Research On Recommendation Method Of Incremental Models Blending And Cascaded Filtering

With the rapid development of the Internet, the amount of information and the number of users are also growing at a rapid rate, so our human have stepped into the age of information overload, the recommendation information filtering technology have come necessarily. Recently, recommendation technology has become a hot research topic in the academic field, and it has been widely used in industry. However, problems of recommendation technology has also been gradually exposed and enlarged, especially the mainstream recommendation algorithm could not identify invalid samples. The generalization ability of the recommendation algorithm model is limited. Its real-time performance is too poor to fit the user’s recent interactive information and the ability to reflect the weak features is all so limited which will seriously affect the further application and promotion of the recommendation system.To solve these problems mentioned above, an recommendation method based on cascade filtering and incremental model blending is proposed. It’s shown as follows:(1) In view of the current mainstream recommendation algorithm can not identify invalid samples, this paper puts forward the method of cascade filter samples. Firstly, a user-item preference model is constructed on the sample set. We will use logistic regression to fit the user interest and return the sample processing to maximize the heads of weak positive samples to enhance degree of filtration. Then the paper will use the secondary logic regression model to filter out the sample with the dummy variable and the unique thermal coding system to find the maximum positive sample and use the cascade model which can effectively filter out the noise samples, outliers and weak contribution rate to provide the subsequent model for more effective sample.(2) Most of the current recommendation algorithms are single model algorithm, which is easy to cause over-fitting, especially, when the time distribution of training set and prediction set are different,it shows poor generalization ability. To solve this problem, Multi-model blending method is presented in this paper, which mainly takes the recommendation problem as binary classification problem and regression problem, so using Random Forests based on bagging and Gradient Boosting Regression Trees based on boosting to fit them respectively. Due to these two algorithms are based on tree model, Logistic Regression is also introduced to obtain a better blending learning result. Then training several models of algorithms above and using Logistic Regression algorithm to blend each sub score prediction models.(3) The poor real-time performance of recommendation system hinders the fitting of users’ recent interaction information. To solve this problem, we design a framework based on online incremental model blending of Random Forests. Due to the new data produced continuously, new models continue to increase, and new model of relative displacement of the new model will become the old model, the constant circulation, enabling the model to growing, forming a complete non Markov chain.To validate the effectiveness of this paper’s method, six experiments have been done. The experimental results demonstrate this method not only filter out outlier samples, noise samples and weak-contribution samples, also overcomes over-fitting and weak ability of generalization, what is more, it shows a better real-time to fit users’recent interactive data. So this method enhances the recommendation precision and has certain practical value.