It　is　already　well　known　that　recommender　systems　usually　suffer　from　data　sparsity　issue　of　user-item　interactions.　However,　representation　learning　can　efficiently　measure　correlations　between　objects,　which　presents　an　unprecedented　opportunity　to　alleviate　this　issue.　In　this　paper,　we　propose　a　new　distributional　vector　space　model,　Tag2Vec,　for　capturing　meaningful　relationships　of　users　and　items　to　improve　the　performance　of　recommender　systems.　First,　we　represent　users　and　items　as　vectors　respectively　using　tag　embedding.　With　this　innovative　representation,　the　semantic　relationships　between　users　and　items　can　be　captured.　To　be　specific,　tag2vec　learns　representations　of　users　and　items　in　low-dimensional　space　from　user-tag-item　interactions　using　the　skip-gram　model.　Second,　we　measure　similarity　between　both　users　and　items,　and　collaborative　filtering　can　then　be　performed　in　the　learned　embedding　space.　To　evaluate　the　performance　of　Tag2Vec,　we　conduct　extensive　experiments　with　two　real　world　datasets　for　Top-N　recommendation　tasks.　The　results　demonstrate　that　our　proposed　method　significantly　outperforms　existing　approaches.　©　2020,　Springer　Nature　Switzerland　AG.