Dynamic User Interest Modeling And Research On Recommendation Algorithm

With the coming of information age,the information people may contact has an explosive growth.How to overcome the "information overload" problem to find satisfied products from vast amounts of information has become urgent in the field of electronic commerce.Personalized recommendation system can quickly provide users optional recommendation products that users may prefer.In the personalized recommendation system,user interest modeling plays an important role,which is the main knowledge source of the personalized recommendation.Thus to get the user's interest accurately,is the foundation of good recommendation.In this dissertation,we study the dynamic changes of user interest,through the construction of dynamic communities that users belong to,the simulated interest forgetting process,and the process of learning knowledge,to depict the user's changeable interests.The main contributions of this dissertation are summarized as follows:Firstly,according to the phenomenon that users may have multiple interests and the interest may changeable with time,a personalized recommendation model is proposed based on users' dynamic temporal interests and multi-memberships in their overlapping communities.In this model,a temporal method to detect overlapping communities is developed based on a user-user graph with time-weighted links,and a new time-weighted association rule mining algorithm is specially designed based on temporal overlapping communities to model user interest drift over time.The experimental results show that the proposed approach outperforms several existing methods in recommendation precision and diversity.Secondly,the nature of the user interest evolution is interpreted by using the psychological theory of “Yerkes-Dodson law” and the individual absorptive capacity in adoption theory.Then the evolution curve is depicted and the interest evolution function is determined by the weighted least squares regression method.A recommendation algorithm is proposed according to the interest evolution function.Experimental results show that the new interest evolution-based recommendation algorithm shows better recommendation accuracy,higher time efficiency and higher interpretability than some conventional static algorithms and dynamic algorithms.Finally,a hybrid personalized recommendation method is developed by considering both user interest evolution and interest drift.The simulation of user interest evolution is executed.Then the interest drift simulation is made according to a modified activated-decayed interest function.The hybrid dynamic recommendation algorithm uses these two parts to represent the score residual space in order to capture user's dynamic interest.The experimental results show that the proposed hybrid dynamic recommendation algorithm has performed better in the accuracy than other dynamic algorithms.