.

555

17.

T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess,
T. Erez, Y. Tassa, D. Silver, and D. Wierstra, “Continuous
556

control with deep reinforcement learning,” arXiv preprint arXiv: 1509
–
02971, 2015.

557

18.

G. Dulac
–
Arnold, R. Evans, H. van Hasselt, P. Sunehag, T. Lillicrap, J. Hunt, T. Mann, T. Weber, T. Degris,
558

and B.

Coppin, “Deep reinforcement learning in large discrete action spaces,” arXiv preprint
559

arXiv:1512.07679, 2015

560

19.

Kuo Chun Tsai, Student Member, IEEE, Li Wang, Senior Member, IEEE, and Zhu Han, Fellow, IEEE”
561

Caching for Mobile Social Networks with Deep Learnin
g: Twitter Analysis for 2016 U.S. Election” IEEE
562

transactions on network science and engineering, october 2017, tnsesi
–
2017
–
10
–
016

563

20.

Shailendra Rathore, Jung Hyun Ryu, Pradip Kumar Sharma, and Jong Hyuk Park” DeepCachNet: A
564

Proactive Caching Framework Based
on Deep Learning in Cellular Networks” This article has been
565

accepted for inclusion in a future issue of this magazine. Content is final as presented, with the exception
566

of pagination. 2019

567

21.

N. Ben Hassine, R. Milocco, and P. Minet, ‘‘ARMA based popularity
prediction for caching in content
568

delivery networks,’’ in Proc. IEEE Wireless Days, Mar. 2017, pp. 113
–
120.

569

22.

Chen Zhong, M. Cenk Gursoy, and Senem Velipasalar” A Deep Reinforcement Learning
–
Based Framework
570

for Content Caching” 2018 52nd Annual Conference on

Information Sciences and Systems (CISS)

571

23.

Wei Jiang , Gang Feng, IEEE, Shuang Qin, Tak Shing Peter Yum,Guohong Cao” Multi
–
Agent Reinforcement
572

Learning for Efficient Content Caching in Mobile D2D Networks” IEEE transactions on wireless
573

communications, vol. 1
8, no. 3, march 2019

574

24.

Yifei Wei†, Zhiqiang Zhang†, F. Richard Yu‡,Fellow,IEEE, Zhu Han§,Fellow,IEEE” Joint User Scheduling
575

and Content Caching Strategy for Mobile Edge Networks Using Deep Reinforcement Learning”

576

25.

Hao Hao
∗
, Changqiao Xu
∗
, Mu Wang
∗
, Haiyong X
ie
†
, Yifeng Liu
†

and Dapeng Oliver Wu
‡”

Knowledge
–
577

centric Proactive Edge Caching Over Mobile Content Distribution Network
”

2018 IEEE Conference on
578

Computer Communications Workshops (INFOCOM WKSHPS):

KCN 2018: Knowledge Centric
579

Networking.

580

26.

Yan Niu, Shen Gao
, Nan Liu, Zhiwen Pan, Xiaohu You” Clustered Small Base Stations for Cache
–
Enabled
581

Wireless Networks” 2017 9th
Cross
national Conference on Wireless Communications and Signal
582

Processing (WCSP)

583

27.

Gao Shen, Li Pei, Pan Zhiwen, Liu Nan, You Xiaohu” Machine Learn
ing based Small Cell Cache Strategy
584

for Ultra Dense Networks” 2017 9th
Cross
national Conference on Wireless Communications and Signal
585

Processing (WCSP)

586

587

© 20
20

by the authors. Submitted for possible open access publication under the terms
and conditions of the Creative Commons Attribution (CC BY) license
(http://creativecommons.org/licenses/by/4.0/).

588