ICE-DEN: A click-through rate prediction method based on interest contribution extraction of dynamic attention intensity

Abstract

The goal of click-through rate prediction is to predict the probability of users clicking on items, and the improvement of its performance is still one of the core tasks in e-commerce. With the development of online social networks, users have access to a large amount of historical behavior data. The use of historical data for extracting user interests is crucial to the improvement of the click-through rate prediction performance. However, most existing methods ignore the importance of dynamic attention intensity. Therefore, this paper proposes an interest enhancement method based on dynamic relationships called Interest Contribution Extraction (ICE). This method first obtains the dynamic interest intensity by considering the dynamic changes in interest. Then, an adaptive interest attention unit is constructed to collect users’ different interest and attention strengths for different candidate items. Based on the ICE method, a new click-through rate prediction model called the Deep Extraction Network based on Interest Contribution Extraction (ICE-DEN) is proposed. The model obtains a user’s low-dimensional features through the embedding layer. Mini-batch perception regularization and the Dice activation function are used to help train deep learning networks with a large number of parameters. The ICE-DEN model has achieved better prediction results than other models on three real datasets.