Date of Award
Master of Science (MS)
Outlier detection is one of the crucial tasks in data mining which can lead to the finding of valuable and meaningful information within the data. An outlier is a data point that is notably dissimilar from other data points in the data set. As such, the methods for outlier detection play an important role in identifying and removing the outliers, thereby increasing the performance and accuracy of the prediction systems. Outlier detection is used in many areas like financial fraud detection, disease prediction, and network intrusion detection.
Traditional outlier detection methods are founded on the use of different distance measures to estimate the similarity between the points and are confined to data sets that are purely continuous or categorical. These methods, though effective, lack in elucidating the relationship between outliers and known clusters/classes in the data set. We refer to this relationship as the context for any reported outlier. Alternate outlier detection methods establish the context of a reported outlier using underlying contextual beliefs of the data. Contextual beliefs are the established relationships between the attributes of the data set. Various studies have been recently conducted where they explore the contextual beliefs to determine outlier behavior. However, these methods do not scale in the situations where the data points and their respective contexts are sparse. Thus, the outliers reported by these methods tend to lose meaning. Another limitation of these methods is that they assume all features are equally important and do not consider nor determine subspaces among the features for identifying the outliers. Furthermore, determining subspaces is computationally exacerbated, as the number of possible subspaces increases with increasing dimensionality. This makes searching through all the possible subspaces impractical.
In this thesis, we propose a Hybrid Bayesian Network approach to capture the underlying contextual beliefs to detect meaningful outliers in mixed attribute data sets. Hybrid Bayesian Networks utilize their probability distributions to encode the information of the data and outliers are those points which violate this information. To deal with the sparse contexts, we use an angle-based similarity method which is then combined with the joint probability distributions of the Hybrid Bayesian Network in a robust manner. With regards to the subspace selection, we employ a feature engineering method that consists of two-stage feature selection using Maximal Information Coefficient and Markov blankets of Hybrid Bayesian Networks to select highly correlated feature subspaces.
This proposed method was tested on a real world medical record data set. The results indicate that the algorithm was able to identify meaningful outliers successfully. Moreover, we compare the performance of our algorithm with the existing baseline outlier detection algorithms. We also present a detailed analysis of the reported outliers using our method and demonstrate its efficiency when handling data points with sparse contexts.
Rentala, Girish Srivatsa, "" (2018). Thesis. 8.