Personalization using machine learning methods in recommender systems has received a lot of attention in recent years. In practical issues, the data related to the transaction of users and items is usually sparse. Therefore, the performance of recommender systems could be improper in the beginning, which will cause customer dissatisfaction and financial losses. This problem is called cold-start and various methods have been proposed to solve this problem in learning models. One of the most promising methods that has been considered in recent years to solve this problem is considering ideas based on meta learning, which tries to expand what has been learned from transactions and use it to predict in future. The proposed methods to solve this problem have weaknesses that have caused insufficient efficiency in some cases.
In this research, we intend to investigate the problem of cold start in personalized deep recommender networks, to present a new solution based on meta learning to improve the performance of these neural networks. Improvements in these networks would be examined in terms of increasing the accuracy of the model, reducing size of data required for proper recommendation and improving privacy at various levels of the system.