Restaurant Recommender System with Review Sentiment Analysis

• Main Authors: Zi-HsuanHung, 洪梓軒
• Other Authors: Jung-Hsien Chiang
• Format: Others
• Language: en_US
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/kqpc9q

碩士 === 國立成功大學 === 資訊工程學系 === 106 === Recently, the social network has been well developed. It is easier and easier that users can express their opinions on the Internet. In Taiwan, there are much restaurant review data on Google Map. However, there are few recommender systems based on review sentiment analysis in Taiwan. We want to build a recommender system by analyzing review sentiment, and test if sentiment analysis affects the recommendation performance. This study focuses on analyzing users’ preference and recommending restaurants to users. Traditional recommender system methods used the rating to an item from users to analyze their preference. Nowadays, many platforms let users rate the items along with text reviews. We wonder if review texts can represent the user preference more. We collected all Taiwan restaurants review data on Google Map for analysis. The rating scale is 1 to 5 stars. We take the 4 and 5 stars rating data as user preference data, 1,687,390 data in total. Sentiment analysis, we pick all 5-star reviews from restaurants that received more than 100 reviews as positive sentiment dataset, and all 1-star reviews as negative sentiment dataset. We used 66,357 sentences of negative reviews and 64,998 sentences of positive reviews to train our sentiment classifier. In order to analyze the sentiment of user reviews, we try three kinds of the classifier, which are support vector machine, Naïve Bayes, and long short-term memory network, to classify review data into positive or negative. With the sentiment analysis, we can know the preference of users about restaurants. After sentiment analysis, we update the user-restaurant rating matrix. Then, we use two recommender system, weighted matrix factorization and matrix factorization with item co-occurrence, as baseline method to test if sentiment analysis is beneficial for preference prediction. At last, we use three evaluation metrics, mean average precision (MAP), Recall, normalized discounted cumulative gain (NDCG), to compare our system to baseline method. We first split the dataset into 8:2 ratio as training and testing dataset. By predicting testing data’s label, we can know how our system performs. As a result, our restaurant recommender system with sentiment analysis enhance 5.77% on MAP, 8.26% on NDCG and 8.81% on recall. The results show that sentiment analysis on review texts for recommender system enhance the recommendation performance.