Recommendation systems are needed to handle large-scale user-item interactions on digital platforms, but they often face severe data sparsity and cold-start issues. Current collaborative filtering and deep-learning models are frequently oblivious to complex contextual dependencies and difficult to scale computationally. To address these limitations, the paper presents a context-sensitive recommendation framework built on deep learning that integrates autoencoder-based latent feature generation with a neural interaction model. The proposed model effectively captures high-order user-item-context correlations in a sparse environment and maintains computational efficiency suitable for large-scale deployment. Contextual cues, including time, category, and geographic ones, are combined through a hybrid manifold that optimizes reconstruction and prediction losses simultaneously. The model consistently outperforms state-of-the-art baselines by up to 15% in Recall@10 and NDCG@10 across three benchmark datasets (MovieLens-1M, Amazon, and Yelp), showing greater robustness in cold-start and high-sparsity situations. The implementation demonstrates effective GPU scalability, affirming the potential of the proposed framework as a robust and computationally feasible solution for real-world intelligent recommendation systems.