Behaviour modelling with data obtained from the Internet and contributions to cluster validation

[EN]This PhD thesis makes contributions in modelling behaviours found in different types of data acquired from the Internet and in the field of clustering evaluation. Two different types of Internet data were processed, on the one hand, internet traffic with the objective of attack detection and on...

Descripción completa

Detalles Bibliográficos
Autor: Perona Balda, Iñigo
Tipo de recurso: tesis doctoral
Fecha de publicación:2016
País:España
Institución:Universidad del País Vasco
Repositorio:Addi. Archivo Digital para la Docencia y la Investigación
OAI Identifier:oai:addi.ehu.eus:10810/20620
Acceso en línea:http://hdl.handle.net/10810/20620
Access Level:acceso abierto
Palabra clave:behavioural modelling
data mining
pattern discovery
machine learning
unsupervised anomaly detection
clustering
cluster validity indices
suffix trees
sequential pattern mining
web personalization
user profiling
collaborative filtering
web mining
web log analysis
link prediction
network Intrusion Detection Systems (nIDS)
tourism website mining
profiling disabled people
Descripción
Sumario:[EN]This PhD thesis makes contributions in modelling behaviours found in different types of data acquired from the Internet and in the field of clustering evaluation. Two different types of Internet data were processed, on the one hand, internet traffic with the objective of attack detection and on the other hand, web surfing activity with the objective of web personalization, both data being of sequential nature. To this aim, machine learning techniques were applied, mostly unsupervised techniques. Moreover, contributions were made in cluster evaluation, in order to make easier the selection of the best partition in clustering problems. With regard to network attack detection, first, gureKDDCup database was generated which adds payload data to KDDCup99 connection attributes because it is essential to detect non-flood attacks. Then, by modelling this data a network Intrusion Detection System (nIDS) was proposed where context-independent payload processing was done obtaining satisfying detection rates. In the web mining context web surfing activity was modelled for web personalization. In this context, generic and non-invasive systems to extract knowledge were proposed just using the information stored in webserver log files. Contributions were done in two senses: in problem detection and in link suggestion. In the first application a meaningful list of navigation attributes was proposed for each user session to group and detect different navigation profiles. In the latter, a general and non-invasive link suggestion system was proposed which was evaluated with satisfactory results in a link prediction context. With regard to the analysis of Cluster Validity Indices (CVI), the most extensive CVI comparison found up to a moment was carried out using a partition similarity measure based evaluation methodology. Moreover, we analysed the behaviour of CVIs in a real web mining application with elevated number of clusters in which they tend to be unstable. We proposed a procedure which automatically selects the best partition analysing the slope of different CVI values.