Comparative Analysis of Outlier Detection Methods with Emphasis on DBSCAN

Outlier detection, also called anomaly detection or novelty detection, is a cornerstone in data analysis and machine learning, aiming to identify data instances that do not conform to expected behavior. Numerous approaches have been widely used in numerical domains; their performance on categorical datasets remains limited due to the complexity of distance computation and density estimation. This paper presents a comparative review of outlier detection methods on multilevel data. First, the major families of outlier detection methods –statistical-based, clustering-based, density-based, and ensemble-based are compared based on their working principle, strength, and limitations. Second, Clustering-based methods of outlier detection are examined. Third, several clustering-based algorithms such as K-Means, K-Modes, ROCK, DBSCAN, LOF, Hierarchical, CLIQUE, STING, and GMM are compared. A detailed comparison of DBSCAN with the well-established categorical clustering approaches ROCK and K-Modes is also presented to determine their underlying clustering, applicability to categorical data and outlier detection mechanism. To contextualize the comparison, several benchmark datasets for detection of outliers in categorical datasets are presented along with their characteristics. The current study presents a comparative analysis of outlier detection methods at multiple levels, and among these, the DBSCAN seems particularly suitable for identifying outlier instances. Its density-based measure makes it better suited to handling noise and clusters of arbitrary shape, and is independent of predefined clusters. Although DBSCAN is primarily designed for numerical data, its density-based framework can be extended to categorical datasets by incorporating an appropriate categorical similarity or dissimilarity measure. Developing such an adaptation provides a potential direction for applying DBSCAN-based outlier detection to categorical data.