Mining Users Rely on E-Commerce

Abstract

In recent years, photo-based social media has become one of the most common social media platforms. Understanding user preferences in user-generated images and making suggestions has become a major necessity due to the large number of images uploaded daily. Several types of hybrids have been suggested to improve the performance of the recommendations by combining different types of third-party information (e.g., image representation, interaction) with user object history. Previous research, however, has failed to incorporate complex factors that affect user preferences into the corresponding framework due to various image features created by users on social media. In addition, many of these hybrid models have used pre-defined weights to combine different types of data, resulting in less favorable performance. To this end, we present a consistent model for capturing public imagery in this paper. We define three key elements (i.e., upload history, social exposure, and proprietary information) that affect each user's preferences, where each item summarizes the content aspect from complex interactions between users and images, in addition to the basic matrix interest model matrix factorization proposal. After that, we create a consecutive natural attention network that demonstrates a consistent relationship between hidden user interests and known key elements (elements at each level and feature level). A sequential attention network will learn to pay attention to more or less content using embedding from higher learning models designed for each type of data. Finally, the availability of extensive tests on real-world information indicates that our proposed model is superior.