Virtual worlds (VWs) are becoming effective interactive platforms in the fields of education, social sciences and humanities. Computing similarity among users is a technique commonly used to make friend recommendations in social networks. However, user communities in virtual worlds tend to have fewer real world linkages and more entertainment-related goals than those in social networks. The above characteristics result in an ineffective modality with respect to applying existing friend recommendation methods in virtual worlds. This study develops a virtual friend recommendation approach based on user similarity and contact strengths in virtual worlds. In the proposed approach, users' contact activities in virtual worlds are characterized into dynamic features and contact types to derive their contact strengths in communication-based, social-based, transaction-based, quest-based and relationship-based contact types. Classification approaches were developed to predict friend relationships based on user similarity and contact strengths among users. A novel friend recommendation approach is further developed herein to recommend friends as regards certain virtual worlds based on friend-classifiers. The evaluation uses mass data collected from an online virtual world in Taiwan, and validates the effectiveness of the proposed methodology. The experiment results show that the friend classifier that takes into account user similarity and contact strengths can elicit stronger prediction performance than the friend-classifier that considers only user similarity. Moreover, the proposed friend recommendation method outperforms the traditional friend of friend (FOF) method of friend recommendation in virtual worlds. User contact activities are characterized into dynamic features and contact types to derive their contact strengths.Classification approaches are developed to predict friend relationships based on user similarity and contact strengths.Novel friend recommendation approaches are further developed to recommend virtual friends based on friend-classifiers.