Research On Incomplete Multi-Domain Transfer Learning For Cross-domain Recommendation Method

With the development of mobile Internet and the rapid growth of data scale,the recommendation system not only improves the efficiency of user selection,but also solves the individual needs of data.It has increasingly become an important branch of machine learning,and has been rapidly developed in both theory and practice.Collaborative filtering algorithm has been applied to recommendation systems for a long time.However,data sparsity is one of the most important problems in traditional single domain collaborative filtering.Extremely sparse high-dimensional rating data in matrices can lead to similarity characterization problems.On the one hand,the recommendation accuracy decreases rapidly with the increase of data sparsity in practice.On the other hand,data with 95% sparsity are widely exist in the real world.Therefore,alleviating the performance degradation caused by data sparsity is one of the most important research areas in the recommendation system.Recently,some research attempts to apply the transfer learning method to solve the data sparsity problem of the traditional collaborative filtering.Transfer learning is the basic method to solve the learning task with scarcity of labeled data,and its research remains in a challenging stage.Hence,this paper aims to apply the transfer learning to the recommendation task and systematically studies the problem challenges and solutions in the cross-domain recommendation of transfer learning based on rating patterns sharing model.When the transfer learning algorithm is introduced into recommender system,it forms a new research branch called cross-domain recommendation.However,transfer learning involves several critical problems such as under-fitting,over-fitting,under-adaptation,incompleteness and negative transfer are still waiting to be solved.Firstly,the single domain transfer learning model suffers from underfitting problem;Secondly,existing multiple domains transfer learning model suffers from overfitting problem;Thirdly,The under-adaptation problem means that the codebook cannot be properly adapted to the size and characteristics of the source domain,therefore,the recommendation algorithm cannot obtain effective knowledge gain when the data size of the source domain is expanded;Fourthly,incompleteness is the case where the source domain data is incomplete in transfer learning.In the real world,data with blank items and missing values is ubiquitous and these data are incomplete data;Fifthly,negative transfer refers to the side effects of source domain knowledge on the target domain learning tasks.This thesis focuses on addressing the under-fitting,over-fitting,under-adaptation,incompleteness,and negative transfer issues,analyzing the intrinsic causes,and designs specific learning models.The novel contributions are summarized as follows.1.For the incompleteness problem,this thesis proposes the Incomplete Orthogonal Nonnegative Matrix Tri-factorizations method in chapter two which extends the existing Orthogonal Nonnegative Matrix Tri-factorizations method.Therefore,the data completeness limitation in the original transfer learning model is removed,and the source domain selection for transfer learning in cross-domain recommendation is extended from the complete domain to the incomplete domain,which increases both the optional range of source domains and the available data size.2.Aiming to solve the under-adaptation problem,this thesis proposes the Low-Dimension Latent-Factor Representation Selection algorithm in chapter three which heuristically learns the appropriate codebook according to the size and characteristics of the source domain data,so that the transfer learning model can adapt the source domain seasonably.In the previous cross-domain recommendation model,the fixed-scale codebook is not adapted to the source domain data.Our algorithm overcomes that issue and solve the rating patterns expressiveness problem which is caused by the averaging trend of codebook in transfer learning.3.Aiming to solve the under-fitting and over-fitting problems,the Multiple Incomplete Domains Regularized Transfer Learning model is proposed to solve the under-fitting problem of the single-source transfer learning model by transferring multiple source domains' rating patterns.By adding regularization constraints to the target domain adaptation,the over-fitting problem in existing multiple source domains transfer learning model is solved by which the transfer learning model is more robust.4.For addressing the negative transfer issue,the restoration learning process after the transfer learning is proposed.In order to implement the restoration learning process,a context restoration model is proposed in chapter five,and then extended to the multiple gradient layer restoration model in chapter six to mine the unique image,metadata,context,category,attributes and other information specific to the target domain.By restoring the ratings of the transfer learning results that do not fit the characteristics of the target domain,the negative transfer phenomenon caused by transfer learning is overcome by which the rating prediction accuracy of transfer learning and the precision of cross-domain recommendation are both improved.5.Chapter six proposes latent factor restoration model for the case where data sets in some domains do not contain available context knowledge.By extending the existing non-negative matrix factorization to the partial-adaptive non-negative matrix factorization method,the model learns the target domain specific latent rating factors according to the existing ratings in the target domain.Then,the transfer learning results are restored according to the latent rating factors in the target domain.