With the massive amount of social media data becoming available, there is a rising interest in automatic metaphor detection and interpretation from open social text. One of the most well-known approaches to this subject is identifying the violation of selectional preference. The basic concept of selectional preference is that verbs tend to have semantic preferences of their arguments and that violations of these preferences are strong indicators of metaphorical language use. Nevertheless, previously, few works have focused on metaphor detection of social media data. In response to this problem, we propose a three-step framework that is based on the technology of selection preference modeling to detect metaphors in social media data. We conduct a pilot study of this framework on the data of a real-world online support group. Furthermore, to improve our approach, we also leverage topical analysis techniques in our framework. As a result, we address the challenges of the task of metaphor detection in social media data, provide qualitative analysis for our experiments, and illustrate our insight based on the results.