A Customized Real Time Restaurant Recommendation System

Open Access
Jiang, Rui
Graduate Program:
Computer Science and Engineering
Master of Science
Document Type:
Master Thesis
Date of Defense:
August 03, 2015
Committee Members:
  • Sencun Zhu, Thesis Advisor
  • Wang Chien Lee, Thesis Advisor
  • machine learning
  • natural language process
  • real-time recommendation
  • customized
According to a Search Engine Land survey conducted in 2014, 88% of consumers have read online reviews to determine the quality of a local business and 72% consumers say that positive reviews make them trust a local business more. What is a local business? It can be a medical practice, a restaurant, or a gym. The result of this survey shows the important role of reviews in deciding to choose a local business. Among various types of business, restaurant is the most searched category by users in Yelp. However, oftentimes people just search a restaurant by using word “restaurant”, while the word “restaurant” means differently to different individuals. For an Asian, it can mean a “Chinese restaurant” or “Thai restaurant”. How to correctly interpret search requests based on people’s preference is a challenge. Building a machine-learning model based on activity history of a registered user can solve this problem. The activity histories used by this thesis are reviews and ratings from users. This thesis introduces a data processing pipeline, which uses reviews from registered users to generate a machine-learning model for each business and each registered user. This thesis also defines an architecture, which uses the generated machine-learning models to support real-time personalized recommendations for restaurant searching. In order to find a good machine-learning model, we have tried several collaborating filtering methods to predict ratings between restaurants and users. The methods we have implemented are Slope One, k-Nearest Neighbors algorithm, and multiclass SVM classification. Our evaluation shows that the multiclass SVM classification method outperforms the other methods. For rating prediction, we compare user-based and item-based collaborative filtering algorithms. Finally, architecture is given to support the building of a real-time recommendation service.