Received April 4, 2018, accepted May 13, 2018, date of publication May 17, 2018, date of current version

June 29, 2018.

Digital Object Identifier 10.1109/ACCESS.2018.2837692

Big Data Analytics, Machine Learning, and

Artificial Intelligence in Next-Generation
Wireless Networks
MIRZA GOLAM KIBRIA , (Member, IEEE), KIEN NGUYEN , (Senior Member, IEEE),
GABRIEL PORTO VILLARDI , (Senior Member, IEEE), OU ZHAO, KENTARO ISHIZU,
AND FUMIHIDE KOJIMA, (Member, IEEE)
Wireless Systems Laboratory, Wireless Networks Research Center, National Institute of Information and Communications Technology, Yokosuka 239-0847, Japan
Corresponding author: Mirza Golam Kibria ([email protected])

ABSTRACT The next-generation wireless networks are evolving into very complex systems because
of the very diversified service requirements, heterogeneity in applications, devices, and networks. The
network operators need to make the best use of the available resources, for example, power, spectrum,
as well as infrastructures. Traditional networking approaches, i.e., reactive, centrally-managed, one-size-
fits-all approaches, and conventional data analysis tools that have limited capability (space and time) are
not competent anymore and cannot satisfy and serve that future complex networks regarding operation
and optimization cost effectively. A novel paradigm of proactive, self-aware, self-adaptive, and predictive
networking is much needed. The network operators have access to large amounts of data, especially from
the network and the subscribers. Systematic exploitation of the big data dramatically helps in making
the system smart, intelligent, and facilitates efficient as well as cost-effective operation and optimization.
We envision data-driven next-generation wireless networks, where the network operators employ advanced
data analytics, machine learning (ML), and artificial intelligence. We discuss the data sources and strong
drivers for the adoption of the data analytics, and the role of ML, artificial intelligence in making the system
intelligent regarding being self-aware, self-adaptive, proactive and prescriptive. A set of network design
and optimization schemes are presented concerning data analytics. This paper concludes with a discussion
of challenges and the benefits of adopting big data analytics, ML, and artificial intelligence in the next-
generation communication systems.

INDEX TERMS Big data analytics, machine learning, artificial intelligence, next-generation wireless.

I. INTRODUCTION of this, the network operators face considerable challenges

In a service-driven next-generation network, a single infras- in extending their coverages and keeping up with the ever-
tructure needs to efficiently and flexibly provide diversified increasing capacity demands with a limited pool of capital
services such as enhanced mobile broadband, ultra-reliable and scarcity of resources such as spectrum. Manual config-
and low-latency communications and massive machine type uration for network planning, control, and optimization will
communications. It should also support coexistent accesses make things even more complicated. Moreover, the human-
of multiple standards such the fifth generation (5G), long- machine interaction can, sometimes, be time-consuming, sus-
term evolution (LTE) and Wi-Fi. Also, it should coordinate ceptible to error and expensive. Consequently, automation of
a heterogeneous network with different types of base sta- various entities and functions of the cellular networks has
tions (BSs), for example, macro, micro, femto, pico BSs and been one of the principal concerns of the network operators
diverse user devices as well as applications [1]. The chal- in consideration of reducing the operational expenses.
lenge to efficiently operate a network capable of facilitating Operators have been optimizing their networks all along,
such flexibility while satisfying the demands from diversi- but even today, the prevailing approach is to independently
fied services is enormous for a network operator. On top optimize single key performance indicators (KPIs), or an

2169-3536
2018 IEEE. Translations and content mining are permitted for academic research only.
32328 Personal use is also permitted, but republication/redistribution requires IEEE permission. VOLUME 6, 2018
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
M. G. Kibria et al.: Big Data Analytics, ML, and Artificial Intelligence in Next-Generation Wireless Networks

FIGURE 1. Graphical illustration of the next-generation communication system with some technological elements. Slicing, virtualization,
(mobile) edge-computing, massive multiple-input-multiple-output (MIMO), 3-dimensional (3-D) beamforming, (ultra-dense) small cell
networks, device-to-device(D2D) cell-free massive MIMO, multi-connectivity, cloud-radio access network (RAN), millimetre-wave,
cloud-architecture/computing, etc., are the fundamental technologies to achieve the targets of the next-generation network.

element within the network independently [2], thus using a slice and traffic steering and splitting (between 3GPP and
small number of data sources. The network operators mostly non-3GPP access) related analytics automatically [4]. The
depend on KPIs accumulated at different locations/parts of European telecommunications standards institute (ETSI) has
the network to make decisions employing various data anal- created the industry specification group called experimental
ysis tools. Network monitoring and optimization are still network intelligence (ENI) that defines a cognitive network
predominantly performed on old/recorded data, but this management architecture based on artificial intelligence (AI)
dramatically restricts their capacity. The network opera- techniques and context-aware policies. The ENI model helps
tors, in general, have/can have access to a vast amount the network operators in automating the network configura-
of data from their networks and subscribers. With the tion and monitoring process.
appropriate analytics, big data can convey broader intu- From the operational expenses point of view, the system
itiveness and understanding since it draws from multiple needs to be smart, self-aware, self-adaptive and must be
sources to reveal previously unknown patterns and corre- able to run the network services economically and manage
lations [3]. It benefits to acquire a thorough knowledge and operate the networks autonomously [5]. Conventional
of various unknown values and delivers new measures in reactive maintenance is no more efficient. With big data
enhancing the performance from different levels of wireless analytics, the predictive and proactive maintenance of the
networks. network elements can be performed. With the volume of the
The value that analytics brings to optimization comes from data, the speed of data flowing in and the range and type of
expanding the range of data sources and taking a customer- data sources, the network even go beyond prediction, i.e., it
centric, quality of experience (QoE)-based approach to opti- can assist and or prescribe the operations and maintenance
mizing end-to-end network performance. In widening the unit with decision options and impacts of the actions, etc.
variety of data sources, analytics requires more effort than Machine learning (ML) and AI can help in uncovering the
traditional optimization, but it also provides a unified and unknown properties of wireless networks, identify correla-
converged platform for multiple targets of optimization. Now, tions and anomalies that we cannot see by inspection, and
within the 3rd generation partnership project (3GPP), net- suggest novel ways to optimize network deployments and
work data analytics (NWDA) has been introduced to deliver operations.

