A Study Of Review-based Neural Recommendation Algorithm

With the rapid development of information technology,people spend more and more time online.Recommender system tries to improve the online service by gathering user historical records and locating user interest by information filtering.Recommender system is the core technique to solve the problem of information overload in machine learning.Recommender system has been proven to produce huge commercial profits in the field of electronic commerce,news,video and advertisement.It can attracts a lot of attention from academic world.However,due to the sparsity problem and long-tail effects of user history data,it is still difficult to accurately predict user tastes.To solve this problem,people try to utilize auxiliary information to capture the relations between user behaviors and tastes as well as boosting recommendation accuracies.Generally,massive auxiliary information data exists in large-scale commercial recommender system,such as user age,regions and tags etc.In addition,multi-media data,such as item description,pictures and user review texts,abound in on-line recommender community,which can be utilized to help better understand user behavior and improve recommendation accuracy.The auxiliary information is easy to collect and can represent the relation between user behavior and user taste.Among those auxiliary information,review may be the most informative one,because it can reflect both user preferences and item characteristics.Besides,it can help generate plausible recommendation explanations.This thesis analyzes the possible reasons why current recommendation algorithm fails to give satisfied recommendation results and proposes a deep learning based recommendation algorithm that jointly learns from ratings and reviews.The proposed methods can boost recommendation accuracy as well as provide plausible explanations.The major contributions of this thesis can be summarized as follows:(1)Based on modern natural language technique,this thesis put forward a text feature extraction module which fits the recommender system context.It can effectively extracts semantic feature from review text data and help boost recommendation accuracy.With the help of attention mechanism,the proposed text feature extraction module can highlight different words and phrases according to the different weights assigned by attention mechanism and provides plausible explanation.Meanwhile,massive experiments are designed to prove the effectiveness of the proposed text feature extraction module.(2)With the help of deep learning technique,this thesis implements a deep learning based recommendation algorithm that fuses both ratings and reviews.The proposed framework can learn user tastes and semantics from reviews simultaneously and has clear architecture which is easy to implement.It can be trained end to end,so it’s easy to train,test and employ.(3)This thesis proposes a novel technique to fuse feature from heterogeneous data,which can learn user and item heterogeneous feature simultaneously in an effective way.Based on the proposed framework,extensive experiments are conducted to prove the effectiveness of the proposed feature fusion mechanism by comparing to various kings of feature fusion technique.Several latest and state-of-the-art recommendation algorithm are include for performance comparisons.In the end,this thesis analyzes the why the proposed method show superior performance and possible future works.