The successful use of recommendation systems involves web applications (e-commerce, including Amazon), streaming media applications (including Netflix), and social media applications (including Facebook). The collaborative filtering techniques, however, have 2 areas that restrict their performance. The paucity of the data is one of the largest problems. In addition, more than 90 percent of cases lack details about users' actions on the item. The cold-start problem arises when the information about a new user or item is very limited. Although more recent approaches, which use association rule mining coupled with matrix factorization to achieve their goals, are promising, it is difficult to find the optimal compromise between cost and prediction accuracy. The new combination recommendation model suggested by the authors in this research proposal is based on Probabilistic Data Augmentation, which is the most common pattern mining method (including Apriori and FP-Growth), and on latent factor models (including SVD, SVD++, and NMF). A naive Bayes classifier that is confidence-sensitive gives ratings and confidence values. Confidence levels of 0.8 or higher are rated 5. A higher sparsity of 92.12-93.69 was achieved when the top quarter of the generational data was used as a sparse estimator. Experiments on the MovieLens-100K data have shown that the Apriori-SVD fusion model, with an augmentation proportion of 25%, is better than the baseline hybrid model (F1-score = 0.896). The disaggregation of the efficiencies indicates that NMF is extremely high, with an F1-score of 0.889 and a much higher cost (> 3.5 hrs), whereas the normal SVD requires about 14.1 secs to execute the run. Lastly, the paper will provide useful recommendations for developing high-performance, fast recommendation systems with limited resources while ensuring high quality in the real world.