VOLUME 6, 2018 32329

 M. G. Kibria et al.: Big Data Analytics, ML, and Artificial Intelligence in Next-Generation Wireless Networks

II. DRIVERS AND EVOLUTION OF ANALYTICS IN

NEXT-GENERATION WIRELESS SYSTEMS AND
COMPUTATIONAL INTELLIGENCE
A. DRIVERS TO ADOPTION OF ANALYTICS
The ever-increasing complexity of the networks and compli-
cated traffic patterns make the big data analytics appealing
and very important for the network operators. The network
operators were earlier very cautious about the adoption of
big data analytics. However, multiple drivers are turning the
network operators conservative stance towards the compre-
hension that in-depth optimization of the networks and the
services are essential for near future. As a consequence, there
exists a consistent and rational commitment to capturing
a thorough knowledge and understanding of the network
dynamics and make the best use of them through optimiza-
tion. Three predominant drivers strengthening the adoption
of big data analytics [2] can be identified as Cost and
Service drivers, Usage drivers and Technology drivers. In the
following, we discuss them in more details.

FIGURE 2. Data sets and the sources of data available to the network
1) COST AND SERVICE DRIVERS operators for big data analytics, machine learning and artificial
The subscribers, in general, are more demanding but less intelligence.
eager to raise the wireless payout. In such environment, there
is an urgent need for optimization of the usage of network
resources. Furthermore, the network-centric service model is need robust analytics framework to orchestrate the virtualized
transforming into a user-centric service model based on the network resources efficiently. The analytics also help the
QoE. As a result, the network operators need to better under- network operators to balance the centralized and distributed
stand the QoE and its relationship with the network’s KPIs. functionality. The data analytics facilitates the network opera-
Also, the network operators need to retain its customers. As a tors to figure out the most competent way to slice the network
result, the network operators need to (i) manage its traffic and traffic, i.e., the number of slices, splitting traffic across
based on service and application, (ii) improve efficiency to slices, etc., which depend on the type of traffic and how varies
retain profit margins, (iii) improve network performance and over time and space.
QoE without increasing cost and (iv) keep churn as low as
possible, etc.
B. TYPES OF ANALYTICS
There exists a succession of evolution in big data analytics,
2) USAGE DRIVERS
starting from descriptive analytics to diagnostic analytics to
The traffic patterns, subscriber equipment, and subscribers’ predictive analytics, and excelling towards prescriptive ana-
profiles are all heterogeneous. In a user-oriented service lytics as shown in Fig. 3, out of which three (descriptive, pre-
model, analytics supports the network operators maintain dictive and prescriptive analytics) are dominant. The network
and regulate traffic types, wireless devices, and subscribers operators currently are in descriptive phase and use mainly
diversely based on the network operators’ strategies and the visualization tools to get insights on what has happened,
each’s requirements. Furthermore, the wireless traffic load is the network performance, traffic profile, etc. The network
growing faster than the capacity, and the network operators operators can make use of the diagnostic analytics to fig-
are facing severe challenges to increase network capacity ure out the root-causes of the network anomalies and find out
cost-effectively. Therefore, intensifying the resource utiliza- the faulty KPIs and network functions/elements. To get the
tion is required. Analytics take the network load into account diagnostic analytics, the analytics tool employs techniques
and helps the network operators to manage network traffics like drill-down, deep learning, data discovery, correlations,
more efficiently in real time. etc.
Predictive analytics is an excellent tool for making pre-
3) TECHNOLOGY DRIVERS dictions. Note that it can never report or be precise about
The next-generation wireless networks have many technol- what will happen, however, predictive analytics can only
ogy components such as network resource virtualization, produce forecasting about what might occur, for example,
edge-computing, mobile edge-computing, network-slicing, future locations of the subscribers, future traffic pattern and
etc. It integrates multiple air-interfaces, network layers and network congestion, etc. Predictive analytics deliver predic-
accommodates a range of use-cases. The network operators tions events based on the real-time and archived data by

32330 VOLUME 6, 2018

M. G. Kibria et al.: Big Data Analytics, ML, and Artificial Intelligence in Next-Generation Wireless Networks

FIGURE 3. Drivers from different directions that strengthen the case for the adoption of analytics in next-generation communication systems. The
flow of processes/actions of the network operator that employs analytics. Here, the Analytics Function collects the data and analyze them. The
Policy Function obtains the analytics reports produced by the Analytics Function and may dynamically and intelligently deliver analytics-based
policy rules for UE and the network functions.

