A Study Of Recommendation Based On Collaborative Filtering With User Reviews

In recent years,with the development of e-commerce,recommendation system has been widely used to help user's activities on the e-commerce.Recommendation system has been widely used to help user's activities on the e-commerce.Recommendation is a task to suggest new items automatically,such as products,food,videos,musics,books etc.in which a target user(active user)may be interested.Although in many approaches to develop recommendation systems,the collaborative filtering(CF)is an effective and widely used technique for recommendation.In the CF,a rating of an“active user”for an unknown item is guessed by ratings of the“active user”for other items as well as ratings of other users of other items.On the other hand,in many e-commerce web sites,users can write a review and express their opinions for the item they bought.Such user reviews are another useful resource to guess a rating for a new item.However,the use of user reviews in collaborative filtering has not been paid much attention in the previous work.This thesis proposed a new method of the collaborative filtering that considers both user ratings marks by others and user reviews written by others.In the traditional CF systems,the similarity about users is measured by comparing the ratings given by two users,and then the preferred items of similar users are recommended to active users.We also obtain similar users by considering user reviews.Finally,we explore a way to combine these two collaborative filtering methods to train a more accurate recommendation system model.In this study,the recommendation task is defined as follows.Let us suppose that there is a collection of items and users.Each user gives ratings for not all but several items.In addition,we suppose that a user can write user reviews for items.Therefore,there exists a collection of quadruplet:(user_i,item_j,rating_ij,review_ij),where rating_ijstands for a rating given by user_i for item_j,and review_ijstands for a review text for item_j written by user_i.Hereafter,we call this dataset as“rating review history dataset”.The goal of our task is to predict ratings of an“active user”for unknown items when a rating review history dataset is given.Once we have guessed the ratings of the unknown items,we can recommend the highly rated items to the active user.Our method composed of three modules:the collaborative filtering with rating,the collaborative filtering with review,and the hybrid method of these two CF methods.The first method is the CF with rating,which uses the rating of other users.It is the same method in the previous work.First,guess the similarity between two users.The user similarity evaluates how ratings for items given by one user is similar to that by the other user.It can be calculated from a rating review history dataset.Next,a rating for a target item(an unknown item)of an“active user”is guessed.Roughly saying,it is calculated by sum of the rating for the predicted item given by the other users,where the weights are defined by the user similarity.The second method is the CF with review,which is a system that considers user reviews.The basic idea is to measure the similarity between users by the similarity of user reviews written by them.In this method,if two users write similar reviews and express similar opinions in their comments,the user similarity becomes high.First,the review texts written by a user for different items are concatenated as a single document.Then,the document embedding,which is an abstract vector representation of a document,is calculated for the concatenated review for each user.We use doc2vec tool to obtain the document embedding.The user similarity is measured by the cosine similarity between two document vectors.Finally,the rating for an unknown item of an“active user”is guessed in the same way of the CF with rating.The third method is the hybrid method that combines the CF with rating and the CF with review.Since the above two CF methods guess ratings from different points of view,we can expect that the use of both methods can improve the performance of the recommendation system.In this method,the rating of an“active user”for a target item is guessed by the linear combination of the rating scores guessed by two methods.In other words,the rating is estimated by the weighted sum of two predicted scores.The weight parameters are optimized using a development data.The objective(loss)function of this optimization is the sum of the difference between the predicted and gold rating scores.The weight parameters are determined so that the loss function on the development data is minimized.An experiment is conducted to evaluate our proposed hybrid recommendation method.We use two rating review history datasets.One is Amazon product data for electronics that consists of 192K users,448K items,375K reviews,and 375K ratings.The other is Amazon Instant Video dataset that consists of 5.2K users,1.7K items,33K reviews,and 33K ratings.The quadruplets(user_i,item_j,rating_ij,review_ij)in the datasets are divided into the training,development and test data.The evaluation criteria used in this experiment is the root-mean-square error(RMSE).Since it is a good way to evaluate the difference between the predicted and gold ratings,the smaller RMSE means the better result.We found that the CF with rating and the CF with review were comparable,but the combination of them could improve the performance.The overall results on the Amazon Instant Video dataset are better than that on the Amazon electronics dataset.Again,our proposed method can improve the performance.From the above results,we can conclude that the use of both rating and reviews in the collaborative filtering is effective.In general,the collaborative filtering is not effective to solve a cold-start problem.It is difficult to guess a rating of a really new user who has not given any rating for items.In the future,we try to refine and apply our method to solve the cold-start problem.