As one of the effective methods to ease the information overload problem, recommender systems have become extremely popular in social media. However, recommender methods suffer from the cold-start problems in new item recommendations and new user recommendations. To combat the cold-start problems in new item recommendations, the concept of user time weights is proposed to characterize the time interval between the user evaluating time and item distributing time. According to the weights, it can determine whether the user is a positive user or a negative user, and the degree of the user's preference for new items. Tripartite graphs are used to picture relations among user-item-tag, and user-item-attribute. Combing information among users, tags, attributes of items and time weights, functions for predicting the rating are defined and a new personalized recommendation algorithm is constructed. Overall experimental results show that the proposed method not only brings satisfied personalized items but also pleasantly surprises different users with different preferences. Comparative experiments illustrate the proposed method is much higher in accuracy and novelty. Cross-validation experiments demonstrate that the new method is effective to solve the cold-start problem in new item recommendations.