making use of various statistical techniques such as machine subscriber related. The external data is collected from the
learning, data mining, modeling as some analytical process third parties. Both the internal and external data can be further
and game-theoretic analysis. Prescriptive analytics goes steps classified into two categories, which are structured data and
ahead of just predicting the future events by suggesting deci- unstructured data. The structured is stored in a relational
sion options for slicing (i.e., how to slice, how many slices), database, i.e., each field in the database has a name, and
virtualization, edge-computing, etc., along with the impli- the relationship between the areas are well-defined. On the
cations of each decision option. Therefore, the prescriptive other hand, the unstructured data (for example, call center
analytics need a useful predictive model, actionable data and transcripts, messages, etc.) is not usually saved in a rela-
a feedback system for tracking down the results generated tional database. Comprehensive coverage of the features and
by the action taken. The decision options (e.g., for network sources of mobile big data can be found in [6] and [7].
expansion, resource usage) are produced considering the net-
work operators preferences, system constraints (backhaul, C. COMPUTATIONAL INTELLIGENCE
fronthaul, spectrum, transmission power), etc. Prescriptive network operators have access to a collection of data sets
analytics can also suggest the finest course of actions for any (i.e., these data can be highly dimensional, heterogeneous,
pre-defined target, for example, of a particular KPI. complex, unstructured and unpredictable) that are so large
The network operators have access to large amounts of data and complex that the traditional data processing and analysis
which can be categorized into two classes such as internal approaches cannot be employed due to their limited process-
data and external data as shown in Fig. 2. The internal data ing space and processing time. Computational intelligence,
corresponded to data belonging to the network operators a set of nature-influenced computational techniques and
and produced in the network, which is network related and methods, play a very crucial role in the big data analysis [8].

VOLUME 6, 2018 32331

 M. G. Kibria et al.: Big Data Analytics, ML, and Artificial Intelligence in Next-Generation Wireless Networks

It enables the analytics agent to process and analyze the AI goes beyond predictions and prescribe plans/suggestions
historical and real-time data computationally, and eventually with implications to realize a benefit.
finds out and explain the underlying patterns, correlations, Managing wireless networks that grow in size and com-
as well as to intensely understand the specific tasks. The plexity becomes very difficult since there is need to integrate
computational analysis tools and methodologies convert the new elements and technologies to benefit from the techno-
network operators’ massive amount of raw data (unprocessed, logical advances. The amount of data such large and complex
structured/unstructured) into meaningful data/information. networks produces is too large and too complicated. Machine
For feature selection, data-size and feature space com- learning and artificial intelligence are useful for analytics as
pliance, active-incremental-manifold-imbalance learning on they can extract valuable information from the raw data and
big data, uncertainty modeling, sample selection, classifi- generate insightful advice and predictions. ML and AI are
cation/clustering, etc., many tools and methodologies can expected to assume the primary role in the development and
be applied for big data analysis. For example, fuzzy logic, evolution of analytics, but analytics will not reduce to them.
neural algorithms, rough sets, swarm intelligence, evolution- ML is developed mainly from AI, hence the two overlap.
ary computing, stochastic algorithms, physical algorithms, ML has tools to extract relevant information, suggestions,
immune algorithms, learning theory, probabilistic methods and predictions from the data sets that are too large and too
are the tool and methodologies that the network operators’ complex. AI has a broader scope: to replicate human intel-
big data analytics agents can employ for computationally ligence or some aspects of it and other cognitive functions.
processing and analyzing the available data. The differentiation among big data analytics, ML, and AI,
In general, for big data analytics, the network opera- and their mutual dependence relationships are discussed in
tor can follow two distinct approaches, namely top-down details in [2].
approach and bottom-up approach [9]. In the top-down Furthermore, for non-recurring events, there is no histori-
approach, the network operators define their targets to cal data to rely on. Hence the real behavior of the network
be achieved or problems to be resolved, and then decide will diverge from the predictions [2]. The ML and AI are
what data sets are required. Whereas, in the bottom-up becoming potential to help network operators to address areas
approach, the network operators already have access to mas- which are new, and there is no historical data, or too compli-
sive amounts of data and then exploit the big data on hand to cated to understand with traditional approaches. The ML and
get the insights. The top-down approach delivers incremental AI tools can correlate multiple sources of data and find what
benefits and it is very challenging to execute. It also, in most is relevant. They may also reveal interrelations and depen-
of the cases, does not bring on surprising and adventitious dencies that were not previously identified because their
results. On the other hand, the bottom-up approach facilitates automated mechanisms have the capability of anatomizing
a more outright and transparent view of the network perfor- and inspecting data more intensely and more methodically.
mance, subscribers’ behaviors, resource utilization, etc., and Although human expertise is useful in confining the focus to
may bring on completely new opportunities for the network produce solutions and to manage complex problems, it has
operators. The bottom-up approach is also likely to capture limited capability in finding new answers and insights. The
the subscribers’ perspectives the RAN perspectives and may future of wireless networks will undoubtedly rely on AI.
beget new business opportunities for the network operators. In [11], the authors have provided a panned overview of
the range of wireless communication problems and issues
III. MACHINE LEARNING AND ARTIFICIAL INTELLIGENCE that can be efficiently addressed using AI while delivering
FOR MANAGING COMPLEXITY detailed examples for the use-case scenarios.
The ML and AI are two compelling tools that are emerging
as solutions for managing large amounts of data, especially IV. DATA-DRIVEN COVERAGE AND CAPACITY
for making predictions and providing suggestions based on OPTIMIZATION OF NEXT-GENERATION CELLULAR
the data sets. They are, however, very often appear to be used WIRELESS NETWORKS
interchangeably in spite of some parallels. ML is sometimes The conventional network-centric architecture cannot capture
brought up as a subspace of AI based on the concept that all of the nuances that can affect service quality. Mobile
we can let the machines learn for themselves by providing operators need solutions that provide them with an analysis
them access to large amounts of data. On the other hand, AI is capability that captures all the information relating to the
the widespread and broader perception of devices becoming network and subscribers into a single enterprise geolocation
capable of carrying out tasks in an intelligent way. Compared platform that can help remove the assumptions involved in
to the generalized AI (a generalized AI system, in theory, fault isolation and reduce mean time to repair. The network
can handle any task), applied AI is more suitable for next- operators are suitably positioned to exploit big data analytics
generation communication systems as the applied AI sys- because of their access to huge amounts of data. The big
tem can be devised to adeptly controlling and optimizing data analytics engine/agent can produce/predict the following
the wireless networks. Unlike ML models, AI models reach analytics based on its data, primarily from two sources, such
out the world, accustomed to the changes and rebuild them- as the network data and the subscriber data, which are then
selves [2]. While ML is excellent for predictive analytics [10], exploited to design and optimize the network.

