Graph neural network (GNN)- based collaborative filtering has recently become a crucial method in recommender systems, as it can model higher-order user-item interactions. However, many existing GNN-based models primarily rely on interaction data, where they gather information from neighbors using predefined weight matrices. These methods are prone to major data sparsity issues because they do not effectively identify useful neighbors. Furthermore, they often overlook side information, even when it is available, despite its ability to reduce sparsity and improve item representations. To overcome these issues, we propose a heterogeneous GNN-based collaborative filtering model that combines interaction data and item attribute information. We use a random-walk-based similarity measure to find meaningful neighbors. We also introduce an attribute completion strategy that fills in missing item features using a transformer-based self-attention mechanism. This combination ensures that both structural and semantic signals are preserved that is helpful in case of high sparsity. Extensive tests on four diverse benchmark datasets show that our model consistently outperforms leading baselines by an average margin of 6.4 % and 5.3 % for NDCG and Recall measures respectively. The data and source code are available at https://github.com/tinkal87/GNNACRec/.