Recommendation System with a Pre-Training Stage Based on Neural Networks for Event Sequence Processing

Recommendation systems are now playing a vital role in modern business due to their ability to increase customer engagement and sales. However, most of the existing recommendation models handle only the history of user behavior, neglecting such an essential factor as the interrelationships between them. The study developed a deep learning model capable of processing the sequence of events and taking into account the interrelationships between users to predict their subsequent actions. Various pre-training algorithms have been developed for different tasks to help extract useful information from spatial-temporal data on user behavior and their social relationships. The pre-training phase was added to the model training algorithm for the target task of predicting the user's next action. Metrics describing the performance of recommendation systems have shown that this approach improves recommendation quality. The proposed method was compared with existing methods, and it was shown to be superior to existing metrics solutions with a statistically significant difference. The recommendation system was developed and tested on the basis of the best algorithm. This work consists of 79 pages, 4 chapters, 25 figures, 34 tables, 8 formulas and 2 appendix. 22 sources were used. Keywords: recommendation system; deep learning; event sequence; pre-training stage; representation learning; social graph; spatial-temporal data; embedding