In this paper, we present an original mechanism based on a novelty detector filter for modelling user's profile in the context of content-based filtering systems. This filter learns possibly evolving user's need by means of positive and negative user's relevance feedback. An experiment on the behaviour of this filter that has been conduced on a corpus of several thousands of web sites taken from one of the main categories of the open directory DMOZ is also described. This experiment demonstrates the accuracy of our filter both for analyzing user's need and for highlighting useful alternatives for representing this need.