Dual Sparse Variational Autoencoder-driven Interactive Estimation of Distribution Algorithm with User Generated Contents

Personalized search and recommendation tasks in a big data environment have attracted wide attention from researchers while also presenting significant challenges.This paper proposed a dual sparse variational autoencoder-driven interactive estimation of distribution algorithm with user generated contents, which considers digging deeper into user generated contents and combining with intelligent evolutionary optimization methods. Firstly, user generated contents in web are extensively collected to build a user interest preference model based on dual sparse variational autoencoders to extract users' preference features. Secondly, the distribution probability is reconstructed according to the neurons of the output layer of the sparse variational autoencoders in the user interest preference model. Combining with Gaussian probability density function, a estimation of distribution sampling probability model based on the user interest preferences is designed in the framework of an interactive estimation of distribution algorithm to generate new individuals with the user preference information. Meanwhile, a user evaluation surrogate model based on the user interest preferences is established by integrating matrix factorizations to participate in a personalized evolutionary search process. Finally, according to new user generated contents and model management mechanisms, the user interest preference model and related models are dynamically updated to guide the interactive personalized evolutionary search for personalized recommendation task. Extensive experiments on various actual public datasets are executed to validate the feasibility, effectiveness and interpretability of the proposed algorithm. It improves the performance of personalized search and recommendation methods.