Research On Recommendation Algorithms Based On Temporal Dynamic Characteristics

The traditional recommendation algorithms mainly researches the binary relationship between users and items.And they always recommend the items that meet users' preferences or predict the users' interest in items.But these algorithms ignore the user-related context which has impact on recommendation,and time is the most important context.User behavior based on time context reflects the law of changes about user interest.And considering the user behavior sequence,recommendation algorithm can achieve a better effect.Based on the temporal dynamic characteristic,we study the impact of time-varying on recommendation algorithms,and carry out the following work.First,we study the characteristics of user's short-term interest fading over time and the time effects of user's cycle behavior.In the recommendation system,time dynamically affects user interests and item lifecycles,and user behavior and item consumption continue to change over time.In addition,time is periodic,such as weekends,holidays,seasonal variation have a great affect to user' behavior and needs.According the different modes,time context can be modeled as continuous value and discrete value.In the neighborhood-based recommendation algorithm,considering the characteristics of short-term interest fading over time,we propose a novel neighborhood recommendation method based on temporal dynamics.In this method,we regard time as a continuous value to reflect the continually change of user interest.For simulating the process of user interest changing over time,we adopt “Newton's law of cooling” as a time decay function to enhance the weight of user recent interest and weaken the long-term interest.In addition,when calculating the user similarity,the Jac UOD method is introduced,which properly handles dimension-number difference for different vector spaces.Finally,we selected several algorithms for comparison experiments on the Movie Lens dataset and Netflix dataset.The experimental results show that our algorithm has a better performance on Top N recommendation.The scoring prediction problem is the focus of recommendation algorithm research,and Matrix Factorization Model is the most commonly method to solve this problem.The 3-dimensional tensor factorization,which considers the period of time,is difficult to apply in practice due to the high complexity.In this regard,a Probability Matrix Factorization Model based on improved similarity measure is proposed in this thesis.This model not only considers the periodic effects of time,but also maintains the matrix factorization in two dimensions.For the problem of score prediction,our model regard time as discrete value with limited values,reflecting the periodicity of user behavior.Then taking time and categories of items into the recommendation process,we define and construct a new usercontext rating matrix.This matrix represents the degree of user interest in the discrete time value and the item category.Next we build a novel context-dependent similarity matrix by mining the implicit feature vector about user from the user-context rating matrix.Finally,we implement the improved model by taking the similarity matrix into the Probability Matrix Factorization Model.The comparison experiments on Movie Lens and Movie Tweetings dataset show that our method can improve the accuracy of scoring prediction.