Personalized recommender systems have become increasingly popular in recent years, as they have the ability to make appropriate choices for each active user. Collaborative filtering (CF) is the most successful and widely used technique in recommender systems, which aims at discovering similar users or items based on the history user rating records, i.e., user-item matrix. However, CF may not generate good recommendations when user-item matrix is very sparse. To address this problem, we explore the property category and semantic content to reduce the amount of items, which lead to more accurate performance when estimating user similarity. In addition, since the amount of users is quite huge, we first profile similar users with the aid of clustering algorithm before recommendation. Then, for each active user, the CF recommender system returns top recommendations from the narrow-down cluster the same as the active user by calculating user similarity with the help of item semantic information. The experiments have been performed on the benchmark dataset in NLPCC 2017 to recommend point-of-interest (POI) for each active user. The comparative results demonstrate that our proposed model outperforms the two baselines (i.e., a user-based CF system and an item-based CF system).