ONLINE CHANGE POINT DETECTION USING BAYESIAN METHOD WITH MODEL SELECTION

• Main Author: Stephanie
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/47785
• Institution: Institut Teknologi Bandung

Changepoint detection is a process of detecting changes that occur in the process of generative sequential data (time series data). Analysis of change point detection has been applied in various fields of research including climate change, robotics, and telecommunications. Changepoint detection methods can be divided into offline and online methods. In its use, online methods are used more widely by the manufacturing industry in the process of quality control and in the medical world to diagnose diseases. This study aims to build an online change point detection model using Bayesian method by providing additional model selection factor, which is called BOCPDMS model (Bayesian Online Changepoint Detection with Model Selection). This model is a modification of the BOCPD model developed by Adams and Mackay (2017). The main difference between BOCPDMS is the addition of several time series models, namely AR-1, AR-2, and AR-3 and there is a model selection factor at all times. To find out the level of goodness of BOCPD and BOCPDMS, a random simulation of change point detection was done using artificial data. The sensitivity analysis is carried out to see the effect of each hyperparameter on the model and comparative analysis of the two models using the same simulation data is done to see which model performs better. The random simulation results of the detection of change points show that BOCPD has an accuracy value of 97%, while BOCPDMS has an accuracy value of 97.5%. Based on sensitivity analysis, BOCPD hyperparameters that affect the number of detected points are ????!, ????!, ????!, ???? and BOCPDMS hyperparameters that affect the number of detected points are ????!", ????!, ????!. After comparing the results on the same simulation data, BOCPDMS model was able to detect the change points more optimally than BOCPD model. In addition, with BOCPDMS model, several hyperparameters initiation values can be included in each data distribution model package.