| With the internet gradually infiltrating into our daily life,network information has been producing,recording and disseminating all the time.It's harder for users to find the information that they truly want.Because of this,recommender system is used to help people find the information they want as efficiently as possible.Collaborative Filtering has become the most popular recommender algorithm because of its high efficiency and robustness,which uses the similarity of users to recommend information.This kind of similarity can be generated from analyzing user-item rating matrix,which is the evaluation that the user gave to the commodity and service they bought or the shop they visited.But in the reality that the numbers of users and items are keep increasing,user-item matrix is becoming sparser and the similarity between users is harder to calculate precisely.At the same time,the cold-start problem is becoming more serious.It can be inaccurate to recommend to inactive users because the lack of their record and it is also difficult to recommend new item to users since they barely be rated.Unlike collaborative filtering algorithm,knowledge-based recommendations attach great importance to the knowledge that has relations with the user or the item,even though the knowledge is from multi-source heterogeneous data.With the help of the knowledge of users and items,using knowledge-based recommendations can solve the sparsity problem and cold-start problem.As popular applications of electronic commerce,local-service and online shopping platform have stored a huge amount of information of users,items and interactions between them.Users not only rate the item they visited,but also wrote reviews for them.These reviews include how users'feel about items and reflects the emotion and sentiment changes of user while they visited the item.The way items attracts users can provoke different emotional reaction of users.For example,the item which uses novelty to attract users will have reviews include surprise emotion,and the item which attract users with its'high quality may have reviews include trust emotion.Uses'characteristics can also influent their emotional reactions toward item.For example,a traditional person may be less interested of novel items.This paper proposes a neural network collaborative filtering recommender model based on NRC sentiment analysis.Under the guidance of knowledge-based recommendations,NRC lexicon based sentiment analysis is used to build knowledge models for users and items,and the models are integrated into neural network collaborative filtering recommender algorithm.The recommender model proposed is used to recommend items that meet the sentiment requirement of users,history data is used to training models and the ratings users give to items are predicted.The experiments of this paper uses the city data of Inverness,Montreal and Scarborough from Yelp dataset,realized currently popular NMF,SVD and SVD++based collaborative filtering recommender algorithms.The neural network collaborative filtering model based on NRC sentiment analysis which this paper proposed has performed better than these three algorithms on all the three datasets Compared with NMF、SVD and SVD++,the mean RMSE value of the model proposed decreased 14.60%,2.96%and 4.60%respectivel. |
