Machine learning has shown outstanding results on the enormous amount of data. Nevertheless, the accuracy of these methods often depends on the quality of the data including the amount of outliers - the objects, which are generated by behaviour different from the rest of the data objects. It is crucial to detect and treat outliers because machine learning methods can learn their deviant behaviour and provide inaccurate results. Our research provides an overview of existing outlier detection methods and proposes a practical guide for outlier detection in merchant vessel data. In this thesis, we perform exploratory data analysis and compare well-established and novel outlier detection methods using real-world data. We use optimization techniques for choosing parameters in DBSCAN, HDBSCAN and Bayesian DP-GMM. The performance of these methods is evaluated by using dimensionality reduction techniques including PCA and UMAP. NAPA Oy provides an internal static tabular dataset of 54.4k merchant vessels, which allows us to analyze almost all operating merchant vessels in the world according to Statista.com.
