Analysis Of User Preference Modeling With Convolutional Neural Networks

With the rapid growth of data,recommendation system has become an important bridge connecting users and data.By analyzing user's historical behavior,recommendation system can effectively dig out user's preferences and complete various types of personalized recommendation tasks.Research in the field of personalized recommendation has greatly improved the user's ability to obtain information on the one hand,and has also brought huge commercial value on the other hand,which is of great significance to the promotion of the market economy and the development of the country.In addition,the relevant research results of the recommended fields come from multiple fields and multiple disciplines,which have promoted various related disciplines.While recommendation system is so important and hot,it also faces many challenges.With the rapid growth of Internet data,there is a lack of interactive data for a single user,which brings a serious data sparse problem,and it becomes very difficult to accurately and adequately characterize each user's static preferences.Secondly,we are currently in a dynamically changing world,with millions or even tens of millions of data being generated every second.It is difficult for traditional static recommendation systems to capture changes in data and changes in user preferences.Therefore,effectively dealing with the challenges brought about by data sparseness and at the same time aiming at dynamic data modeling is the focus of existing recommendation system research.In this paper,aiming at the problems of user preference pattern mining in recommendation scenarios,the author studies the user preference pattern mining methods from two aspects:static user preference pattern mining and dynamic user preference pattern mining.The main contributions are as follows:(1)static user preference pattern mining:The core of recommendation system is to accurately describe user's preferences.User's static preference refers to user's essential attributes.However,the extremely sparse user item interaction information brings a lot of challenges to the recommendation system.Starting from the consistency and complementarity between different source data,this paper proposes a deep recommendation model CMF based on the convolution matrix decomposition model.Specifically,for information in the two different fields of ratings and reviews,the author introduced a projection layer to dig out the unique and common parts of the two,so as to more accurately describe user preferences.Experimental results on three real data sets show that the CMF model can model sparse user behavior data and improve prediction accuracy.At the same time,thanks to the mining of more effective information,the CMF model also has excellent performance on long-tail data.(2)dynamic user preference pattern mining:The number and internal structure of users and data are changing over time.It is not advisable to use constant thinking to predict dynamic changes.This paper draws on the idea of distillation learning,trains a model on historical data and another on current data separately,proposes a deep incremental recommendation model DIR.Without the need of retraining all the data,the author mainly focused on the new data and the difference between history data and it.In addition,the author first proposed a way to measure the variation of data which is called data migration degree.A series of experiments on three real-world datasets have shown that DIR outperforms the existing popular incremental recommendation methods.