32332 VOLUME 6, 2018

M. G. Kibria et al.: Big Data Analytics, ML, and Artificial Intelligence in Next-Generation Wireless Networks

FIGURE 4. Left: Example of the trajectory of a mobile user who visited the vicinity of during an observational period. The circles correspond to the
serving BSs, i.e., give the approximate locations. The gray lines depict the Voronoi lattice. The width of the line between two BSs is proportional to how
often the user moves between the BSs, and the sizes of the circles are proportional to the resources/traffic consumed. Middle: Example of the
small-scale (e.g., within few seconds/minutes) trajectories of the user for more accurate/exact location of the user. Right: The circles represent the BSs
and sizes of the rings are proportional to the traffic loads on the BSs, i.e., the traffic pattern and congestion status in the network can be measured.
Note that results here do not come from real field experience.

• Subscriber Profile (SP): In this context, the subscriber metrics, the network operator can generate heat map
profile consists of the device profile, service-level agree- for coverage and determine the RAN quality. Advanced
ment (SLA), subscriber’s affordability (price per unit of cell mining that statistically analyzes the performance
data-rate), quality of service (QoS)/policy, behavioral data enables the network operator to identify radio cell
profile, etc. It plays a vital role in the abovementioned irregularities and other negative syndromes via anomaly
controlling and optimization process. The priority of the (i.e., SLA violation) diagnosis and trend study of the
subscriber in the network is defined in the subscriber time series data, and control traffic and RAN conges-
profile when resource allocation, congestion control and tion problems. With RAN Perspective, the full end-to-
traffic offloading is performed. Behavioral data provides end subscriber experience can be measured regarding
information how the user behaves in using various appli- Service Availability and correspondingly mapped to the
cations/services. For example, how frequently and when exact location in the network. The network operator can
the user makes video/audio call and the average length also use Subscriber Satisfaction Coefficient to define the
of the call duration? Through analytics, we can speculate RAN perspective. Note that the signaling metric cannot
on a lot of these user attributes. be easily retrieved from mobile gateways or retrieved by
• Subscriber Perspective (Sub-P): Subscriber perspec- network probes. An efficient retrieving method is dis-
tive is an attribute/measure that associates network oper- cussed in [13] that uses SIM-based applet stored in users
ator’s network activity with the user’s SLA, pricing, devices to collect the signal strength and quality metrics.
QoS, QoE, etc., and it delivers a subscriber-centric out- Thus the subscriber devices act as network probes in
look of the network for analytics [12]. Subscriber per- measuring the RAN perspective.
spective is, in general, defined by the Cost Over Quality • Subscriber Mobility Pattern (SMP): To guarantee the
Ratio, which sometimes gets polished through a variety QoS requirements and to efficiently maintain resources
of attributes linked to the requested service class and utilization, traffic offloading and routing, knowing the
perceived friendliness to the service, i.e., QoS violation, mobility information of a user in advance is very crucial.
delay violation, etc. It enables the network operator to Human travel pattern analyses reveal that people travel
measure or make a perception about the RAN quality along specific paths with reasonably high predictabil-
from the subscriber point of view and put them in a better ity [14]. The trajectory of a mobile user can be pre-
position to provide a high QoE. dicted based on user’s present location, the movement
• RAN Perspective (RAN-P): RAN perspective is a mea- direction and the aggregate history of SMP. It is possi-
sure that provides the network operator the subscriber- ble to predict the spatiotemporal trajectory (trajectory
centric RAN quality, i.e., the RAN performance from the with both spatial and temporal information), i.e., not
subscriber’s point of view [13]. The user equipment’s only the mobile user’s future location but also the time
view of the signaling information such as signal strength, of arrival and the duration of stay can be predicted.
error codes, available networks, etc., are beneficial to the Mobility pattern is based on user positioning, which
network operator for analytics. From the user’s predicted can be estimated using the signals from the cellular
trajectory, spatial deployment of the BSs and signaling system.

