Matrix Decomposition Method Merged With Rank Model In Recommender System And Research Of Related Problems

With the rapid development of Internet industry, the Internet has become the tool of human development. In work and life, the Internet is changing the way people live, provides people with more convenient and comfortable service. But with the rapid development of Internet,more and more problems are also beginning to emerge. Among them,the information overload problem is new problem along with the development of the Internet. The vigorous development of the Internet has created fast and convenient living and working environment. But also because of the explosive growth of information, it has brought people the misery of filtering information. It has become a major challenge in the era of information explosion to figure out how to filter the large,complicated and disordered information and recommend the most interested information or item to the users.The traditional search engine, although to a certain extent,can solve the problem of information screening, but it can only be matched to the user's explicit demand and provide the indiscriminate information search service for users, it is unable to mining potential interests and hobbies and provide personalized information recommendation. Under the background of the era of big data and the requirements of personalized information, recommendation system arises at the historic moment.At the same time, in recent years, social networks and audio and video service provider on the Internet is developing at a surprising speed,Facebook, Twitter, sina microblog, Netflix and Youtube application has become an important part of life and entertainment for a lot of people.People left many user behavior records when they browse microblog,chat with friends, watch movies. This information will promote the recommendation system of algorithm research, and the research achievements of recommendation algorithm also greatly improved the service to the user and the user experience.In this dissertation ,we focus on some key problems in recommendation system for research, mainly includes: for SVD++model how to select the appropriate implicit feedback information; For recommendation system itself can be understood as a top - N problem and how to design reasonable sorting model. On the basis of summarizing the most mainstream models of recommendation algorithm,this paper mainly makes some research achievements:(1) At present, in the field of recommender systems SVD++ model is one of the highest accuracy algorithm of single model. In this paper,combined with the classical algorithm principle and the characteristic and meaning of the specific data, we made a reasonable design and implementation of the model. On the selection of implicit feedback,made a reasonable choice, to achieve the ideal effect model. This article will use the SVD++ model as the one of a basic model of the final model(2) In the recommendation system we care about more of whether users may accept the recommendation or not rather than a specific user rating prediction, we can make a certain degree of compromise,especially for the items that user didn't accept. Recommendation system,therefore, to a certain extent, can be understood as a top - N application,while the recommendation algorithm can be understood as a Learning to Rank (LTR) problem. In order to solve this problem, this paper combined with the specific data, and the most commonly used two main methods in LTR PairWise and ListWise method, designed a gradient PairWise model, the model combine the advantage of List Wise method and PairWise method in a certain degree.(3) As a branch of machine learning, ensemble learning is more and more hot in various applications. In this paper we compared the most popular ensembles learning thoughts and decide to use the method of secondary learning model. Finally we use Logistic Regression (LR,Logistic Regression) to ensemble the SVD++ model and gradient PairWise model, and achieved good result.