Research On Exploiting Multidimensional Domain-specific Features For News Recommendation Models

It is necessary daily activity for most users to read news.Readers can understand quickly what happened outside.In the past decade,with the rapid growth of network technology,more and more people tend to read online news.The large number of digital news articles from websites are accumulated,and the readers are submerged in the ocean of information.It is difficult to find quickly interesting articles from the enormous amount of news,which may lead to heavy news information overload problem.Fortunately,personalized news recommender systems are developing rapidly towards resolving news in-formation overload problem.Personalized news recommendation can recom-mend different news to different users according to their reading preferences.Although personalized news recommender systems have been studied widely,most existing studies do not exploit domain-specific features resided in news domain to improve news recommendation effectiveness and remit the extreme data sparsity in news domain.The main contributions in the thesis are summa-rized as follows:(1)The hybrid recommendation model study based on category feature and user behavior sequence feature.We first find that category feature is very helpful to model user preferences.We use category feature(i.e.,general news and special news have different life cycles,and they play different roles in user preferences)and other features to capture user preferences more accurately.As a result,we improve the traditional content-based news recommendation.Then,we propose an effective similarity computation strategy to compute the similar-ity between users.In more detail,based on user behavior sequence feature,we use sequence feature vectors to compute the similarity between users.Accord-ingly,the traditional collaborative filtering news recommendation is improved.Finally,we propose a hybrid news recommendation model fusing multidimen-sional features,and the hybrid model incorporates the advantages of content-based and collaborative filtering techniques.The experimental results on real news datasets demonstrate our proposed hybrid model outperforms the state-of-the-art models.(2)News recommendation model study based on life cycle feature.In the above,we notice that special news and general news have different life cycles,and special news reflect better user preferences than general news.Further,through analyzing news datasets,we discover that different news topics have different life cycles.This is very helpful to improve news recommendation,especially in the situation of data sparsity.Based on our findings,we propose life cycle-aware topic model to incorporate the influence of life cycle feature.Each life cycle is denoted by Poisson distribution over time interval.Mean-while,we also utilize worth feature to further relieve the data sparsity and cold start issues.The experiments on news datasets show the improvement of our proposed model over existing models in terms of recall and nDCG.(3)News recommendation model adapting the dynamic change of reader consumption behaviors.Through using statistics significance testing method,we investigate the dynamic change of reader consumption behaviors on real news datasets.We find that reader consumption behaviors are influenced by other factors such as breaking news in addition to user interests.Based on this conclusion,we propose a topic model to adapt the dynamic change of reader consumption behaviors.User interests and crowd effects are used to adapt the dynamic change of consumption behaviors,and the reading sequences are em-ployed to adapt the evolving of user interests.Moreover,in order to remit the extreme sparsity in news recommendation,we capture reading sequences on a group level instead of on a personal level.The experiments show that the proposed model outperforms significantly baselines.(4)Latent semantic model based on group sequence.In the third study,the group division strategy is coarse,which degrades the accuracy of group division.More importantly,it is difficult to find the representative sequences of each group.To address these problems,we propose a group-sequence-aware topic model,where each group is regarded as a potential variable,and it can be learned automatically and represented by a multinomial distribution over users.Likewise,each group's sequence pattern is also viewed as a potential variable,and it can be learned automatically and denoted by a multinomial distribution over individual reading sequences.The experiments with significance testing demonstrate the superiority of the proposed model.