VOLUME 6, 2018 32333

 M. G. Kibria et al.: Big Data Analytics, ML, and Artificial Intelligence in Next-Generation Wireless Networks

• Resource Allocation Strategy Advanced resource allo-

cation is very crucial for enhancing the spectrum and
power utilization efficiency of the communication sys-
tems. Leveraging the big data analytics based prediction
ability in optimizing the resource allocation has been
reported to be very advantageous. With the help of
SMP and TP, the network operator can approximate the
typical resource usage per-cell per-user in the network.
Because (i) the average channel gains are predictable
from the trajectory of the user and (ii) based on content
popularity, user’s behavioral profile and currently run-
ning application, the preferred contents can be predicted
even before the individual users put forward their service
requests. As a result, with the help of big data predictive
analytics, the operators can predict changes in the users’
service demands and thus can manage and optimize the
resource allocation in real time. Integrated backhaul and
access in mmWave.
FIGURE 5. Some data analytics and their applications in control and
• Subscriber-Centric Traffic Routing Providing the best
optimization of next-generation wireless communication systems. QoE as the end users’ subjective perception is one of
the most important requirements. Service delay, jitter
affect the mobile users’ QoEs very severely. Data-driven
solutions can deliver traffic to different users depending
• Radio Environment Map (REM): The network oper-
on their subscription profile, types of applications, and
ators can better plan, build, control and optimize their
preferences. A QoE-aware network continuously adopts
networks conforming to the spatiotemporal radio atmo-
the changing environment to provide acceptable QoE.
sphere, through prediction of radio signal attenuation.
The SMP, the network utilization profile, and TP can
Many schemes have been developed that give the net-
help the operator to devise efficient routing protocol
work operators the means to predict the distribution of
while considering the backhaul load, the SLA, and the
radio signal attenuation at different operating frequen-
corresponding cost. Depending on users preferences and
cies and in many different radio environments. The radio
interests, and currently running application, the system
map along with the mobile user’s predicted trajectory
can proactively cache the favorite content, and use the
facilitates the prediction of average channel gains. There
backhaul route that is closer to the local caching server.
are several different methods to construct the radio map,
• Subscriber-Centric Wireless Offload Due to an expo-
for example, radio map based on drive test measure-
nential surge in mobile data traffic carrier over macro-
ments, radio map based on measurements through user
cell layer, the network operators are more and more
terminal equipped with global positioning system [15].
finding out approaches to optimize the traffic in the
• Traffic Profile (TP): To attain as well as predict the
network while ensuring seamless connectivity and mini-
network’s congestion status, tempo-spatial traffic load
mum guaranteed QoS to its subscribers. Traffic offload-
variation needs to be known, i.e., the knowledge of
ing from macrocell layer to small cell layer (specifically
temporal traffic trace, BS spatial deployment and BSs’
towards WiFi networks) is a great way to relieve conges-
operating characteristics (transmission power, height,
tion in the macro layer and enhance the overall network
etc.) are critical. The authors in [16] report that the
throughput. Blindly offloading the mobile users may
network’s traffic load dynamics demonstrates periodical
result in dissatisfaction of the subscribers of the higher
characteristics over days and hours, thus implying high
tier and breaching SLAs. Therefore, it is necessary to
predictability of the traffic load. The traffic profile along
devise practical solutions that aid the network operators
with the SMP can be used to estimate and predict the
to decide and offload mobile users to WiFi, based on user
traffic arrival rate, congestion status of the network with
profile and network congestion conditions. Data-driven
required time resolution/granularity.
contextual intelligence originated from correlating the
It is very crucial to have a sturdy, well-balanced load- customer profile (types of application, spending pattern,
distributed cellular system in a dynamic network and radio SLA) with SMP, TP, and REM, can decide which cus-
environment with mobile users using bursty applications and tomer should be offloaded, and even to which small
services. The next-generation network can employ the sys- cell/WiFi the customer needs to be unloaded.
tems analytics, user, and service analytics, radio analytics for • Optimized Cell Placement Small cell placement plays
control and optimization of the network [17] in the following a vital role in defining the capacity of a heterogeneous
scenarios. network. Strategic small cell placement is crucial in

32334 VOLUME 6, 2018

M. G. Kibria et al.: Big Data Analytics, ML, and Artificial Intelligence in Next-Generation Wireless Networks

areas where subscribers concentrate while taking care of such TP and SMP can not only help in understanding the
coverage goals, radio frequency interference issues and cells’ current load situation, but also in identifying the
its potential in relieving congestion from the macro layer heavily loaded parts of the network and predicting the
by offloading traffic. Rapidly placing the small cells traffic variation in advance. Consequently, the network
at the very best locations is a complex problem as the operators can perform advanced load balancing and cell
number of small cells is much larger. Traditional macro planning by adding capacity, expanding the coverage of
cell layer management tool and even the self-organized unloaded or lightly loaded cells to unburden the neigh-
networking tool may not compensate for improper cell boring overburdened/overloaded cell. The data-driven
placement. However, the data-driven solution can effi- advanced load balancing will enable the network oper-
ciently administer the small cell placement issue exploit- ators to optimize the utilization of available network
ing the knowledge of long-term user density, traffic resources.
intensity. Data-driven solutions incorporating the long- • Advanced Beamforming Beamforming is an integral
term TP, REM can devise optimized dynamic small technology component in next-generation communica-
cell placement strategy that identifies key locations tion systems for enhancing the coverage and data rates.
where small cells need to be deployed and re-arranged A BS with multiple antennas can generate many beams
to enhance the network capacity, minimize interfer- simultaneously [18]. Under static beamforming (fixed
ence and improve the traffic offloading capability. The beam pattern without beam-steering), for a mobile user,
3D geolocator tool that uses predictive ‘‘fingerprinting’’ the quality of the serving beam may deteriorate, and
algorithms to locate traffic hotspots can simplify the cell hence a different beam from the same BS (from same
placement task. sector or different sector) or an adjacent BS that serves
• Radio Access Network Congestion Control The com- the user well needs to be selected. The ML can help the
bination of limited network resources and ever-growing serving node to choose the best beam for the user dynam-
demands result in unavoidable RAN congestion, which ically. The ML also enables dynamic switching ON/OFF
degrades users’ quality of experience. Expansion of the beams based on TP and SMP for energy and inter-
existing RAN provides a solution to this problem, but ference minimization. Holographic beamforming1 (with
it is expensive. A flexible, as well as a cost-effective electronic speed beam-switching/beam-steering) [19]
solution, is to deploy a proactive policy control mech- along with data analytics and ML can help in dynami-
anism preventing deficiency of RAN resources. Smart cally rerouting the traffic, dynamic adjacent cell access,
congestion control solution considering location infor- steering coverage where it is needed to accommodate
mation, the load level of network elements and users’ usage patterns, for example, rush hour traffic, events,
service level agreements can deliver perceptibility at a etc., as shown in Fig. 6.
particular sub-cell level and caters priority to some set Fig. 6(a) contemplates adaptive/dynamic rerouting of
of subscribers based on their tiers. The congestion events traffic when there is an obstacle or physical interference
are short-lived (typically congestion occurs at busy times between two communicating nodes. The rerouting path
of the day) and users future locations are predictable. can be dynamically selected based on TP and other infor-
With the help of data-driven predictive analytics incor- mation such as resource availability. Dynamic adjacent
porating the correlation between SMP, radio map and cell assistant in Fig. 6(b) with holographic beamforming
traffic profile, advanced proactive RAN congestion con- facilitates serving a distant user outside the general cov-
trol mechanism can be deployed where the occurrence of erage area of the assisting BS when the original serving
RAN congestion is predicted. RAN congestion control- BS has bandwidth shortage/overloaded or becomes non-
ling can be done in many ways, for example, by reducing functional. Depending on the received information and
the QoS for subscribers belonging to the lowest tier of making use of TP and REM, the holographic beamform-
users, rejecting new session establishment, terminating ing antenna can dynamically configure a high direc-
specific sessions. tivity beam towards the distant user [20]. Similarly,
• Advanced Load Balancing Note that the profile of dynamically steering coverages where it is required as
mobile users and traffic in each cell is distinctive, and shown in Fig. 6(c) combined with 3D geolocator tool
the patterns change from time to time. When some users and configuring long-range, high-capacity links along
disassociate with one cell and move to the neighboring with electronic speed beam-switching to provide access
cell, the network’s traffic load distribution, i.e., the traf- in motion as shown in Fig. 6(d) can be enhanced by
fic profile may change severely, and as a consequence, analytics from internal data, external data and ML.
some cells in the network may get overloaded caus- Apart from the control and optimization scenarios men-
ing service downgrade. Currently, the load balancing tioned above, accurately and efficiently accomplishing the
methods employed by the network operators are almost
1 In holographic beamforming, the complex propagating wave across sur-
manual, thus not efficient, and at the same time, they
face scattering antenna or the transmitting aperture becomes a holographic
are not accurate enough. Predictive analytics by data profile, i.e., the collective profile across the antenna-array elements repre-
mining and correlating the network and subscriber data sents the desired hologram for transmission.

