Research On Tensor Theory-Based Mobile Recommender System

In this age of information overload, people use a variety of strategies to make choices about what to buy, how to spend their leisure time. As a tool, Recommender systems automate these strategies with the goal of providing personal and high-quality recommendations.Context has been recognized as an important factor to consider in personalized Recommender Systems. However, most model-based collaborative filtering approaches such as matrix factorization do not provide a simple and effective way of integrating context information into the model. In this paper, we will introduce a collaborative filtering method based on tensor factorization, which is an extension of matrix factorization, and model the data to form a User-Item-Context N-dimensional tensor instead of the traditional User-Item2-dimensional matrix, so it can be more flexible to integrate context information. In tins paper, in the proposed model, which is called tensor decomposition-based multiverse recommendation, different types of context will be considered as an additional dimension in the presentation of data as a tensor. In this paper, tensor factorization model uses a compact data model that can be used to provide context-aware recommendations.In addition, before explain the details of tensor factorization-based model presented in this paper, we also do some other work.First, briefly descript the basic concepts of recommendation systems and several basic categories of traditional recommendation system:collaborative filtering recommendation, content-based recommendation, knowledge-based recommendation and hybrid recommendation method. Introduce the various application areas and some of the challenges and difficulties of the recommendation system. Describes some of the development in the mobile recommendation system field, and provide a detailed explanation of the basis algorithm for the proposed tensor factorization model——matrix factorization. And also introduce several improved matrix factorization algorithm:singular value decomposition, non-negative matrix factorization and margin-maximum matrix factorization.Second, provide a detailed description of the core algorithm of this paper-tensor factorization algorithm, include how we model the given data, propose the loss function, optimize the parameters in our model and also gives the pseudo-code of the algorithm to solve the given model.Finally, operation the simulation experiment on tensor factorization model. We used three data sets:semi-synthetic Yahoo WebScope dataset, real-world Adorn movie data set and a food menu dataset. In the experiment, we also used three other recommendation algorithm in which we compare the result to our model and then analyzed the result.According to results of experiment, the algorithm we provide can solve the problem of the decomposition of the N-dimensional matrix, and the result show that our model improves upon non-contextual matrix factorization up to30%in terms of the Mean Absolute Error (MAE). We also compared our method with other two context-aware methods. The result showed that tensor factorization consistently outperforms them both in semi-synthetic and real-world data and improvements range at least2.5%depending on the dataset. It is worth noting that, whenever there is more contextual information is available, our method will improve in a greater range than other methods.According to above description, our model has the following features:●Able to extend matrix factorization to the N-dimensional case——tensor factorization.●Able to include any number of contextual dimension into the model.●Using a variety of functions and a plurality of parameters for continuous fine-tuning in order to achieve better results.●Use relatively fast and simple algorithm to train the model.●Take advantages of the sparse data, while exploit the interaction between all users-items and context.