The existing social tag recommending technology has the problems of data sparsity, high time complexity and low interpretability. To solve these problems, this paper proposed a tag recommending approach called TagRec-UPMF, which jointly factorizes the latent feature factor based on PMF. The approach jointly builds the corresponding feature vector in the form of probability, combining latent features of the three different facets of users, resources and tags, and then produces the top-N recommendation according to the linear combination of the inner products between the feature vectors of each pair. The proposed algorithm improves its accuracy in the case of the large size and sparse data, and it can be used for large-scale data due to the linear complexity. Experimental results show that our method has higher accuracy and lower time consuming than TagRec-CF, and Tucker, NMF, etc. Meanwhile, the proposed method has better precision than PITF algorithm when their complexity is of little difference. And our method shows lower complexity compared with TTD algorithm while their precision are nearly the same.