VOLUME 6, 2018 32335

 M. G. Kibria et al.: Big Data Analytics, ML, and Artificial Intelligence in Next-Generation Wireless Networks

FIGURE 6. Data analytics and ML, and AI techniques can be used in analog, digital, and hybrid beamforming in terms of generating the optimal
beam patterns, dynamically selecting the most suitable beam, and performing beam-steering operation. In this figure, holographic beamforming
has been taken as the use-case scenario.

maintenance of the network elements, backhaul monitor- V. CHALLENGES AND BENEFITS

ing (potential bottlenecks in backhaul networks). Fronthaul Although employing big data analytics for control and opti-
management and orchestration, intelligent network slicing, mization of wireless networks is very attracting to the net-
energy optimization, monitoring of critical network health work operators, it comes with some challenges. The process
variables are some of the key pain and troublesome issues of managing and leveraging of a massive amount of data,
the network operators are very often challenged with. With designing algorithms for dynamic and efficient processing
big data analytics, the predictive maintenance of the net- of sizable data sets and then exploiting the insights from
work elements can be performed. The predictive mainte- the data analytics in networks can pose unique challenges.
nance inspects the operational status of the network elements The prime concerns for the network operators emerge from
through sensors in real-time. With the help of big data ana- the extent of effort, skills, and workforce needed to manage
lytics, the potential risks can be identified. Thus the possible and operate a big data platform. However, the most critical
faults are found earlier. This helps the operation and mainte- and challenging task is more likely to stem from the loss
nance team of the network operator to become proactive to of direct control that the network operators still have over
work out predictive maintenance planning. the wireless network. The loss of direct control is incurred
Furthermore, there are plenty of other applications of data from the combination of automation and real-time operations
analytics, ML, and AI in next-generation communication within the big data analytics framework. However, the consid-
systems. For example, the network operators can employ erable complexity of the next-generation networks makes the
analytics for obtaining useful insights about the physical automation inevitable, and handover or relinquish that level
layer [21] and the medium access control (MAC) layer. Also, of direct control is imperative. On top of these, a substantial
the optimal constellations in interference channels where the investment is necessary.
optimal schemes are unknown, the best beamformer, pre- Despite all the challenges, the network operators are more
caching/buffering, the most suitable forward error correction considerate towards data analytics platform since the chal-
code, the optimum MAC protocols, predictive scheduling, lenges are outweighed by the benefits. The big data analytics
etc., can be performed. Intelligent wireless network archi- infuses efficiency into the provisioning of services and end-
tecture, RAN optimization regarding transmission control to-end network. Analytics facilitates the network operators to
protocol (TCP) window optimization, mobility management gain from the better planning, increased utilization of network
optimization can also be achieved through the use of big data resources, efficient maintenance of the network elements
analytics [22]. Data-aided transmission, network optimiza- and lower operating costs. It gives the network operators
tion, for example, channel modeling, multiple user access the flexibility to define and execute their network utilization
and novel applications such as unmanned aerial vehicle/drone strategy. It helps the operators to make new service and offer
communications, smart grid, etc., have been discussed plans that are suited to subscribers’ needs. Although the net-
in [23]. work operators are already performing these kinds of service

