A proposed memory-based collaborative filtering technique based on a new similarity and MADM methods (CF-NSMA) for improving the recommendation accuracy

The collaborative filtering (CF), as one of the most widely used and most successful approaches to provide service of recommendations, provides users with a set of recommendations related to what they need (their interests). These recommendations will be generated based on the correlation among the users’ preferences such as ratings and behaviour. Nevertheless, the number of users and items available on the Internet has increased dramatically, and most of the users do not give enough ratings for the items. Moreover, this vast growth has made the user-item rating matrix very large and sparse. This is considered a problem in the current traditional memory-based CF recommender system because the similarity calculation process between users/items becomes very difficult or may lead to locating unsuccessful neighbours which in turn to a weak recommendation. Therefore, formulating a right similarity method to identify the successful neighborhoods is a one key of memory-based CF. Similarly, the prediction method has the same level of importance in the process of improving the CF accuracy. Unfortunately, most studies on improving the accuracy of conventional CF systems have focused solely on enhancing the similarity measure. In contrast, improving the prediction method has been somewhat neglected. Consequently, the prediction method is still an open area for improvement to get better candidate items ranking and in turn increase the accuracy of CF. In the prediction process, the system predicts a user score for each item in the candidate set and promotes the highest-rated items as recommendations. This process of evaluating and ranking candidate items is therefore quite significant to the performance accuracy of the CF. Therefore, in this work, a new memory-based Collaborative Filtering (CF) technique is proposed to address the issue of sparsity data and improve the accuracy of recommendations, it is called CF-NSMA technique. The proposed technique consists of three main steps: 1- Constructing a new normalized matrix to overcome the sparsity issue; 2- Formulating a new similarity measure, based on adopting the fairness and the proportion of common rating factors to locate the accurate neighbours; 3- Applying the MADM method to get better evaluating and ranking list of candidate items. These phases were carefully designed and implemented to solve the issues that were mentioned earlier. Moreover, to assess the accuracy of CF-NSMA technique, several experiments were conducted using a public dataset (MovieLens 100K, MovieLens 1M benchmark datasets). The evaluation process was performed to measure the accuracy of the proposed technique using Mean Absolute Error (MAE) to measure the prediction accuracy and Precision, Recall and F-measure to measure the performance accuracy. These selected metrics are considered as the most common metrics to be used in an accuracy evaluation process of the CF techniques. The result of the experiments revealed that the accuracy of the proposed technique is better compared to the common base memory-based CF methods. The prediction accuracy percentage in terms of MAE was around 0.76 and 0.74 via 100K and 1M datasets, respectively. While, the improvement of the CF-NSMA technique in terms of performance accuracy was around more than three-fold in term precision, around four-fold in term of recall, and around three-fold in term of F-measure. In conclusion, this work contributes significantly to the field of improving the accuracy of memory-based CF by developing the critical phases of traditional memory-based CF, including re-representing the rating matrix, formulating a new similarity method and replacing the prediction method with the MADM method. Furthermore, MADM successfully minimizes the negative effect of the prediction method in evaluating and ranking the candidate items and significantly improves the accuracy of memory-based CF. Therefore, the primary objectives of this research were achieved