Due to information overload, we are used to the personalization and recommendation techniques that are present in most e-commerce, information, or music systems, etc. They aim to filter out irrelevant information and show only accurate items. In a longer perspective, such an approach can provide a narrowing of the results, generating a filter-bubble effect and imposing polarization in social networks. The most popular solution to the problem is information or item diversification, but it can cause the accuracy of the recommendation to decrease. In the paper, we describe an overall idea of the Collective Recommendation System that aims to keep up-to-date information about users interests that allows to recommend accurate items to them. The system clusters users with similar interests and dynamics. The main contribution is to analyze the conditions to rearrange the clusters of dynamic users and discuss their usefulness in the re-clustering task.