Research And Application On Collaborative Filtering Recommendation Algorithms Based On Multi-type User Feedback

With the rapid development of science and information technology,various innovative Internet applications have developed into an essential part of coexistence with human beings.Along with the explosive growth of the data and user scale all over the world,a new era of information overload is dawning upon the world now.However,it becomes increasingly difficult for retrieval useful and relevant information.Recommender systems(RSs)have been investigated for decades of years,and collaborative filtering(CF)is considered to be the most popular and successful technique,which has attracted more and more attention from both academia and industry.But under the circumstances of the richer data type,the bigger data size,and the complex application environment,there are still some new problems in traditional CF algorithms,for example,the similarity measurement,the assumption of pairwise preferences,the lacking of negative feedback and the uncertainty of the implicit feedback.With the above four aspects,this dissertation focuses on two kinds of collaborative filtering algorithms including the neighborhood-based collaborative filtering and the learning to rank with implicit feedback.The main research contents and contributions are listed as follows:1.A new collaborative filtering algorithm with the overlapping-dependency of multiple user behavior.The similarity computation is one of the most critical aspects of the neighborhood-based collaborative filtering,as it has a significant impact on both the performance and the quality of recommender systems.Traditional similarity computation methods suffer from the sparsity problem of rating data,and can result in the biased similarity caused by only measuring the common ratings.Thus,a new similarity computation approach named overlapping-dependency is proposed,which can be used to compute similarities by combining the overlapping similarity and the global dependency similarity.Then,a new collaborative filtering algorithm with the overlapping-dependency of multiple user behaviors is presented.The weighting factor is introduced in the similarity computing,and it can measure the effects of different user behavior with regard to the recommendation quality.The proposed algorithm is tested on the Top-md dataset,which is collected from the healthcare service website of “www.topmd.cn”.The experimental results show that the new algorithm can outperform the traditional methods on the prediction of ratings,and effectively improve the recommendation accuracy.2.A novel Bayesian Personalized Ranking algorithm via learning pairwise preferences over Mixed-Type Item-sets.In some recent works,the problem of one-class collaborative filtering has rised and been widely concerned.This approach can make use of users' behavior with “one-class” feedback form coming from different services to improve the recommendation performance.Some previous works resolve the issue based on the assumptions of pointwise preference on one item,and the assumption of pairwise preference on items or item-sets based on relative score over two item sets.In the pairwise approaches,BPR(Bayesian Personalized Ranking)and CoFiSet(collaborative filtering via learning pairwise preferences over item-sets)perform empirically much better by utilizing such one-class data well.Nevertheless,such pairwise preference assumption with regard to items or item-sets is always invalid in real-world applications.We adopt an alternative assumption of pairwise preferences over mixed-type item-sets by defining the preference on two item sets with different type instead of on a single type of item set.One is an item set with the same-type feedback relationship,and another is a mixed-type item set.With this new assumption,we then develop a novel algorithm named Bayesian Personalized Ranking via learning pairwise preferences over Mixed-Type Item-sets(MT-BPR).A series of experiments with several state-of-the-art methods are conducted on three real-world datasets,and find that our new assumption is more general and reasonable.The new algorithm of MT-BPR can make use of the one-class data more effectively and achieve better recommendation performance.3.A novel Bayesian Personalized Ranking algorithm with multiple implicit feedback.The recommendation methods can infer users' potential requirements and interests with implicit feedback,which can be derived more easily and the data volume is bigger than explicit feedback.Due to the lacking of negative samples in implicit feedback,it becomes more challenging to reflect the users' preferences directly and completely.To deal with this problem,this paper focuses on utilizing multiple feedback for better recommendation models based on the Bayesian Personalized Ranking,given the assumption regarding a new class of items referred to as “auxiliary feedback”.A special coefficient is introduced to measure the preference distance between multiple actions of the users.Then,we propose a new algorithm called Bayesian Personalized Ranking with multiple implicit feedback(MBPR).A series of experiments are conducted on three real-world datasets,and the empirical results show that the better performance can be achieved in comparison with the other counterparts.4.A novel Bayesian Personalized Ranking algorithm based on the Confidence of Multi-type Auxiliary Implicit Feedback.Due to the uncertainty of multiple types of implicit feedback from user behavior,it becomes more challenging to mine the users' preferences precisely.To deal with this problem,this paper focuses on utilizing multiple feedback for better recommendation models based on the Bayesian Personalized Ranking.The confidence of the multi-type auxiliary feedback is learned by the two methods,including Logistic Regression and Tree-based feature selection.Then,the confidence can be used to select “certain auxiliary feedback”.We propose a new algorithm called Multi-type Implicit Feedback Confidence Learned Bayesian Personalized Ranking(MTC-BPR)which can select more effective training samples from the multiple types of implicit feedback.A series of experiments are conducted on two real-world datasets,and the empirical results show that our method outperforms the other counterparts.5.An online healthcare services platform(www.topmd.cn)was designed and developed by the laboratory which the author works in.The four recommendation algorithms proposed in this dissertation are applied to this website,and the empirical results show that our methods can achieve better recommendation performance.