Session

Computer Science and Communication Engineering

Description

The data size is increasing dramatically every day, therefore, it has emerged the need of detecting abnormal behaviors, which can harm seriously our systems. Outlier detection refers to the process of identifying outlying activities, which diverge from the remaining group of data. This process, an integral part of data mining field, has experienced recently a substantial interest from the data mining community. An outlying activity or an outlier refers to a data point, which significantly deviates and appears to be inconsistent compared to other data members. Ensemble-based outlier detection is a line of research employed in order to reduce the model dependence from datasets or data locality by raising the robustness of the data mining procedures. The key principle of an ensemble approach is using the combination of individual detection results, which do not contain the same list of outliers in order to come up with a consensus finding. In this paper, we propose a novel strategy of constructing randomized ensemble outlier detection. This approach is an extension of the heuristic greedy ensemble construction previously built by the research community. We will focus on the core components of constructing an ensemble –based algorithm for outlier detection. The randomization will be performed by intervening into the pseudo code of greedy ensemble and implementing randomization in the respective java code through the ELKI data-mining platform. The key purpose of our approach is to improve the greedy ensemble and to overcome its local maxima problem. In order to induce diversity, it is performed randomization by initializing the search with a random outlier detector from the pool of detectors. Finally, the paper provides strong insights regarding the ongoing work of our randomized ensemble-based approach for outlier detection. Empirical results indicate that due to inducing diversity by employing various outlier detection algorithms, the randomized ensemble approach performs better than using only one outlier detector.

Keywords:

outlier detection, ensemble outlier detection, greedy ensemble, randomized ensemble, ELKI

Proceedings Editor

Edmond Hajrizi

ISBN

978-9951-550-14-7

First Page

115

Last Page

119

Location

Durres, Albania

Start Date

7-11-2015 9:00 AM

End Date

7-11-2015 5:00 PM

DOI

10.33107/ubt-ic.2015.98

Share

COinS
 
Nov 7th, 9:00 AM Nov 7th, 5:00 PM

Randomizing Ensemble-based approaches for Outlier

Durres, Albania

The data size is increasing dramatically every day, therefore, it has emerged the need of detecting abnormal behaviors, which can harm seriously our systems. Outlier detection refers to the process of identifying outlying activities, which diverge from the remaining group of data. This process, an integral part of data mining field, has experienced recently a substantial interest from the data mining community. An outlying activity or an outlier refers to a data point, which significantly deviates and appears to be inconsistent compared to other data members. Ensemble-based outlier detection is a line of research employed in order to reduce the model dependence from datasets or data locality by raising the robustness of the data mining procedures. The key principle of an ensemble approach is using the combination of individual detection results, which do not contain the same list of outliers in order to come up with a consensus finding. In this paper, we propose a novel strategy of constructing randomized ensemble outlier detection. This approach is an extension of the heuristic greedy ensemble construction previously built by the research community. We will focus on the core components of constructing an ensemble –based algorithm for outlier detection. The randomization will be performed by intervening into the pseudo code of greedy ensemble and implementing randomization in the respective java code through the ELKI data-mining platform. The key purpose of our approach is to improve the greedy ensemble and to overcome its local maxima problem. In order to induce diversity, it is performed randomization by initializing the search with a random outlier detector from the pool of detectors. Finally, the paper provides strong insights regarding the ongoing work of our randomized ensemble-based approach for outlier detection. Empirical results indicate that due to inducing diversity by employing various outlier detection algorithms, the randomized ensemble approach performs better than using only one outlier detector.