Community detection aims to identify groups of densely connected nodes in a graph, with modularity maximization serving as the most widely used objective. Modern networks are often richly annotated with node attributes, offering an opportunity to explain community structure in terms of the observable node characteristics. However, existing methods either ignore node attributes or rely on graph-dependent latent representations that offer limited interpretability. In this work, we propose a novel approach for community detection in attributed graphs that maximizes modularity while learning an inherently interpretable mapping from node attributes to communities. Specifically, we present a gradient boosting framework based on a new differentiable objective, exponentiated soft modularity, designed to improve gradient-based modularity maximization. To ensure interpretability, we enforce an additive structure that yields transparent attribute-level explanations. Experiments on eleven real-world networks show that our boosting framework achieves higher modularity than existing mapping-based neural baselines, while producing inherently interpretable mappings.