In recommender systems, determining user preferences to suggest relevant products is key. Context-aware recommendation systems have gained traction for providing more accurate, personalized suggestions than traditional methods. However, they face challenges such as limited data and the cold start problem. Traditional methods, like singular value decomposition (SVD), often yield lower precision due to their limited ability to extract user- and item-feature vectors. To address these issues, a novel context-aware recommendation algorithm called CSSVD (context similarity singular value decomposition) is proposed. CSSVD enhances recommendation performance by tackling sparse data and cold starts. It uses two similarity criteria: IFPCC (item-feature Pearson correlation coefficient) and DPCC (demographic Pearson correlation coefficient) to form the SSVD matrix, addressing the cold start issue. A context matrix is built using the CWP (context weight performance) criterion to resolve sparse data challenges. These matrices create a 3D matrix by leveraging tensor characteristics. The algorithm is tested using the Stanford Sentiment Treebank (STS) dataset, incorporating user features, object features, and contextual data. Experimental results demonstrate that CSSVD outperforms other algorithms, such as TF, HOSVD, BPR, and CTLSVD, in precision, recall, F1-score, and NDCG measures, showing significant improvements in recommendations by resolving cold start and sparse data problems.