The Technology Of User Novelty Based Personalized Recommendation

Due to its huge potential economic value,personalized recommendation has attracted a lot of attention from both research and industrial communities,and it is everywhere in people's lives.As an important quantitative measurement of the openness in people's traits of characters,the novelty of user determines her willingness to explore the items of unfamiliar interest,and further influences people's preference for items.Thus,the study of personalized recommendation relies on the investigation of user novelty which facilitates the prediction of user interest and improves the effectiveness of recommender systems.This study starts with the discriminative analysis of user novelty,based on which it develops different personalized recommendation methods that meet users' temporal changes in preference.Then,it investigates the more difficult problem of modeling users'intransitive preference with respect to multiple angles of user novelty.The primary contributions of this study are summarized below:(1)The problem of the discriminative analysis of user novelty is formulated for the first time.Four generic temporal behavioral features including average normalized item popularity,average normalized item reconsumption ratio,user reconsumption ratio and window repeat ratio,are extracted based on user's item selection history.The relationship between these behavioral features and user novelty is evaluated on four real-world datasets.Besides,two fast methods are proposed to predict user novelty in real time.(2)The problem of recommendation for repeat consumption is formally defined for the first time,which is barely studied before.A time-sensitive personalized pairwise ranking(TS-PPR)model is brought forward,which factorizes the temporal interactions between users and items.Four temporal behavioral features including normalized item quality,item reconsumption ratio,recency feature and dynamic familiarity,are extracted and further combined with TS-PPR to recommend repeat items.A negative sampling method is also proposed to decrease the time and space cost in training.The evaluation on real-world datasets shows encouraging results.(3)This study incorporates the features of human memory into recommender system for the first time.A personalized framework that consists of the process of interest forgetting and interest accumulation,as well as the variable-order Markov model,is brought forward to comprehensively model the dynamic change of user interest.The effectiveness of the proposed model is evaluated on a real-world music listening dataset.(4)This work investigates the more difficult problem of modeling personalized intransitive preference.This is the first work to explore the causes of users' personalized intransitive preference.The multi-angle preference series models are proposed,which consist of five different sub-models based on the definition of preference angles.A new dataset of users' pairwise preference comparisons on images is constructed in this work.The evaluation is conducted on this new dataset as well as another two real-world datasets.