In this paper we present a novel approach for anonymous recommendation that combines the knowledge-based recommendation with a learning component. The novelty of the approach is specially expressed through a new approach for cascade combination of two systems, that enables better the usage of prior knowledge in the learning process. Secondly, the training data for learning is based on a complex description of a recommendation situation, that enables significant improvement in the quality of recommendations in unseen situations. The approach has been implemented in a real e-commerce environment and tested in large dataset based on the twenty months usage data. The results of the evaluation studies show that this kind of hybrid systems is very promising for ephemeral recommendation.