Graph Neural Networks (GNNs) have performed well in collaborative filtering by capturing complex user-item interactions. Still, most graph-based recommender systems rely solely on interaction data and overlook valuable information such as item attributes, relationships, and hidden community structures. In this study, we introduce a new community-aware recommendation approach that leverages behavioral and semantic item relationships to enhance graph-based recommendations. We start by building a combined item graph that merges global item transition patterns with content-based semantic similarities. To find meaningful item communities, we use modularity-guided Louvain community detection on this graph. We then add the discovered community structure to a community-aware Light-GCN model to improve how items and users are represented. Tests on two benchmark datasets, Yelp and MovieLens-1M, show that our method consistently beats several leading baselines in Recall and NDCG. Further analysis shows that modularity-guided pruning makes the graph better and directly leads to more accurate recommendations. These findings show that adding item community knowledge can make graph neural recommendation models more effective.