Towards a general cellular frustration approach to anomaly detection

• Autor(a) principal: Gonçalves, Rodrigo Faria
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/34080
• topic: Cellular frustration; Machine learning; Unsupervised learning; Anomaly detection; One-class support vector machines; Isolation forest; k-means

Cellular Frustrated Systems model the interactions between presenter and detector agents, with the goal of detecting anomalies in data sets. These two types of agents follow a frustrated dynamic (i.e., unstable), in which they continuously change the agent of the other type they are paired with, when a normal sample is presented by the presenter agents. In order for CFSs to make detections, presenters must show abnormal samples that have one or more abnormal features, which leads the agents to pair for longer times, leading to stable pairs, hence detections. This work improves upon the previous versions of this model by allowing detectors to see two regions of feature space as abnormal, with a normal region of feature space in-between. The K-means clustering technique is also used to cluster data in data sets, so that detectors are able to partition the feature space and be assigned to a certain region which they will see as normal, with the rest being seen as abnormal. It is shown that there is no need to train separate populations of detectors in order to make detections with data sets that have abnormal samples in-between normal ones. The current version of the model is compared with previous versions of the Cellular Frustration model, and also with two well known anomaly detection methods, the One-Class Support Vector Machines, and the Isolation Forest. The results show that it has comparable performance in relation to the competing methods, whereas regarding the previous versions, it is able to achieve the same results while being a more robust model applicable in more situations with less effort. Finally, some ideas for future work are discussed in order to further improve the model, which still has some issues when certain unfavorable conditions arise in data sets.