Improving the Performance of Collaborative Filtering Using Outlier Labeling, Clustering, and Association Rule Mining

Collaborative Filtering (CF) is a popular recommendation method because it can provide recommendations personally. Under conditions of data sparsity, CF recommendation systems are known to have low accuracy because available historical information is not enough to properly identify preferences. Based on experiments conducted by researchers, the factor that limits the accuracy of the CF recommendation system is the number of recommendation items that exceed the system requirements. The existence of outliers with too much items also affect the recommendation results. This study attempted to apply model-based CFs to improve the accuracy of CF recommendations under data sparsity conditions. Conduct outlier labeling, clustering, and association rule mining implemented from preprocessing as a combination of data processing methods to generate recommendation items. Experiments conducted on Groceries dataset, a real-world point-of-sale transactions data from grocery outlet. The results of the evaluation indicate that the proposed method can improve the accuracy of basic CF by 26% with the quality improvement of the recommendation result by 24%.