Consumption Behavior Analysis of Over the Top Services: Incremental Learning or Traditional Methods?

• Main Authors: Juan Sebastian Rojas, Alvaro Rendon, Juan Carlos Corrales
• Format: Article
• Language: English
• Published: IEEE 2019-01-01
• Series: IEEE Access
• Subjects: Classification algorithms; dataset; incremental learning; machine learning; OTT applications; supervised learning
• Online Access: https://ieeexplore.ieee.org/document/8845576/

Network monitoring and analysis of consumption behavior are important aspects for network operators. The information obtained about consumption trends allows to offer new data plans aimed at specific users and obtain an adequate perspective of the network. Over The Top applications are known by their large consumption of network resources. Service degradation is a common mechanism that applies limits to the amount of information that can be transferred and it is usually applied in a generalized way, affecting the performance of applications consumed by users while leaving aside their behavior and preferences. With this in mind, a proposal of personalizing service degradation policies applied to users has been considered through data mining and traditional machine learning. However, such approach is incapable of considering the swift changes a user can present in their consumption behavior over time. In order to observe which approach is capable of a continuous model adaptation while maintaining their usefulness over time, this paper introduces a performance comparison of traditional and incremental machine learning algorithms applied to information about users' Over The Top consumption behavior. Two datasets are implemented for the tests: the first one is built through a real network experiment holding 1,581 instances, and the second one holds 150,000 instances generated in a synthetic way. After analyzing the obtained results, the best algorithm from the traditional approach was a Support Vector Machine while the best classifier from the incremental approach was an ensemble method composed by Oza Bagging and the K-Nearest Neighbor algorithm.