Combining user preferences and user opinions for accurate recommendation
Highlights
► We proposed a novel method for opinion-feature extraction from online reviews. ► We proposed a method to combine user preference and opinion for recommendation. ► Experimental results demonstrate the effectiveness of the proposed methods.
Introduction
Given the development and popularization of the Internet, the amount of available information has increased dramatically, resulting in the problem of information overload. A personalized recommendation system (Adomavicius and Tuzhilin 2005) is one effective way to solve this problem and has been used in many applications (Ali and Van Stam, 2004, Bennett and Lanningm, 2007, Goldberg et al., 2001, Linden et al., 2003). However, the sparseness of a data set makes many recommendation methods such as collaborative filtering suffer from low accuracy. Collaborative filtering (Breese et al., 1998, Resnick et al., 1994) and multi-criteria recommendation methods (Adomavicius and Kwon, 2007, Adomavicius et al., 2010, Tang and McCalla, 2009) are usually based on commonality among customers. Similar users or items are found by measuring the similarity of the common rating scores of users. However, the insufficiency of data makes this method unsatisfactory, because the number of items rated by both users is usually very small.
We propose a very different method to infer the preference of the user, and this method is called preference and opinion-based recommendation. This method recommends items based on the preferences of the user on each feature of an item, and the preferences are inferred based on the difference of the opinions of the user from others’ opinions. The concept behind this method is that if a user frequently gives a lower score about a feature than other users after eliminating the effect of the rating habit of the user, he may have a higher requirement for this feature. Thus, the features and opinions for which users expressed their comments (feature-opinion extraction problem) need to be extracted, and the polarity of the opinions (sentiment analysis problem) should be analyzed. Effective measures to describe the preference of the user and a design recommendation strategy (recommendation problem) based on the preferred features of the user also need to be developed.
Therefore, we will focus on two problems: the feature-opinion extraction problem and the recommendation problem. We discuss these issues based on online restaurant reviews written in Chinese. Existing studies on the feature-opinion extraction problem were mainly conducted on reviews written in English (Ding and Liu, 2007, Hu and Liu, 2004) and other languages such as Japanese (Kobayashi et al., 2006, Morinaga et al., 2002). Little work has been done on Chinese reviews.
The contributions of this paper are as follows. (1) We propose a new method to extract features and opinions from online user reviews. The method is called the adverb-based opinion-feature extraction method, which uses Chinese expression characteristics to enhance extraction accuracy. (2) We develop two novel measures, concern and requirement, to capture the preference of the user toward features of an item. Based on these measures, we design a recommendation method. (3) We propose an algorithm called PORE to incorporate these above-mentioned methods for online review-based recommendation. We also conduct an empirical study to evaluate the performance of the algorithm. Experimental results obtained from real online reviews show that the methods proposed are effective in dealing with insufficient data and are more accurate and efficient than existing traditional methods.
The rest of the paper is organized as follows. In Section 2, we formally define two research problem, and introduce some definitions and notation. The feature-opinion extraction method and the proposed recommendation method are described in Sections 3 Opinion-feature extraction, 4 Combining user preference and opinion for recommendation. The performance of the proposed methods is evaluated in Section 5. Related work is discussed in Section 6. We conclude in Section 7.
Section snippets
Problem definition
An online review is about an object, such as a product like a camera, or about a service, such as that of a restaurant. Each object is called an item. We consider online restaurant reviews, and each restaurant that commented by a user is an item. Suppose we have a collection of M reviews, D = {d1, d2, … , dM}, about a certain kind of item. Let U be the set of N users who wrote these reviews, U = {u1, u2, … , uN}, and R be the set of K items users commented on, R = {r1, r2, … , rK }. Let F be a set of L features
Opinion-feature extraction
Existing feature opinion extraction methods can be classified into two types, namely, supervised learning (Liu et al., 2005, Ghani et al., 2006) and unsupervised learning (Hu and Liu 2004, Popescu and Etzioni 2005). We focus on unsupervised learning, which is domain-independent and also saves time spent manually labeling training data sets.