32336 VOLUME 6, 2018

M. G. Kibria et al.: Big Data Analytics, ML, and Artificial Intelligence in Next-Generation Wireless Networks

provisioning, analytics delivers richer insights. Analytics [11] M. Chen, U. Challita, W. Saad, C. Yin, and M. Debbah. (2018). ‘‘Machine
helps to improve subscriber management and policy imple- learning for wireless networks with artificial intelligence: A tutorial on
neural networks.’’ [Online]. Available: https://arxiv.org/abs/1710.02913
mentation. With the aid of natural language processing [12] S. A. Kyriazakos and G. T. Karetsos, Practical Radio Resource Manage-
and interfacing with the smart digital assistants in the ment in Wireless Systems. Boston, MA, USA: Artech House, 2004.
user devices, autonomous customer care can be facilitated. [13] Procera Networks. RAN Perspectives: RAN Analytics & Enforcement.
Accessed: Oct. 13, 2017. [Online]. Available: https://www.
Another benefit the network operators get by employing ana- proceranetworks.com/hubfs/Resource%20Downloads/Datasheets/
lytics is differentiation, which is compelling to strengthen- Procera_DS_RAN_Perspectives.pdf?t=1481193315415
ing an network operator’s market positioning. Analytics can [14] X. Lu, E. Wetter, N. Bharti, A. J. Tatem, and L. Bengtsson, ‘‘Approaching
the limit of predictability in human mobility,’’ Sci. Rep., vol. 3, p. 324,
support the network operators to employ new techniques to Nov. 2013.
traffic handling such as network-slicing (i.e., the way to slice [15] R. Atawia, H. S. Hassanein, and A. Noureldin, ‘‘Fair robust predictive
the network) and edge-computing (i.e., the way to balance resource allocation for video streaming under rate uncertainties,’’ in Proc.
IEEE Globecom, Dec. 2016, pp. 1–6.
centralized and distributed functionality). [16] E. Oh, B. Krishnamachari, X. Liu, and Z. Niu, ‘‘Toward dynamic energy-
efficient operation of cellular network infrastructure,’’ IEEE Commun.
VI. CONCLUSION Mag., vol. 49, no. 6, pp. 56–61, Jun. 2011.
We consider a data-driven next-generation wireless network [17] A. Banerjee, ‘‘Advanced predictive network analytics: Optimize your net-
work investments & transform customer experience,’’ Heavy Reading,
model, where the network operators employs advanced data New York, NY, USA, White Paper, Feb. 2014.
analytics, ML and AI for efficient operation, control, and [18] W. Hong et al., ‘‘Multibeam Antenna Technologies for 5G Wireless
optimization. We present the main drivers of big data ana- Communications,’’ IEEE Trans. Antennas Propag., vol. 65, no. 12,
pp. 6231–6249, Dec. 2017.
lytics adoption and discuss how ML, AI and computational [19] E. J. Black et al., ‘‘Broadband surface scattering antennas,’’ U.S. Patent
intelligence play their important roles in data analytics for 0 187 123 A1, Jun. 2017.
next-generation wireless networks. We present a set of net- [20] E. J. Black et al., ‘‘Methods and systems for communication with beam-
forming antennas,’’ U.S. Patent 0 373 181 A1, Dec. 2016.
work design and optimization schemes with respect to data [21] T. O’Shea, and J. Hoydis. (2017). ‘‘An introduction to deep learning for
analytics. Finally, we discuss the benefits and challenges the physical layer.’’ [Online]. Available: https://arxiv.org/abs/1702.00832
that the network operators encounter in adopting big data [22] C.-l. I, Q. Sun, Z. Liu, S. Zhang, and S. Han, ‘‘The big-data-driven
intelligent wireless network: Architecture, use cases, solutions, and future
analytics, ML, and AI in next-generation wireless networks. trends,’’ IEEE Veh. Technol. Mag., vol. 12, no. 4, pp. 20–29, Dec. 2017.
[23] Q. Lijun, Z. Jinkang, and S. Zhang, ‘‘Survey of wireless big data,’’
ACKNOWLEDGEMENT J. Commun. Inf. Netw., vol. 2, no. 1, pp. 1–18, 2017.
This research was conducted under a contract of R&D for
Expansion of Radio Wave Resources, organized by the Min-
istry of Internal Affairs and Communications, Japan.
The first author would like to thank Monica Paolini from MIRZA GOLAM KIBRIA (S’11–M’14) received
Senza Fili Consulting for the insightful comments. the B.E. degree from Visveswaraiah Technological
University, India, in 2005, the M.Sc. degree from
REFERENCES the Lund Institute of Technology, Lund Univer-
[1] F. Boccardi, R. W. Heath, Jr., A. Lozano, T. L. Marzetta, and P. Popovski, sity, Sweden, in 2010, and the Ph.D. degree from
‘‘Five disruptive technology directions for 5G,’’ IEEE Commun. Mag., Kyoto University, Japan, in 2014, all in electrical
vol. 52, no. 2, pp. 74–80, Feb. 2014. engineering. In 2014, he joined the National Insti-
[2] M. Paolini. Mastering Analytics: How to Benefit From Big Data and tute of Information and Communications Technol-
Network Complexity. Accessed: Nov. 2, 2017. [Online]. Available: ogy, where he is currently a Researcher with the
http://content.rcrwireless.com/20170620_Mastering_Analytics_Report Wireless Systems Laboratory, Wireless Network
[3] S. Bi, R. Zhang, Z. Ding, and S. Cui, ‘‘Wireless communications in the Research Center, Yokosuka Research Park, Japan. His research interests
era of big data,’’ IEEE Commun. Mag., vol. 53, no. 10, pp. 190–199, include resource allocation optimization, wireless signal processing, shared
Oct. 2015. spectrum access communications, small cell networks, and stochastic geom-
[4] Study on Access Traffic Steering, Switching and Splitting Support in the 5G etry. He was a recipient of the Japanese Government (Monbukagakusho,
System Architecture, document TR 23.793 V0.1.0, 3GPP, Aug. 2017. formerly known as Monbusho) Scholarship for his Ph.D. study. He was a
[5] S. Han, C.-L. I, G. Li, S. Wang, and Q. Sun, ‘‘Big data enabled mobile recipient of the IEICE WBS Student Paper Award in 2013, the IEEE WPMC
network design for 5G and beyond,’’ IEEE Commun. Mag., vol. 55, no. 9, Best Paper Award in 2015, and the Young Researcher’s Encouragement
pp. 150–157, Sep. 2017.
Award from the Japan chapter of the IEEE Vehicular Technology Society
[6] X. Cheng, L. Fang, L. Yang, and S. Cui, ‘‘Mobile big data: The fuel for
in 2012.
data-driven wireless,’’ IEEE Intenet Things J., vol. 4, no. 5, pp. 1489–1516,
Oct. 2017.
[7] X. Cheng, L. Fang, X. Hong, and L. Yang, ‘‘Exploiting mobile big data:
Sources, features, and applications,’’ IEEE Netw., vol. 31, no. 1, pp. 72–79,
Jan./Feb. 2017. KIEN NGUYEN (SM’16) received the B.E. degree
[8] A. Engelbrecht, Computational Intelligence: An Introduction, 2nd ed. in electronics and telecommunication from the
New York, NY, USA: Wiley, 2007.
Hanoi University of Science and Technology, Viet-
[9] O. Acker, A. Blockus, and F. Pötscher. Benefiting From Big Data:
nam, in 2004, and the Ph.D. degree in informatics
A New Approach for the Telecom Industry. Accessed: Dec. 5, 2017.
[Online]. Available: https://www.strategyand.pwc.com/reports/ from the Graduate University for Advanced Stud-
benefiting-big-data and https://www.strategyand.pwc.com/media/file/ ies, Japan, in 2012. In 2014, he joined the National
Strategyand_Benefiting-from-Big-Data_A-New-Approach-for-the- Institute of Information Communications Technol-
Telecom-Industry.pdf ogy, Japan, as a Researcher. His research interests
[10] C. Jiang, H. Zhang, Y. Ren, Z. Han, K.-C. Chen, and L. Hanzo, ‘‘Machine include novel, software-based, and evolvable net-
learning paradigms for next-generation wireless networks,’’ IEEE Wireless working technologies for the next generation of
Commun., vol. 24, no. 2, pp. 98–105, Apr. 2017. mobile networks and Internet of Things. He is a member of the IEICE.

