In order to find out more location information from users, the user feature is dug out from location semantic similarity. The location topic model is built for user sign-in information by using LDA algorithm, and distribution functions can be calculated by the Gibbs sampling algorithm. The user similarity feature vector based on sign-in location semantic is put forward by these distribution functions. Then, the supervised machine learning algorithm is put forward to make link prediction by multi-dimensional similarity feature vector from fusing LBSN network structure information, sign-in location information and location semantic similarity. The experiments result on Gowalla databases shows that the link prediction algorithm using more similarity feature as subsidiary information can improve performance of LBSN link prediction significantly comparing with the traditional algorithm.