Service (API) discovery and recommendation is key to the wide spread of service oriented architecture and service oriented software engineering. Service recommendation typically relies on service linkage prediction calculated by the semantic distances (or similarities) among services based on theircollection of inherent attributes. Given a specific context (mashup goal), however, different attributes may contribute differently to a service linkage. In this paper, instead of training a model for all attributes as a whole, a novel approach is presented to simultaneously train separate models for individual attributes. Meanwhile, a latent attribute modeling method is developed to reveal context-aware attribute distribution. Experiments over real-world datasets have demonstrated that this fine-grained method yields higher link prediction accuracy.