Enhancing item-based collaborative filtering by users’ similarities injection and low-quality data handling

With the increasing amount of the commercial items (movies, music, books, cars, etc.) produced each day by companies, it becomes very difficult for customers to find the suitable products satisfying their needs. Generally, Recommendation Systems (RSs) were used to fit this necessary requirement by solving the problem of information overload, specially on the web. Indeed, RS are designed to provide relevant resources to a client using certain information about users and resources. To the best of our knowledge, RS remains providing modest performances in many domains. In this context, we proposed a new model named CSWMC that combines two different techniques: item-based and user-based Collaborative Filtering. In fact, our proposed algorithm starts with the estimation of the suitable number of the user’s neighbors that offers to the Recommender System the optimal efficiency. Then, the system integrates this knowledge about users in the ‘Mean Centered’ aggregation method. Also, we proposed a simple method for handling the cold start and data sparsity problems that used mean value of the training datasets. The proposed models were validated through an experimental study on three standards datasets and compared with six well-known models. The obtained results demonstrated that our proposed model (in its two versions: with and without cold start handling) outperforms all the other models in terms of three evaluation metrics: RMSE, MAE and R2.