Research On Multi-network Relation Prediction And Recommendation-Based On Active Transfer Learning

Traditional machine learning usually requires sufficient labeled data,however,the requirement is ofen unsatisfied in some scenarious.For example,there is a cold start problem in multi-network relationship prediction and system recommendation.In other words,there is less edge and relationaship data in the new network,which makes it difficult for the general supervised learning framework to obtain good results.Transfer learning and active leaning are two important ways to solve the problem of data sparseness.In multi-network relationship prediction,transfer learning uses the relationship information in the source network to help the relationship prediction task of the target network,and active learning queries the real labels of valuable node samples in the target network,then adds the results to the dataset to train the model again.In multi-source data recommendation system,collaborative recommendation using the auxiliary data(collaborative recommendation with auxiliary data,named as CRAD)comprehensives user's social networks and browsing history to improve the recommendation effect.Transfer learning makes use of auxiliary datas to learn user's preference information,and active learning queries the item's rating which can reflect user's preference better.This thesis proposes the following two independent methods for the above tasks:(1)multi-network relation prediction method TAQIL based on active transfer learning;(2)recommendation system TACF based on active transfer learning.The first research transfers the node information in the source network to the target network to train an initial model.Then,the data with the largest amount of information is selected from unlabeled dataset in the target network for annotation,and it is added to the labeled dataset for futher training.The above steps are repeated until the query cost is exhausted.The second research obtains users' preference information from browsing history by using transfer learning,and uses the information in the collaborative filtering to lean an initial model.At the same time,allowing the system to query users' rating of certain items,and adding them to the score matrix to learn users' s preference information according to the newest evaluation information.For the multi-network relationship prediction,experimental results on non-network dataset,network dataset show that the TAQIL method has higher classification accuracy,compared with the existing methods.For multi-source recommendation system,experimental results show that TACF method has better effect in the recommendation datasets MovieLens and MovieRating,according to the two evaluation criteria MAE and RMSE.