Knowledge Graph Completion (KGC) has gained significant popularity for improving the completeness of knowledge graphs by predicting missing relations between entities. However, traditional methods do not adequately capture contextual information and interactions between entities and relations. In this paper, we propose a model called Data-Augmented Hierarchical Feature Aggregation Graph Neural Network (DA-HFA), which integrates graph-structured information through multiple independent encoders to explore various concrete and interpretable knowledge compositions. Specifically, in DA-HFA, each encoder is explicitly modeled using neighborhood relationships and effectively combined through multi-layer aggregation, resulting in a more comprehensive representation of information. Moreover, we design a data augmentation method that utilizes random walks on the knowledge graph to increase the number of plausible triples. By providing richer and more diverse training datasets, this method overcomes the limitations of traditional methods in dealing with data sparsity and diversity. Finally, DA-HFA has proven superior to state-of-the-art baselines through extensive experiments on numerous benchmark datasets, including FB15k-237 and WN18RR, demonstrating its effectiveness across various domains and scenarios.