Clustering in cellular frustration models

• Autor(a) principal: Carvalho, Diana Barros de
• Data de Publicação: 2018
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/25202
• topic: Data mining; Anomaly detection; Cellular frustration

Cellular frustrated systems are models of interacting agents displaying complex dynamics which can be used for anomaly detection applications. In their simplest versions, these models consist of two agent types, called presenters and detectors. Presenters display information from data samples. Detectors read this information and perceive it in a binary signal, depending on its frequency of appearance. The type of signal perceived will have an impact on the agents' decision dynamics. In particular, the presence of anomalies leads to less frustrated dynamics, i.e., more stable. In this thesis it is questioned if the mapping in binary signals could not bene t from the knowledge of the existence of clusters in the data set. To this end, a clustering technique was developed that gives particular attention to the fact that cellular frustrated systems discriminate samples depending on the number of features displaying rare values. The clusters obtained with this technique are also compared with those obtained using k-means or hierarchical agglomerative clustering. It is shown that using a clustering technique prior to application of cellular frustration system can improve anomaly detection rates. However, it is also shown that depending on the type of anomalies, this may not be generally the case, and therefore simpler cellular frustration algorithms may have the advantage of being simpler. It is believed that this study proposes new directions on how to improve the cellular frustration technique in a broader context.