Graph convolution has been the fundamental operator of contemporary network architecture for skeleton-based action recognition tasks. In previous approaches, graph convolution focuses on aggregating local features and learning the discriminative motion pattern from different semantic representations. However, these local processing methods are inefficient for capturing long-range dependencies between distant nodes, such as clapping. Moreover, the combination of different semantic representations is usually implemented through direct operations, which can become a bottleneck for the network performance. To alleviate the above issues, we propose a novel Graph Involution (GI) operator for capturing richer dependencies and a Dynamic Feature Fusion (DFF) module to enlarge the receptive fields and adaptively fuse the different semantic representations. We leverage the GI operator and DFF module to construct an effective feature extractor and light weight model, DFF-GIN, which achieves comparable results on NTU RGB+D 60 datasets for skeleton-based action recognition.