Page 31 - Proceedings of the 2018 ITU Kaleidoscope

Machine learning for a 5G future

5.3 QoS criteria and KQIs Importance of QoS criteria for users
The data collected from the probes and the surveys shed
light on the important differences between the users’
influence factors in the different scenarios. In addition, the
results of the surveys about the user’s requirements for each
of the scenarios also showed that the relevant QoS criteria
and KQIs differ from one scenario to another (Figure 10). Commercial
Although the use of ML to automatically update the KQIs
has not been yet tested, it has been proven that it is essential
given the number and changing nature of the influence
factors.

5.4 Control plane: Anomaly detector and predictor

The data collected from the probes have been used, not only
to learn about the users' behavior and extract the context
information, but also to detect anomalies and enhance the Campus
channel selection process through unsupervised ML.

5.5 User plane: Supervised ML

Based on previous research studies [10], inductive
supervised learning has been employed for the NP-QoE
correlation model. The results of the surveys about user’s
QoE and satisfaction have fed the model to infer the rules to
automatically predict the QoE based on NP and the
influence factors. In future stages of the study ML Residential
techniques will be also implemented to enhance the
satisfaction model (CSAT) described in [14].

5.6 Results: Corrective actions to enhance QoE
Figure 10. KQI relevance in different scenarios
Though the validation of the methodology is still at a
premature stage, the case study has revealed that the
proposed methodology can be very useful to deploy the 6. CONCLUSIONS
QoS management model and enhance the user´s QoE. In
fact, the results have indicated several corrective actions In this paper a methodology to implement a global QoS
that could be implemented through ML in the scenarios management model for the next-generation wireless
under study: ecosystem, taking advantage of big data and ML techniques,
has been presented.
 Commercial scenario: one of the most relevant KQIs
in this scenario (streaming video/audio application Taking into account international standards, the QoE-
performance) is affected for NP problems centric approach makes use of supervised ML techniques in
(continuous disruptions of the service). Corrective order to identify the KQIs relevant for the users.
action: Analysis in NP data of AP capacity and use Unsupervised ML mechanisms are proposed for the
of ML to enhance the channel selection mechanism. identification of the user’s influence factors, network
 Campus Scenario: Some client-association problems performance anomalies, faults detection and channel
were found due to bad AP configuration. selection enhancement. The approach links the NP and QoE
Furthermore, the students were dissatisfied with one via inductive ML techniques and provides the intervention
of the KQIs (the ease of login) so the procedures points where corrective actions are required.
around the login procedure should be revised.
Corrective actions: Customize AP performance Although the definition of the methodology and the
through ML techniques and revise login procedures validation of the approach is still at an early stage, the
according to the learned rules. results of the case study, carried out on a number of
 Residential scenario: Lower cost and higher network different Wi-Fi scenarios, reveal that the methodology may
speed were two of the key requirements in this aid enhanced QoE in next-generation wireless
scenario. Corrective actions: Customize residential environments. In addition, some recent studies suggest that
Wi-Fi APs for optimal throughput through ML and ML techniques may be applied for customer retention to
enhance the business model using supervised ML enhance the QoBiz and this will be also analyzed at future
through the survey's observation set. research stages of this proposal.

– 15 –