Research On Personalized Recommender Algorithms Based-on Machine Learning

With the rapid development of Internet,the data on the network has been grown explosively.On the one hand,massive data makes it easier to get rich information.On the other hand,people have to spend a lot of energy and time to search for the useful information.Information overload problem are becoming more and more serious.In the face of massive data resources,the traditional search engine has been unable to meet the users' needs,and personalized recommender system has become the new darling of the times.Personalized recommender system can supple the users with information or merchandise which they are interested in by analyzing user data and capturing the user's interests.In this thesis,to resolve the problems of cold start,sparseness and low accuracy of the state-of-art recommender system,we introduce the current research hotspots--machine learning,such as cross-domain recommender based on transfer learning,multi-source multi-task interest recommendation,the time series recommendation baded on cucurrent neural networks and the ontology-based tourism recommendation,which complemented the personalized recommendation.The main contributions of the thesis are as follows.In order to achieve a good cross-domain recommendation result,a cross-domain recommender algorithm model CTR-TL based on transfer learning is proposed.The model is based on the numerical scoring data on each domain,and combines the annotation text information on each domain to model the domains in a single domain.And the inter-domain modeling is carried out according to the numerical scoring data on each domain.Finally,combined with intra-domain modeling and inter-domain cross-domain modeling,while the combination of multiple domains modeling,experiments show that the model CTR-TL can solve a single domain sparseness of data.In order to improve the accuracy and recall rate,a multi-source and multi-task recommender algorithm is proposed.First we use the user's multiple personal network of personal information to build a view of the degree of interest between the tree;then based on the Web expression of interest,with the label vocabulary and weight information to represent the text of interest,combined with the user multiple social Network multi-source multi-task for interest recommendation research;finally build interest co-occurrence matrix.In order to obtain the better prediction accuracy,a recommender algorithm based onBP-RNN considering time sequence is proposed.For users in the short-term recommender system,the most likely to depend on the recent behavior of the phenomenon,the establishment of a circular neural network,the use of gated cycle unit to solve the problem of time series.The recurrent neural network treats the user's recent behavior as a sequence,and each hidden layer simulates the behavior or preferences of each user in order.We combine the recurrent neural network with the back propagation neural network.In order to make the recommender algorithm have superior performance under the multi-standard,the ontology-based tourism recommender algorithm is proposed.First,we construct the ontology of tourist attractions suitable for tourism,then use FP-Growth algorithm to excavate the association rules between users who visit different attractions,and then improve the collaborative filtering algorithm into TEUCF algorithm and TAUCF algorithm according to the difference of users.Aiming at the cooperative filtering algorithm with the embedded time factor and the evaluation factor with the characteristic evaluation result,the result of the recommendation is filtered,which is more accurate and consistent with the actual situation.The algorithm proposed in this thesis has different performance in different aspects,which realizes the individuality of the recommended algorithm and promotes the practicality of machine learning in the field of recommender system.