Improvement of recommendation systems based on cellular learning automata

Recommendation systems suggest useful information to the end users. They predict the information demands of online users and offer recommendations to facilitate their navigation. There are many approaches to construct such systems. Most of the recommendation systems use data mining techniques on the web access log or the database of the site and discover user´s access patterns. Afterwards, the recommendation system uses these patterns to recommend useful information to users. However, there is a problem that they need to periodically update extracted patterns to make sure they still reflect the trends of the users. Dynamic recommendation systems can continuously interact with the users and learn their behavior. One group of this kind of systems is based on asynchronous cellular learning automata (ACLA). In this paper, we improve the reward and punishment criterion in ACLA recommender systems. We take into account the similarity of the information that recommender system suggests to the user and the information that the user actually requests. Then, the reward or punish actions in ACLA is computed base on this similarity. This results in a hybrid system that uses two sources of data, content of items and data usage. This system keeps the simplicity of the ACLA recommendation system based on collaborative filtering as well as it can achieve the accuracy and quality of hybrid recommendation systems. This is achieved by considering the higher similarity between recommendation and users need. Our experiments on actual datasets indicate that our hybrid system can improve the quality of recommendations.