Hybrid Method of Recommender System to Decrement Cold Start and Sparse Data Issues

Background and Objectives: The primary purpose of recommender systems is to estimate the users desires and provide a predicted list of items based on relevant data. Recommender systems that suggest items to users face two cold start and sparse data challenges. Methods: This paper aims to propose a novel method to overcome such challenges in recommender systems. Singular value decomposition is a popular method to reduce sparse data in recommender systems by reducing dimensions. However, the basic singular value decomposition can only extract those feature vectors of users and items that may be recommended with lower recommendation precisions. Notably, using the similarity criteria between entities can reduce cold start to resolve the singular value decomposition problem by extracting more refined factor vectors. Besides, considering the context s dimensions as the third dimension of the matrix requires using another flexible algorithm, such as tensor factorization, which offers a viable solution to minimize the sparse data challenge. This study proposes TCSSVD, a novel method to resolve the challenges mentioned above in recommender systems. First, a two-level matrix is obtained using the similarity criteria between the user and the item to reduce the cold start challenge. In the second step, the contextual information is used by tensor in two-level singular value decomposition to reduce the challenge of sparse data. Results: For reviewing the proposed method, these two data sets, IMDB and STS, were used because of applying user and item features and contextual information. The RMSE criterion (95% accuracy) was used to investigate the predictions accuracy. However, since the user s rating of the item is particularly important in recommender systems, compared with other methods, such as tensor factorization, HOSVD, BPR, and CTLSVD, the TCSSVD method uses the following criteria: Precision, Recall, F1-score, and NDCG. Conclusion: The findings indicated the positive effect of using the innovative similarity criteria on the extraction of user and item attributes to reduce the complications deriving from the cold start challenge. Also, the use of contextual information through the tensor in the TCSSVD method reduced the complications related to sparse data. The results improve the recommendation accuracy of the recommender systems.