Existing popular feature opinion extraction methods (Hu and Liu, 2004, Liu et al., 2005) first find features and then make use of them to obtain the opinions
Combining user preference and opinion for recommendation
The traditional recommendation methods such as collaborative filtering are based mainly on the idea that similar customers may have similar interests. In other words, similar customers may like similar products or services. These methods focus on the commonality among customers or among items. However, the difference between customers is another angle toward solving the problem. Besides, collaborative filtering uses only the rank or overall score information. In fact, online customer reviews
Empirical study
To evaluate the performance of the proposed approaches, we crawled a real online restaurant review dataset through a well-known website (beijing.koubei.com). A total of 54,208 reviews were collected. The restaurants with only one review were excluded. The total number of reviews is 53,734, which were written by 9992 users for 3094 restaurants.
Running environment. All experiments were conducted on a computer with a 2.66 GHz Intel Core 2 processor, 4 GB RAM, and 32-bit Windows Server 2008. All
Related work
Recommendation methods are usually categorized into two major classes: content-based and collaborative filtering methods. Content-based methods (Lang, 1995, Mooney and Roy, 2000) recommend items that are similar to those the user liked before. Collaborative filtering (Billsus and Pazzani, 1998, Breese et al., 1998, Goldberg et al., 2001, Marlin, 2004, Nakamura and Abe, 1998, Resnick et al., 1994) recommends items liked by similar users who share the same preferences. Our recommendation method
Conclusion
We have proposed a very different method to infer a user’s preferences and to recommend products to users, which we call PORE. This method recommends items based on the preferences of users on the features of each item. To infer a user’s preferences on the features, we developed a new method to extract features and opinions from online user reviews and proposed measures to infer a user’s preferences. We evaluated these methods by conducting experiments on online restaurant reviews written in
Acknowledgements
This work was supported by the National Natural Science Foundation of China under Grant Nos. 70871068, 70890083 and 71110107027 and the 973 program of China under Grant No. 2012CB316205.
References (40)
NewsWeeder: Learning to filter netnews
- et al.
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) - et al.
New recommendation techniques for multi-criteria rating systems
IEEE Intelligent Systems
(2007) - et al.
Multi-criteria recommender systems
- et al.
Mining Text Data
(2012) - Alexander, P., and Patrick, P. Twitter as a corpus for sentiment analysis and opinion mining. In Proceedings of the...
- Ali, K., and Van Stam, W. TIVo: Making show recommendations using a distributed collaborative filtering architecture....
- Benamara, F., Cesarano, C., Picariello, A., Reforgiato, D., and Subrahmanian, V. Sentiment analysis: Adjectives and...
- Bennett, J., and Lanningm, S. The Netflix prize. In Proceedings of 2007 KDD Cup and Workshop, San Jose, CA,...
- et al.
Learning collaborative information filters
Empirical analysis of predictive algorithms for collaborative filtering
WordNet: An Electronic Lexical Database
The utility of linguistic rules in opinion mining
Pulse: Mining customer opinions from free text
An energy-efficient mobile recommender system
Text mining for product attribute extraction
SIGKDD Explorations
Eigentaste: a constant time collaborative filtering algorithm
Information Retrieval
Conjoint analysis in consumer research: issues and outlook
Journal of Consumer Research
Conjoint analysis in marketing: new developments with implications for research and practice
Journal of Marketing
Cited by (130)
Combining review-based collaborative filtering and matrix factorization: A solution to rating's sparsity problem
2022, Decision Support SystemsAspect-based approach to measure performance of financial services using voice of customer
2022, Journal of King Saud University - Computer and Information SciencesCitation Excerpt :They used FAHP to evaluate relative weights of financial and non-financial variables final ranking was evaluated using TOPSIS technique. Liu et al. (2013) proposed a recommender algorithm based on user opinions. They consider ratings and opinion to create a restaurant recommendation system.
Movie Recommendation Algorithm Based on Sentiment Analysis and LDA
2021, Procedia Computer ScienceIs UGC sentiment helpful for recommendation? An application of sentiment-based recommendation model
2024, Industrial Management and Data SystemsSentiment analysis: A survey on design framework, applications and future scopes
2023, Artificial Intelligence ReviewA deep learning model for online doctor rating prediction
2023, Journal of Forecasting