Capturing Dynamic User Behavior in Recommender Systems by Finetuning

The amount of digital content is increasing significantly. Thus, it is important for companies to show the most relevant digital advertisements to the users to maximize their profits and i ncrease user experience. For this purpose, click through rate (CTR) estimation tasks are used, which track whether the user clicks on the displayed content. The fact that deep learning algorithms work successfully in many areas has inevitably been valid in recommender systems to predict clickthrough rate prediction also. Deep learning algorithms need large amounts of data to work well. Therefore, user data is gathered at specific i ntervals a nd t rains a re r un w ith t his d ata t o m odel user behaviors. Training models for high-traffic o r g anizations c a n be quite costly due to the high volume of data and because the data is collected over a long period of time, user interactions cannot be generalized well due to the dynamic user behavior. To overcome this problem, we present the application of finetuning f o r the click-through rate prediction domain to reduce the computational cost of model training and increase the generalization capability of deep learning-based recommender models. Consequently, the model will be able to capture the latest trends in user behaviors and recommend more relevant content to the users. The results show the superiority of training the model using finetuning compared to conventional batch training.