Quality of service, as mentioned in [33], refers to the collective effect of service performance provided to the user, which determines the level of user satisfaction with the service. Another definition of service quality defined for SLA in [3] refers to the quality of service as the degree of compliance of the service provided to the user following the agreement made in SLA.
A. Quality of service agreement
The overall quality of a telecommunications network is influenced by many factors that are related to the performance parameters of the network.
The SLA has a section on how to measure, monitor, and decide on the thresholds of Quality-of-Service parameters. This section is called the QoS agreement, as shown in Fig. 8. The purpose of this section in the SLA is to adjust the QoS parameters in a way that achieves end-user satisfaction. This section is shown in Fig. 8.
Defining QoS parameters is one of the essential steps of SLA and will affect the QoS agreement. Therefore, the definition of these parameters should be considered from different angles. In this paper, through reverse-engineering the QoS/QoE relationship, we provide a method for finding thresholds used in the QoS agreement section of the SLA. For this purpose, we harnessed the decision tree (DT) classification algorithm in a backward sense; we used the QoS intervals provided by the DT for the prediction of QoE level to set the standards in the SLA contract.
The decision tree is used to specify the permitted range of fluctuation for QoS features to satisfy a certain QoE score. In a sense, it would be a backward path that we take to set the feature thresholds for the corresponding QoE. We performed the backward process for all three MIN, MAX, MEAN databases to determine the service quality threshold in order to set the SLA at different levels. In addition to having these threshold values, customizable QoS features in the network will enable service providers to improve QoS features in order to bring the QoE score to the amount defined and agreed upon in the SLA. Therefore, QoE can (and should) be increased to an acceptable level by improving and changing these thresholds if QoE degradation is predicted by our online QoE prediction model constructed in the previous section. On the other hand, if the QoE score agreed in the SLA is set at a lower level, network productivity can be increased by having the appropriate service quality features threshold.
The decision tree is a much simpler classifier than a neural network. That justifies why we did not use DT for prediction purposes in the previous section. However, the very simplicity allows us to find easy criteria for QoS parameters that otherwise could not be set had we used a neural network or any other complex classifier. Nevertheless, in order to tune the decision tree classifier, we used feature selection. The feature selection mechanism at this stage was solely based on mutual information. After sorting the QoS features, we selected the top 20 features. As mentioned earlier, the number 20 was chosen based on experimental results. The reason we do not want to further filter these features is because we want to facilitate the internet service provider company with a higher degree of freedom in the QoS agreement.
B. QoS agreement adjustment
Through interactions between the Telecommunication Infrastructure Company [35] and providers, bounds and thresholds of Quality of Service to be mandated in SLA are usually determined. As such, IF-THEN rules could serve as a good candidate for describing such bounds and thresholds.
Most classification models in the literature that predict QoE based on QoS parameters are offered in a black-box manner. A common defect in black-box models is that they do not reveal enough of the hidden details. In the present work, however, we strongly suggest the opposite; that we need to explicitly state such information to justify the limits and thresholds declared in the SLA contract. For example, to determine the thresholds of QoS in the Gold contract, we need to know what level of delay could lead to the decline of customer satisfaction from Gold to silver level. Most black-box classification techniques, however, do not offer such granularity. Rule extraction algorithms, on the other hand, bring about such possibilities for us. However, one should realize that not only should the extracted rules be pinpointed, but also, they need to be simple and understandable by the customer. Rules are one of the most well-known symbolic knowledge extraction techniques [34]. In this work, extracted rules from our dataset are used in writing the SLA agreement.
The procedure in this section is to create a decision tree for each database using a number of service quality features that require a threshold. The reason for using the decision tree at this stage is to obtain service quality characteristics for different QoE classes. We will then extract IF-THEN rules using these trees. By combining these rules, we were able to get a good framework for different QoE standards in the QoS agreement. For the sake of convenience, we will mention only 5 of all the rules that can be obtained from decision trees.
In theory, service providers could use different sets of QoS features to tune their QoE. In this study, we assumed that the service provider is able to change and set the threshold for the 12 features listed in Table X. Thus, we made our quality service decision tree using these features.
C. If-Then rule extraction for MAX dataset
Table XI shows the rules extracted for the decision tree by the 12 service quality features mentioned. Many rules have been extracted from the decision tree from which the service providers can use according to their specific limitations as well as the special needs of their customers.
As can be seen from the structure of the decision tree, different rules can be defined for the Maximum QoE Score value, which can be selected to be included in the QoS agreement according to the conditions of service providers and the specific needs of customers. We performed up to 6 levels of pruning here to make it easier to draw the decision tree, but it goes without saying that if we do not prune, the set of rules will be much larger. For example, you can skip the feature selection step and use 42 properties to draw a decision tree, in which case we will have a much larger tree with many more branches and leaves.
D. If-then rule extraction for MEAN dataset
For MEAN dataset, the procedure was similar to MAX datasets. Here, according to the 12 features mentioned in Table XII, we performed the tree drawing operation and achieved the following results:
E. If-then rule extraction for MIN dataset
In MIN database, 12-service quality features used in MAX and MEAN databases were used and the rules corresponding to the QoE score were extracted.
For MIN, Table XIII shows the rules extracted from the decision tree by the 12 service quality features mentioned. The service providers can use any of these rules according to their specific limitations as well as the special needs of their customers. As shown in Tables XI, XII, and XIII, the Service Quality and Line performance features in all three databases had the most interaction with the class labels. This result is somewhat intuitively predictable. Because, for example, you cannot expect a high QoE score from a line with Line Quality of grade 3. In the other selected features for the three databases, there is a slight difference in the final list of selected features. In the next section, we will look at the combination of these rules for setting up a QoS Agreement in SLA.
F. QoS agreement adjustment in SLA
After obtaining the service quality feature thresholds for all three databases, it is time to combine these rules for inclusion in the QoS Agreement. Our proposed structure for setting up a QoS Agreement is illustrated in Table XIV.
Table XIV shows an example for using three criteria namely Maximum, Minimum and Mean. Obviously, the number of these levels can be reduced or increased. The service provider can adjust the SLA at different levels based on these defined levels. Each level can have different conditions in terms of pricing.
Now, we need to define IF-THEN rules for different databases according to different levels of service quality. To combine these rules, we use simple union and intersection as shown in Table XV. Combining these rules for the Gold1 level in QoS Agreement, we will have the rule in the last row of Table XV.
As mentioned before, Service Providers can use different branches of the decision tree to write their rules according to the available facilities and needs of their customers. We have only provided an example here as proof of concept.