Outlier Detection Based on Low Density Models

Most outlier detection algorithms are based on lazy learning or imply quadratic complexity. Both characteristics make them unsuitable for big data and stream data applications and preclude their applicability in systems that must operate autonomously. In this paper we propose a new algorithm—called SDO (Sparse Data Observers)—to estimate outlierness based on low density models of data. SDO is an eager learner; therefore, computational costs in application phases are severely reduced. We perform tests with a wide variation of synthetic datasets as well as the main datasets published in the literature for anomaly detection testing. Results show that SDO satisfactorily competes with the best ranked outlier detection alternatives. The good detection performance coupled with a low complexity makes SDO highly flexible and adaptable to stand-alone frameworks that must detect outliers fast with accuracy rates equivalent to lazy learning algorithms.