With the increasing prevalence of online businesses and social networking services, a huge volume of data about transaction records and social connections between users is accumulated at an unprecedented speed, which enables us to take advantage of electronic word-of-mouth effect embedded in social networks for precision marketing and social recommendations. Different from existing works on social recommendations, our research focuses on discriminating the community-level social influence of different friend groups to enhance the quality of recommendation. To this end, we propose a novel probabilistic topic model integrating community detection with topic discovery to model user behaviors. Based on this model, a recommendation method taking both individual interests and conformity influence into consideration is developed. To evaluate the performance of the proposed model and method, experiments are conducted on two real recommendation applications, and the results demonstrate that the proposed recommendation method exhibits superior performance compared with the state-of-art recommendation methods, and the proposed topic model exhibits good explainablibity of topic semantics and community interests. Furthermore, as some people are more individual interest oriented and some are more conformity oriented demonstrated by the experiments, we explore factors that influence each individual’s conformity tendency, and obtain some meaningful findings.