Research On Tensor Factorization-based Recommendation Algorithm

With the rapid development of information technology, people came to the era of information overload from the era of lacking information. It is difficult for users to find their own interest or valuable information from the massive data. Recommendation system is an effective way to solve the problem of information overload. The model based recommendation algorithm is the state-of-art recommendation method with high recommendation precision, excellent at discovering new interests and scalability.However, the model based recommendation algorithm has some problems. First, these algorithms neglect temporal influence of recommendation system. Traditional recommendation models are stationary with neglecting time factor. Some recommendation algorithms take time factor into consideration, but what they do is using latest data or reducing the weight of past data. They may lose some useful information. Second, the goal of optimization does not agree with the goal of recommendation system. The ultimate aim of recommendation system is recommending list of items to users. Traditional recommendation algorithms predict the user’rating for item firstly and then generate recommendation list of items based on the getting rating. That means the focus of these recommendation algorithms is predicting rating. However, these algorithms with high precision of rating prediction may not give perfect recommendation list of items.To solve these two problems, we propose two improved algorithms on the basis of existing research work. To solve the first problem, we put forward a time-based local low-rank tensor factorization algorithm. The algorithm take time factor into consideration and view rating matrix as 3-dimensional sensor based on the traditional recommendation algorithms which extend the traditional algorithms to tensor field. To solve the second problem, we proposed an improved time-based local low-rank tensor factorization algorithm on ranking recommendation task. Firstly, our algorithm optimizes Mean Reciprocal Rank directly. Secondly, our algorithm modifies the evaluation metric to fit explicit feedback data for making full use of the information in datasets. Thirdly, our algorithm smooths the modified evaluation metric for using standard optimization methods to optimize the evaluation metric directly. Fourthly, our algorithm derives the lower bound of the smoothed evaluation metric for optimizing it tractably. Fifthly, our algorithm takes the evaluation metric as objection function and uses stochastic gradient ascent to optimize it. Experiments show that these two algorithms could improve the efficiency of ranking recommendation.