Multiobjective trees for forecasting
Loading...
Date
2023
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Thesis (M.S.) - Bogazici University. Institute for Graduate Studies in Science and Engineering, 2023.
Abstract
Time series forecasting is a significant task, and it can be used in various fields such as education, finance, medicine, and manufacturing. The wide use of temporal data has resulted in many researches and studies in the field of data mining. Besides time series analysis models, tree-based methods are commonly used for time series forecasting. Tree-based models provide accurate results in complex datasets, and they can explain nonlinear relationships. However, time series data is obtained by measuring a variable at certain time intervals and often contains predictable fluctuations. In the use of tree- based modeling, the temporal relationship and seasonality in time series data should be included in the modeling process. Regression trees consider all data points independently of each other. Thus, preprocessing or postprocessing methods are needed for their use in time series forecasting. However, temporal and seasonality effects can be incorporated into learning process of trees. In this thesis, Multiobjective Trees for Forecasting (MTF) Model is proposed, which is a tree-based ensemble approach that makes changes in the learning phase of decision trees. In addition to objective of decision trees, two new considerations are added to be taken into account during partition. These are temporal closeness in nodes and the seasonal distance between observations. A single decision metric is created by combining these three objectives in different weights, and splitting is done according to this decision metric. In the experiments, five different synthetic datasets and three time series datasets are used. Adding these objectives yields better results than the traditional approach, when the weights are selected with parameter tuning. Finally, multiobjective trees are tested in randomized setting similar to random forest models and compared with the traditional time series forecasting models. More competitive results are obtained when only one decision tree is fitted compared to randomized setting.