Recommender system commonly suffers from data sparsity and cold-start problems, where user-item interactions hinder reliable preference learning. While recent Graph Neural Network based models such as LightGCN and NGCF effectively capture higher-order collaborative signals, they primarily rely on interaction-derived embeddings and remain sensitive to sparse environments. This paper Attribute Enabled Graph Neural Framework (AE-GNF) proposes a Semantic-Aware Graph Refinement Framework that integrates attribute-driven representation learning with graph-based collaborative propagation to address these limitations. The proposed method first encodes heterogeneous user and item attributes using semantic embedding modules to generate informative initial representations independent of interaction density. Dense sematic embeddings are generated using modality specific neural encoders including transformer based text encoder for descriptive attributes, a recurrent attention network for behavioral interaction sequences and temporal contextual feature encoder for metadata signals. These embeddings are then refined through a normalized graph propagation mechanism that jointly models structural connectivity and semantic similarity, enabling robust higher-order preference learning. Unlike conventional recommenders, the framework preserves attribute semantics during message passing and enables inductive cold-start recommendations, where embeddings for newly introduced users or items are generated directly from attributes without requiring prior interaction edges. Experimental evaluation conducted on publicly available benchmark datasets including MovieLens-1M, Amazon Electronics, Amazon Books, and Amazon Prime Movies and LastFM360 demonstrates consistent performance improvements over Matrix Factorization, content-based models, GraphSAGE and Neural Graph Collaborative Filtering (NGCF). Results show notable gains in ranking accuracy, diversity and robustness under varying sparsity levels. The proposed AE-GNF achieves improved recommendation performance reducing RMSE by 3.5 to 6.2% and improving NDCG@10 by 6-11% compared to graph-based baselines across benchmark datasets. The findings confirm that integrating semantic attribute encoding with graph refinement provides a scalable and effective solution for next-generation recommendation systems operating in sparse and heterogeneous environments.