A Multi-task Learning Model Based On Matrix Factorization On Human Behavior Data

With the rapid development of information technology and communication technology, every day there are vast amounts of human behavior data produced, such as Weibo were sent on Sina, applications were downloaded on Apple APP, goods were sell on Taobao, network access occurred on Renren, search queries generated on Baidu, and so on. We can predict human behavior by the data analysis, but how to find business opportunities in these human behavior data? How to recommend personalized service to users? Mining technology is based on an human behavior data need for in-depth study. Data mining technology can dig out the hidden potential of knowledge model for people to acquire the necessary knowledge and information to assist decision making.Matrix factorization model is a method based on hidden feature vectors, it can effectively utilize massive data user interaction information. In practice, each task is not independent and identically distributed, multi-task learning can learn shared structure between a plurality of tasks and the differences between each task, well enhance the good performance of mining technology. In view of this, this thesis proposed mining method based on human behavior data, nicely combines the advantages of matrix factorization techniques and multi-tasking learning technologies for data processing vast amounts of information-based humna behaviors.The main work includes the following three aspects:Firstly, the dissertation proposes linear mining method MF-MTL model based on human behavior data. This model combines statistic characteristic user owned and hidden variable characteristics which got from the decomposition of dynamic human behavior data matrix. Then combines the advantages of multi-task learning technology it create prediction objective function. On the normal condition, the human behavior data are sparse and high dimension. Base on personalized recommendation algorithm of this data, the effect of matrix factorization model is the most outstanding. Therefore, the model in this dissertation utilizes the superiority of matrix factorization and multi-task learning. At the same time, it combines user profile data and behavior data on the aspect of data. It absorbs their advantages and improves the performance of model.Secondly, the relation of data is not simple linear relationship in the practice application. Using the non-linear transformation of inner product kernel function defined map the input space to a high-dimensional space, then resolve problem in the input space. It extends MF_MTL model to a wider range of non-liner areas. The dissertation uses experience minimized principles and addition regularization term avoid fitting phenomenon.Thirdly, the dissertation use least squares technique create MF-MTL model. Through optimized Lagrangian multiplier method, the optimization problem with equality constraints is transformed to the resolve problem of linear system. Compare with traditional optimization algorithm with equality constraints, the optimization algorithm based on alternating least squares reduces computation complexity on a certain extent.