4of16 XU ET AL.
as the postprocessing of the neural network weights to reduce the errors between the predicted and subjective MOSs. It
indicates that traditional neural networks usually suffer from low accuracy and slow coverage.
Most studies mainly focus on one kind of service, such as video, voice, or web service. Zhang et al
43
improved the
perceptual evaluation of speech quality by replacing bark-scale frequency with equivalent rectangular bandwidth–scale
frequency. Oche et al
9
proposed a QoE prediction approach for real-time video service, which mapped multidimension
influence factors into user QoE by using the ordinal regression analysis. Kim and Choi
14
defined a video QoE assessment
model by modifying the existing mathematical model (E-model) to measure the subscriber QoE. Upadhyaya et al
24
pro-
posed a craw ling-based solution to obtain user perception about web service and found the strong positive correlation
between QoS and QoE. For VoIP service over mobile networks, De Pessemier et al
23
found that the used network and
handovers occurred during the voice call have a significant impact on user QoE. We can see that these models severely
depend on service identification in practice and the whole system needs to be retrained to match newly emerging services.
In recent years, the researchers have mainly focused on QoE prediction in terrestrial networks, such as mobile net-
works, IoT networks, cloud-based networks, and cellular networks. In Vizzarri and Davide,
44
the linear QoS/QoE model
of over-the-top–managed services over LET network was derived. In Sun et al,
25
it was validated that the exponential
model is more suitable for QoE/QoS modeling in IoT networks. Hsu and Lo
45
designed a QoS/QoE mapping model for
cloud-based video service with considering buffering time and streaming video discontinuity. The influence of access
bandwidth and latency on the QoE was studied for cellular networks in Casas et al.
26
The specific features of satellite net-
works (such as high bit error rate related to hard environment
46,47
and long propagation delay) make them different from
terrestrial networks; thus, the above models cannot be applied to satellite networks directly. To fill this gap and improve
the performance of QoE/QoS model, we proposed a deep and modular QoE/QoS mapping method.
3 QOE/QOS MAPPING METHOD
The proposed modular and deep QoE/QoS mapping method for multimedia services over satellite networks is described
in this section. As illustrated in F igure 1, a satellite network simulator is used to simulate the service delivery over prac-
tical network, and the subjective test is performed to collected subjective opinion scores. The QoS parameters of s atellite
networks are monitored when multimedia services are delivered over the network and the corresponding subjective opin-
ion scores given by observers are collected. In this way, a dataset containing subjective opinion scores and associated QoS
parameters is constructed to train the modular and deep neural network, which can be used to predict the user QoE based
on measurable QoS parameters, including delay, jitter, and packet loss. More specially, the complex task of QoE prediction
is divided into several subtasks by using the traffic classification–based task decomposition module. Then, each subtask
is to build QoE/QoS correlation model for a certain kind of multimedia service by using a separated and simple DBN.
Finally, the outputs of each neural network are integrated together as the predicted QoE.
3.1 Task decomposition
In this paper, the traffic classification based on K-means clustering (see more details in Erman et al
48
) is exploited to divide
the task of QoE prediction for satellite multimedia services into K different subtasks; that is, the task of QoE prediction
is divided according to the service type. The average packet size, connection duration, packet number, delay, jitter, and
packet loss are taken as the distinctive features of flows to identify their service type. Let X ={x
1
, x
2
, … , x
T
} be a set of
flows and T is flow number. The centers of the K different subtasks are c ={c
1
, c
2
, … , c
K
} , which are determined by
K-means clustering. For a delivered flow x
t
, t ∈{1, 2, … , T} is characterized by a feature vector x
t
={x
tm
|1 ≤ m ≤ M},
where x
tm
is the mth feature of the tth flow and M is the number of measured features. The distance between x
t
and
cluster center c
k
can be calculated by Equation 1. The membership degree that flow x
t
belongs to the kth service type can
be calculated by Equation 2
d
tk
=
√
|x
t
− c
k
|
2
, k = 1, 2, ···, K (1)
𝑓
tk
= e
−
d
2
tk
0.02
. (2)
According to the relationship between membership degree f
tk
and the threshold of membership degree D,theservice
type to which flow x
t
belongs is determined. If f
tk
≥ D,thenx
t
belongs to the kth service type and its cluster center is c
k
.