Packet Discards and User Behaviour Pattern Recognition in Mobile Networks: Unsupervised Temporal Time Series Segmentation and Categorisation of Time Series
Sapozhnikov, Aleksei (2021)
2021
Tietotekniikan DI-ohjelma - Master's Programme in Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. Only for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2021-12-14
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202111158404
https://urn.fi/URN:NBN:fi:tuni-202111158404
Tiivistelmä
Mobile network technologies continue to evolve rapidly and are enabling technological innovation in many fields of human activities, both in the private and business domain. Mobile network users expect differentiated service levels and mobile operators are urged to provide corresponding service levels and maintain their level of services to retain customers. There are, however, a number of challenges impeding the coverage and capacity capabilities of mobile networks, which may lead to the impairment of quality of services and disgruntled customers with unmet expectations. Mobile subscribers do not think of their music listening, video watching or internet browsing in network terms as packet loss, bandwidth or throughput capacities, they rather consider how much time it takes to download the content of a web page or the reliability of video streaming. The better radio conditions mobile services providers supply, the greater the number of contented mobile network users. The aim of the work was to create a tool which would assist in the analysis of subscribers’ behaviour in those conditions when the network is not functioning properly from a perspective of a subscriber.
The thesis work was conducted at Data Collection and Analytics Platform (DCAP) department of Nokia Solutions and Networks Oy (Nokia – FI/Espoo). The department works with different types of mobile networks data, and one of them is “Per Call Measurement Data" (PCMD) which provides technical information on each call. PCMD data allows the monitoring of the state of quality of mobile networks and calls made over the air interface. Completed work was solely based on analysis of PCMD of those calls in which packet discards happened, or in other words when the network is not functioning properly. Although PCMD provides a rather accurate view of user experience, the thesis focuses on overall end user behaviour analysis, to understand how an end user acts when quality of the connection has deteriorated. One of the indicators of quality of connection is number of packet discards. The purpose of the work was to find out what types of user behaviour patterns exist when packet discards occur to categorise drop periods according to their types with the help of a tool for segmentation of calls represented in time series of PCMD data.
The original idea was to use unsupervised learning methods to create a tool which would automatically detect packet discard time periods and further segment soi-disant call. The main methods to create the analysis tool were kernel density estimation and information gain-based temporal segmentation. Kernel density estimation was employed to split a call into the parts of a period before drop, a drop period, and a post-drop period after the drops. Information gain-based temporal segmentation was used to detect changes in the post-drop period and segment it into three consecutive post-drop periods. As a result of this work the segmentation tool was created. The tool helps detect packet discard periods in calls with further segmentation of post discard period into three consecutive periods. This tool facilitates the work with a PCMD database for radio technology experts and helps to analyse calls and consider calls from a subscriber’s perspective.
The thesis work was conducted at Data Collection and Analytics Platform (DCAP) department of Nokia Solutions and Networks Oy (Nokia – FI/Espoo). The department works with different types of mobile networks data, and one of them is “Per Call Measurement Data" (PCMD) which provides technical information on each call. PCMD data allows the monitoring of the state of quality of mobile networks and calls made over the air interface. Completed work was solely based on analysis of PCMD of those calls in which packet discards happened, or in other words when the network is not functioning properly. Although PCMD provides a rather accurate view of user experience, the thesis focuses on overall end user behaviour analysis, to understand how an end user acts when quality of the connection has deteriorated. One of the indicators of quality of connection is number of packet discards. The purpose of the work was to find out what types of user behaviour patterns exist when packet discards occur to categorise drop periods according to their types with the help of a tool for segmentation of calls represented in time series of PCMD data.
The original idea was to use unsupervised learning methods to create a tool which would automatically detect packet discard time periods and further segment soi-disant call. The main methods to create the analysis tool were kernel density estimation and information gain-based temporal segmentation. Kernel density estimation was employed to split a call into the parts of a period before drop, a drop period, and a post-drop period after the drops. Information gain-based temporal segmentation was used to detect changes in the post-drop period and segment it into three consecutive post-drop periods. As a result of this work the segmentation tool was created. The tool helps detect packet discard periods in calls with further segmentation of post discard period into three consecutive periods. This tool facilitates the work with a PCMD database for radio technology experts and helps to analyse calls and consider calls from a subscriber’s perspective.