Changing User Interests through Prior-Learning of Context

Abstract

The paper presents an algorithm for learning drifting and recurring user interests. The algorithm uses a prior-learning level to find out the current context. After that, searches into past observations for episodes that are relevant to the current context, ‘remembers’ them and ‘forgets’ the irrelevant ones. Finally, the algorithm learns only from the selected relevant examples. The experiments conducted with a data set about calendar scheduling recommendations show that the presented algorithm improves significantly the predictive accuracy.