Research On Time-Aware Recommendation Algorithms

With the rapid development of Internet and the popularity of smart phones,people can get online contents and services whenever and wherever possible,such as news,music,videos,restaurant positions and comments.These huge amounts of Internet information bring users convenience and simultaneously a new challenge of how to filter out the contents of interests.Recommender systems arise to solve this so called “Information Overload”problem.As the bridge connecting content consumers and providers,a recommender system not only helps consumers find the relevant items,but also help providers increase sales and user engagement.In the recent decade,recommender systems were widely employed in the major industrial fields,and receive considerable attentions from the research communities of computer science,physics and management.Most of early works are focused on static recommender systems,that is,building time-unaware recommendation models regardless of the occurring time of user activities.However,past study on human dynamics found that user behaviors are usually closely related with the time factor.For example,user interests will shift over time,there is a phenomenon of anchoring bias in users rating behaviors,and the sales of products usually suffer periodical fluctuations.In recent years,the effect of time factor in recommendation systems has attracted a lot of research attentions.In this thesis,we investigate the time-aware recommendation algorithms by analyzing the role of user dynamic behaviors,based on a large amount of user-item interaction records with time stamps.The major contributions of this thesis are listed as follow.(1)The time-aware algorithm based on user explicit ratings.User explicit ratings are usually categorized into positive ratings and negative ratings.Different types of ratings may have different correlations with users future preferences.We empirically discover the different temporal effect of positive and negative ratings,and propose the improved mass-diffusion algorithm by taking into account these temporal effects.This algorithm can significantly improve recommendation accuracy and solve the personalized problem in sparse data sets.In order to save the computation time on traversing the optimal parameters,we make use of linear regression model to fit the different decay factors of positive and negative ratings.This algorithm based on parameter-fitting produces superior recommendation results with lower computation complexity,thus will be of great business value in future.(2)The time-aware algorithm based on implicit feedback.The choosing activities of users may be affected by other people who previously selected the same item,and could also affect the subsequent users.Given a specific target user,we categorize other users into two groups,leaders and followers,according to their shopping sequences.The empirical study shows that the shopping lists of followers are more similar to each other,and those of leaders are more personalized.In this thesis,we propose a time-aware diffusionbased recommendation algorithm,which assigns different time-decay weights to different user groups.Experiment results show that our algorithm achieves higher recommendation accuracy,compared with existing time-unaware recommendation models.(3)The recommendation algorithm based on recent popularity.The item popularity in the existing recommendation models is usually defined by its total purchased times in the whole period of observation,named total popularity of items.Our empirical study finds that the recent popularity of items,defined by the recently purchased times,plays an important role in recommender systems.In this thesis,by taking into account this recent popularity,we improve the heat conduction recommendation algorithms,and propose a resorting method for the recommendation results of mass diffusion model.Experiment results show that,these two recent-popularity based algorithms can achieve good accuracy,and improve the personalized measure in sparse data sets.(4)The recommendation algorithm based on relative time.In the scenarios without absolute time stamps,the relative time of user-item interaction records is also useful for performance improvement of recommender systems.Based on the collecting sequence of the same items by different users,we quantify the influence among users and accordingly assign weights on the edges of user-item bipartite graph,thus propose a user sequence aware recommendation algorithm.What is more,we model each user as a vector,in which the element is the probability of a permutation on his collected items.The similarity between two users is measured by the Jensen-shannon divergence.Based on this,we proposed an item sequence aware recommendation algorithm.Experiment results demonstrate that these two methods can both improve the diversity,novelty and coverage while keeping the high accuracy.