Improving Dual-Regularized Matrix Factorization Using Hierarchical Attention Networks

  • 楊 潓賢

Student thesis: Doctoral Thesis


Recently it is getting tougher and tougher to come to a decision because we get so much information through the internet every hour Therefore recom- mender systems (RSs) are becoming more and more important They can help people choose effectively by suggesting suitable options to people Most of the hy- brid RSs extensively utilize review text information to improve their performance However it is still a challenge for RSs to extract the most informative features from huge numbers of reviews In this thesis a new RS called the HAN-PMF model which combines prob- abilistic matrix factorization and hierarchical attention networks is proposed The aim is to learn which kinds of items users like by extracting ratings and reviews To get an optimal model and parameters the system uses the Amazon 7 dataset to train the HAN-PMF model and the root mean square error as the evaluation metric The difference between the results is influenced by the random seed so we fix the seed and conduct the experiment five times to ensure experimental reliability Our result shows a 1 47% improvement compared with DRMF and a 13 05% improve- ment compared with PMF We then extract the 20 most important words to explain customer insight
Date of Award2019
Original languageEnglish
SupervisorRen-Shiou Liu (Supervisor)

Cite this