Summary: | 碩士 === 國立中正大學 === 資訊工程研究所 === 104 === There are many blogs and food recommendation websites that consist of lots of photos. When choosing restaurants, we usually browse restaurant information on the internet, where visual information provides positive influence in addition to text. However, previous related works didn’t use visual information for restaurant recommendation. In this thesis, we conduct a pilot study about how to use visual information to recommend restaurants for users.
We first introduce two common approaches (content-based filtering approach and collaborative filtering approach) for recommendation and related research challenges. We then study how to reduce these problems by jointly considering visual information and user preference.
We design a framework to extract features from photos and text information in restaurant-related blog articles to describe restaurant attributes and represent user preference. We extract keywords from blog articles as restaurants attributes, and categorize images into four classes to describe visual information in different aspects. We analyze effectiveness of these features, and prove that visual information can present restaurant well and yield better performance.
We integrate visual information into two types of recommendation approaches, and reduces the problems they face to improve recommendation performance.
|