Sequential recommender systems (SRSs) aim to predict the subsequent content in which users may be interested based on their past usage history. Existing solutions on SRSs focus on modeling sequential characteristics of user-service interactions and achieve promising performance. However, they do not take full advantage of one key factor that usually influences user behaviors: the category of services. It is necessary yet challenging to leverage category information due to two significant reasons. Firstly, bundling relationships exist between services/categories, which is vital for the prediction of user behaviors but hard to mine and encode. Secondly, since interest preferences and category preferences are closely related, their dynamic evolution has to be studied simultaneously. To tackle the above challenges, we propose a novel Dual-channel Preference Extraction and Refinement Network (DPERN) to extract users' multi-faceted preferences toward more accurate recommendation. For the former challenge, we leverage the co-occurrence information of services and categories to represent their intrinsic relationships and then adopt the graph embedding method to jointly pre-train their embeddings. For the latter challenge, we design dual preference extractors, each leveraging both service and category information, to capture interest preferences and category preferences, respectively. Moreover, we devise a preference refinement network to model the interaction between two extracted preferences, to enhance preference representations. Experimental results on three public datasets have demonstrated the effectiveness of the proposed DPERN model.