近年來，社群網站的蓬勃發展，讓使用者愈來愈輕易地能在網路上發表自己的言論。在Google Map上台灣的餐廳也已累積到了一定的數量可用來分析台灣用戶的用餐喜好。而目前在台灣基於評論資料情緒分析的推薦系統相對較少，於是我們想要藉由在Google Map上用戶對於餐廳的評價資料來建立一個餐廳推薦系統，並且知道情緒分析是否對於推薦效能有所影響。
推薦系統的核心概念是找出使用者的喜好，針對喜好做個人化推薦。一般都會將使用者對於項目的評分做為喜好分析依據，而現在許多平台上除了有使用者的評分外還連帶著使用者對於項目的評論文字資料。傳統的推薦系統方法主要著重於評分資訊的分析，我們想知道除了分析評分資訊，若進一步加上評論文字資料的情緒是否能夠更代表使用者喜好。我們蒐集了Google Map上台灣所有的餐廳評論資訊做為分析資料，它的評分資訊是以1到5分為尺度，其中取出4分以上的評論當作使用者的喜好餐廳資訊共1,687,390筆資料。在情緒分析方面，我們將擁有100篇評論以上的餐廳評論資訊中挑出所有的5分評論做為正向情緒資料集，1分評論做為負向情緒資料集，共使用了66,357句的負向情緒資料及64,998句的正向情緒資料來訓練評論情緒分類模型。為了分析使用者評論的文字情緒，我們嘗試了三種分類器，分別是傳統的SVM, Naïve Bayes分類器以及類神經網路的LSTM方法來將評論資料分成正向評論或是負向評論，利用評論情緒來判斷使用者對於項目的喜好。分析完喜好後我們將它結合到Weighted Matrix Factorization以及Matrix Factorization with Item Co-occurrence兩種推薦系統方法來測試新的使用者喜好分析對於喜好預測是否有助益。
最後使用了三種評測推薦系統的方法來比較我們系統的效能，分別是MAP (Mean Average Precision)、Recall、NDCG (Normalized Discount Cumulative Gain)，首先將資料集以8:2的比例分出訓練集和測試集，藉由計算預測喜好項目是否有出現在測試集來評估系統好壞。結果發現三個評測指標中加入了情緒分析的推薦系統在MAP的數據上升了5.77%、NDCG上升了8.26%以及Recall上升了8.81%。實驗結果顯示評論文字資料的情緒分析對於整體的推薦效能是有提升的。

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.

1. Adomavicius, G. and A. Tuzhilin, Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions. IEEE transactions on knowledge and data engineering, 2005. 17(6): p. 734-749.
2. Bengio, D.B.K.C.Y., Neural machine translation by jointly learning to align and translate. arXiv, 2014.
3. Cortes, C. and V. Vapnik, Support-vector networks. Machine Learning, 1995. 20(3): p. 273-297.
4. Friedman, N., D. Geiger, and M. Goldszmidt, Bayesian Network Classifiers. Machine Learning, 1997. 29(2): p. 131-163.
5. Hochreiter, S. and J. Schmidhuber, Long Short-Term Memory. Neural Computation, 1997. 9(8): p. 1735-1780.
6. Hu, Y., Y. Koren, and C. Volinsky. Collaborative Filtering for Implicit Feedback Datasets. in 2008 Eighth IEEE International Conference on Data Mining. 2008.
7. Jain, L.C. and L.R. Medsker, Recurrent Neural Networks: Design and Applications. 1999: CRC Press, Inc. 416.
8. Koren, Y., R. Bell, and C. Volinsky, Matrix factorization techniques for recommender systems. Computer, 2009. 42(8).
9. Liang, D., et al., Factorization Meets the Item Embedding. 2016: p. 59-66.
10. Luong, M.-T., H. Pham, and C.D. Manning, Effective approaches to attention-based neural machine translation. arXiv preprint arXiv:1508.04025, 2015.
11. Mikolov, T., et al., Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781, 2013.
12. Mikolov, T., et al. Distributed representations of words and phrases and their compositionality. in Advances in neural information processing systems. 2013.
13. Pang, B. and L. Lee, Opinion Mining and Sentiment Analysis. Found. Trends Inf. Retr., 2008. 2(1-2): p. 1-135.
14. Ricci, F., L. Rokach, and B. Shapira, Introduction to Recommender Systems Handbook. 2011: p. 1-35.
15. Sarwar, B., et al., Analysis of recommendation algorithms for e-commerce, in Proceedings of the 2nd ACM conference on Electronic commerce. 2000, ACM: Minneapolis, Minnesota, USA. p. 158-167.
16. Schafer, J.B., et al., Collaborative filtering recommender systems, in The adaptive web. 2007, Springer. p. 291-324.
17. Schafer, J.B., J. Konstan, and J. Riedl, Recommender systems in e-commerce, in Proceedings of the 1st ACM conference on Electronic commerce. 1999, ACM: Denver, Colorado, USA. p. 158-166.
18. Schafer, J.B., J.A. Konstan, and J. Riedl, E-Commerce Recommendation Applications. Data Min. Knowl. Discov., 2001. 5(1-2): p. 115-153.
19. Schmidhuber, J., Deep Learning in Neural Networks: An Overview. Neural Networks, 2015. 61: p. 85-117.
20. Tang, H., S. Tan, and X. Cheng, A survey on sentiment detection of reviews. Expert Systems with Applications, 2009. 36(7): p. 10760-10773.