The surging data traffic and dynamic user mobility in 5G networks have posed significant demands for mobile operators to increase data processing capacity and ensure user handover quality. Specifically, a cost-effective and quality-aware radio access network (RAN) is in great necessity. With the emergence of fog-computing-based RAN architecture (Fog-RAN), the data processing units (BBUs) can be separated from base stations (RRHs) and hosted in distributed fog servers, where each server accommodates a community of RRHs to handle data traffic and user handover. The key problem in Fog-RAN optimization is how to cluster complementary RRHs into communities and allocate adequate numbers of BBUs for the fog servers, since real-world traffic and mobility patterns are highly dynamic to model, and it is not trivial to find an optimal RRH clustering and BBU allocation scheme from potentially enormous numbers of candidates. In this work, we propose a data-driven framework for cost-effective and quality-aware Fog-RAN optimization. In the RRH clustering phase, we build a weighted graph model to characterize user mobility patterns across RRHs, and propose a size-constrained community detection (SCUD) algorithm to cluster RRHs into communities with frequent internal handover events. In the BBU allocation phase, we formulate BBU allocation in each community fog server as a set partitioning problem, and propose a column-reduced integer programming (CLIP) algorithm to find optimal BBU allocation schemes that maximize BBU utilization rate. Evaluations using two large-scale real-world datasets collected from Ivory Coast and Senegal show that compared to the traditional RAN architecture, our framework effectively reduces the average handover overhead to 12.8% and 27.3%, and increases the average BBU utilization rate to 49.7% and 52.3% in both cities, respectively, which consistently outperforms the state-of-the-art baseline methods.