| Title |
Registration of functional data / |
| Translation of Title |
Funkcinių duomenų registravimas. |
| Authors |
Petronis, Andrius |
| Full Text |
|
| Pages |
46 |
| Keywords [eng] |
Key words: landmark registration, continuous registration, time-warping functions, k-means clustering, non-homogeneous Poisson process. |
| Abstract [eng] |
This paper investigates both landmark and continuous registration methods for functional data that are expected to reduce phase variation between observations. The empirical part is dedicated to the analysis of IT system issues within a company, which satisfies all necessary conditions of the non-homogeneous Poisson process. Results reveal the existence of two clusters that could be registered separately using a two-step approach. This leads to lower variance and more precise representation of the underlying process, compared to unregistered cases, while there are no significant changes in pointwise probability density functions. Also, more than 95% of all data points lies within a 2-sigma range away from the mean function, while only less than 40% of functions are fully covered with these boundaries. Moreover, registration reduces maximum deviations from the mean functions. Therefore, these results correspond to the existing literature and extend knowledge about potential benefits of applied registration. |
| Dissertation Institution |
Vilniaus universitetas. |
| Type |
Master thesis |
| Language |
English |
| Publication date |
2017 |
