Users are overwhelmed by the gigantic amount of information generated instantaneously by shopping sites, social networks, web services and so on. Having a relevant item (document, film, news, business, etc.) on time and without difficulty is a real challenge for developers. Recommendation systems are effective and powerful tools to address this problem. This research Intituled “Bi-embedding and metadata enhancement for collaborative prediction” (Bi-MeCp) proposes a recommendation method based on bipartite graph topology. To recommend an item to a user, we study the evolution of the graph to find the weights of non-existent links. The link weight between an item node and a user node measures the degree of importance for such a recommendation based on score similarity. The idea is to extract and aggregate the implicit and asymmetric connectivity “Bi-embedding” to build vector embeddings. Then, we structure and enrich the content of the item node with metadata (DC, MPEG, IEEE-lom, etc.) for personalized classification. The use of metadata allows for a variety of recommendations and goes beyond the tradition of being interested in recommending a single type of item. In this way, we have taken into account purely negative user judgments to refine and boost the recommendation process. This study has been tested on real datasets widely used in the field of recommendation and has produced encouraging and promising results.