VOLUME 6, 2018 32337

 M. G. Kibria et al.: Big Data Analytics, ML, and Artificial Intelligence in Next-Generation Wireless Networks

GABRIEL PORTO VILLARDI (S’07–M’09– KENTARO ISHIZU received the Ph.D. degree in
SM’12) received the B.E. degree in electrical computer science from Kyushu University, Japan,
engineering with emphasis in telecommunica- in 2005. He has been working with National
tions from the Federal Center of Technological Institute of Information and Communications
Education of Rio de Janeiro, Rio de Janeiro, Technology (NICT), Japan, for 13 years. He has
Brazil, in 2002, and the M.E. and Ph.D. degrees been dedicated to R&D on heterogeneous wireless
(Hons.) as a Japanese Government (Monbuk- networks, cognitive radio systems, and TV white
agakusho) Scholar in physics, electrical, and space systems. He has been leading NICT’s TV
computer engineering from Yokohama National white space trials at various locations in the world.
University, Yokohama, Japan, in 2006 and 2009, He is currently managing the TV white space
respectively. In 2009, he joined the National Institute of Information and projects of NICT. One of the developed systems were sent to the disaster
Communications Technology, where he is currently a Senior Researcher area of the great earthquake in the eastern Japan on March 11, 2011, and
with the Wireless Systems Laboratory, Yokosuka Research Park, Yokosuka, contributed to recover from the damage of network access environment.
Japan. His current research interests span several areas in wireless commu- He also has been involved in wide area of international standardizations,
nications, such as communications theory, statistical modeling, multiple- including the IEEE 1900.4, IEEE 802.11, and IEEE 802.21.
input multiple-output, physical-layer design for white space cognitive radios,
space–time codes, fault tolerance, and energy reduction issues in sensor
networks. He has been actively contributing to the IEEE 802.19.1, IEEE
802.22b, IEEE 802.11af, and IEEE I900.6a task groups since 2010, toward
the standardization of TV white space and cognitive radio technologies. He is
currently a Voting Member of the IEEE 802.22, IEEE 802.11, IEEE 802.15,
and IEEE 802.19 standardization working groups. He has been serving as
the Secretary for the IEEE 802.22 Working Group and the IEEE 802.22b
Task Group since 2014. From 1999 to 2000, he received the Coordination
for the Improvement of Higher Education Personnel/Institute of International
Education Scholarship to pursue his studies at Clemson University, Clemson,
SC, USA. FUMIHIDE KOJIMA (S’96–M’99) received the
B.E., M.E., and D.E. degrees from Osaka Uni-
OU ZHAO received the B.E. degree in electronics versity, Osaka, Japan, in 1996, 1997, and 1999,
and communication engineering from the Nan- respectively, all in electrical communications
jing University of Posts and Telecommunications, engineering. He is currently the Director with
China, in 2005, and the M.E. and Ph.D. degrees in the Wireless Systems Laboratory, Wireless Net-
electronic engineering from the Graduate School work Research Institute, National Institute of
of Informatics, Kyoto University, Kyoto, Japan, Information and Communications Technology,
in 2014 and 2016, respectively. He is currently a Yokosuka, Japan. In 1999, he joined the Commu-
Researcher with the Wireless Networks Research nications Research Laboratory, Ministry of Posts
Center, National Institute of Information and Com- and Telecommunications, where he has been involved in research on various
munications Technology. His major research inter- topics, such as intelligent transportation systems, radio-over-fiber multi-
ests include resource allocation, wireless signal processing and its hardware media transmissions, mobile ad hoc emergency networks, wireless grid
implementation, machine learning, and big data analytics. He received the systems (smart utility networks), and medium access control protocol for
Young Researcher’s Encouragement Award from the Japan chapter of the communications systems.
IEEE Vehicular Technology Society in 2013.

32338 VOLUME 6, 2018