Web Usage Mining: Data Pre-processing Impact on Found Knowledge in Predictive Modelling

Abstract

Abstract We describe the importance of pre-processing data in the web usage mining and the impact of mistakes to the analysis of the data during this phase. We analyse data of commercial bank and accesses of stakeholders to the selected part of the website according to the rules of Basel 2 Pillar 3. These rules are focused on the Market discipline and were created after the financial crisis in 2008. We model the time dependent behaviour of the web user. Modelling the probabilities of the accesses to web categories depending on time was done using the multinomial logit model, which is a part of generalised linear models. We found non-human access to the web portal during the modelling phase, which significantly influenced the obtained knowledge. We had to repeat the pre-processing phase and repeat the modelling. After the identification and removing of the problem accesses to web portal pages, we identified a new model parameter from which we calculated the logit estimates and subsequently estimated the probabilities of accesses to the web parts of the web portal. Our estimates of theoretical logits fit (model) empirical logits. It is essential to not underestimate the data pre-processing phase in the process of web usage mining, as the pre-processing phase directly affects the quality of the acquired knowledge.