Year of Graduation
Nowadays, tree-based algorithms are widely used in practice — from different machine learning competitions such as in the Kaggle service to business usage. It is became on of the first chosen models used by researches in the different data science problems. However, these methods are lack of automatic learning from unstructured data. Thus, researchers need to hand-craft a list of features at first. This study provides a tree-based approach that allows to learn different spatial patterns without need for manual feature engineering. This approach enables to build a Trajectory Decision Tree based on a trajectories data. The detailed algorithm and problem definition for building it is provided. Both, regression and classification problems are taken into account. Also, ensembling technique is applied to this method and led to the Trajectory Random Forest. Since, proposed model takes only trajectory data, the two approaches to mix Trajectory Tree with metadata are considered: stacking and separate training on clusters of similar trajectories. Provided methods are implemented in Python programming language and tested on the real dataset from the one of the Kaggle competition, which objective was to predict the taxi trip final destination. Also, models are compared with different well-known algorithms. The prediction performance is quite good and comparable with popular methods. Moreover, Trajectory Random Forest with separate training led to the best prediction on a local test dataset. Results are shown, that the tree-based approach provided in this study is worth using in practice and further discussion. Also, the possible directions of algorithm improvements are presented.