NeoLOD: A Novel Generalized Coupled Local Outlier Detection Model Embedded Non-IID Similarity Metric.

Traditional generalized local outlier detection model (TraLOD) unifies the abstract methods and steps for classic local outlier detection approaches that are able to capture local behavior to improve detection performance compared to global outlier detection techniques. However, TraLOD still suffers from an inherent limitation for rational data: it uses traditional (Euclidean) similarity metric to pick out the context/reference set ignoring the effect of attribute structure. i.e., it is with the fundamental assumption that attributes and attribute values are independent and identically distributed (IID). To address the issue above, this paper introduces a novel Non-IID generalized coupled local outlier detection model (NeoLOD) and its instance (NeoLOF) for identifying local outliers with strong couplings. Concretely, this paper mainly includes three aspects: (i) captures the underlying attribute relations automatically by using the Bayesian network. (ii) proposes a novel Non-IID similarity metric to capture the intra-coupling and inter-coupling between attributes and attribute values. (iii) unifies the generalized local outlier detection model by incorporating the Non-IID similarity metric and instantiates a novel NeoLOF algorithm. Results obtained from 13 data sets show the proposed similarity metric can utilize the attribute structure effectively and NeoLOF can improve the performance in local outlier detection tasks.