Research On Mobile Context-aware Recommender System Based On Factorization

With the high-speed development of information technology and Internet, the modern people have already moved forward to the period of information overload from the period of information lack. Scale of tens of thousands of users and vast amounts of information makes users and information providers facing a huge test. It is a very difficult thing to find the valuable resources in a huge amounts of information for information users or consumers. On the other hand it is also difficult that how the information providers make their products stand out so that different users could get attention to these products. At this point, recommender system arises at this historic moment, it links the users and the products, making the most useful information to present in front of the users. So that the users and information providers could achieve a win-win situation, strengthening the trust to users, making providers profit.In recent years, recommender systems and their algorithms have very mature research all over the world. However, the rise of mobile communication network let more and more users start to use mobile devices(Smart phone, tables, etc.).Recommender system based on the context of mobile arises in this situation. Modern mobile smart devices could predict users’ preferences by collecting users’ browsing behavior and using personalized recommendation algorithm. Modern mobile smart devices can collect users’ context including the time, place, mood very easily and it’s very useful to improve the accuracy of recommender system. Therefore, it is the key to recommender system for collecting context information accurately and using these indicators to the algorithm.In this paper, the main work includes:The first chapter reviews and expounds the definition, structure, classical algorithm of recommendation system and research achievement at home and abroad. Classic collaborative filtering algorithms are introduced as well as excellent algorithms of machine learning and their application scenarios are summarized and compared.The second chapter expound the classical recommendation algorithm in the field of recommendation system, including user-based CF, Item-based CF, hybrid recommendation technology, recommendations based on the graph, as well as the influence of context to recommend results.The third chapter introduces the factorization theory in detail and explain the basic theories of machine learning and the classification of loss functions. Next it brings two high-performance recommendation algorithm:Multiverse Recommendation and Factorization Machines and introduces the history and theory of factorization model.The fourth chapter mainly carries on the empirical analysis, comparison of the optimization algorithms of Factorization Machines, analysis of time complexity, the influence of the dimensionality k and select a number of key features as context into the system. Moreover, it also compares the performance with the tensor decomposition algorithm Multiverse Recommendation. The empirical analysis is based on three public international data sets and it has proved that Factorization Machines is better than the traditional ones.This paper proposes two new factor models and mainly analyzes the Factorization Machine whose time complexity is linear which can provide a good performance to process large datasets. This algorithm has generality, which can mimic many different models by specifying the input data. Although the entire model still has a lot of improvements, but the Factorization Machine is one of the best algorithms in the recommended field.