Based on our field studies and consultations with field experts, we identified three main problems that are of key importance to online web personalization and customer relationship management: 1) detecting changes in individual behaviour, 2) reporting on user actions that may need special care, and 3) detecting changes in visitation frequency.

Online personalised "my*" services are gaining popularity due to a growing customer need for information filtering and customisation. However, current systems mostly rely on some general usage and customer interaction in selecting components from prespecified blocks of content. The demand is great for high-quality unsupervised services on the customer side and for enabling techniques on the vendor side. Furthermore, individual profiles and, thus, personalised content should reflect changing individual behaviour. How do we efficiently build and maintain up-to-date personalised services for a large number of individuals? A compact and efficient, incrementally updatable representation of individual profiles is crucial. In addition, methods are required for efficient comparison of such profiles. Here we propose a methodology for building up-to-date personalised services. Individual profiles are represented as space-efficient prefix trees that are inherently easy to update incrementally. To measure the similarity of profiles, and also for the purpose of segmentation, we define a support-based metric that exploits the advantages of the tree-based structure. We evaluate our method on anonymised web data of 10,000 customers of an investment bank collected over 1.5 years.

On the Web, the intention of a user is mostly hidden. To approximate user intention and characterise user behaviour researches in Web usage mining mainly exploit two types of information: order and frequency of visited pages. However, several studies in information retrieval and human-computer interaction have suggested that the time spent on Web pages (TSP) is an important measure of user intention and page relevance. In our paper we provide a methodology to preprocess the TSP. In addition, we present a real-world testbed that provides an unbiased environment and representative, real-world data in specific Web domains. The environment can be used to evaluate user interest indicators and importance measures, to validate clustering algorithms and for a broad selection of other validation problems. As a case study, we define a testbed on online retail shop data and evaluate, among others, the relevance of TSP