Food Recommendation System Project Report

This document discusses content-based recommendation systems, i.e., systems that recommend an item to a user based upon a description of the item and a profile of the user’s interests. Basically it gives you a clear example of content-based recommendation systems, specifically a food recommendation system. The dataset that is used in this project is from “yemeksepeti.com”. Similar foods are recommended according to previously constructed user preferences. Keywords— Information Retrieval, Recommendation Systems, ContentBased Approach, Expert Systems, Web Crawling, Feature Similarity 1Introduction and Problem Definition People make decisions everyday. “Which movie should I see?”, “Which city should I visit?”, “What should I eat?”... There are too many choices and a little time to explore them all. Recommendation systems help people make decisions in these complex information spaces. Recommendation systems are a type of information filtering that presents lists of items (films, songs, books, videos, images, products, web pages...) which are likely of user interest. Amazon, Last.fm, Ulike, iLike, Netflix, Pandora are the most popular recommender systems all over the world. Simply they compare user interest acquired from his/her profile with some reference characteristics and predict the rating that the user would give. Those characteristics may be from the item information (content-based approach) or the user's social profile (collaborative filtering approach). I focus on the question “What should I eat?” in the scope of this project. My system uses content-based recommendation technique for producing food recommendations. It is based on similarity of foods. Basically, my system constructs user profiles from the previously rated features and food profiles from the ingredients of the food, then it recommends the most appropriate foods according to the preferences of the users. This document is organized as follows. Overall system design is reviewed in the next section. System Architecture and experimental evaluation appear in Sections 3 and 4. Discussions about the problems I encountered are mentioned in Section 5. Conclusion and future work are presented in Sections 6. 2 Overall System Design 2.1 Recommendation Technique: My system is a content based recommendation system. Food domain can be seen as a set of foods where each food has a set of ingredients. Content-based approaches treat the recommendation problem as a search for related items. Given a rated food, the algorithm constructs a search to find other related items with the same ingredient. If a user likes salmon, for example the system might recommend other foods having salmon like sushi. However, this is the simplest logic behind content-based recommendation systems. In my system foods are defined by their important features and represented by vectors. Thus, feature weights are crucial in these vectors. Similarity is computed based on item attributes using appropriate distance measures. Content-based recommendation systems share in common describing the items that may be recommended, creating a profile of the user that describes the types of items the user likes, and comparing items to the user profile to determine what to recommend. My content-based recommendation system can be seen as a combination of three distinct parts; food profiling, user profiling and recommending foods according to the previous feedback of the users. Moreover, food domain can be seen as a set of features where each feature is actually an ingredient of that food. Therefore, both item and user profiles are kept as vectors of features. As my system is a contentbased recommendation system, it tries to find best matches between the user profile and the food profiles. 2.2 Item Profile Representation In content based recommendation systems, every item is represented by a set of features or an attribute profile. A variety of distance measures between the feature vectors may be used to compute the similarity of two items. For example Euclidian or cosine similarity supposes that all the features have equal importances. However, human judgment of similarity between two items often gives different weights to different attributes. Moreover, document frequency is more commonplace to use for this purpose. Here N denotes for the total number of foods in a collection, and dft is for the total number of foods that have the ingredient “t”. Thus the idf of a rare term is high, whereas the idf of a frequent term is likely to be low. This method is called “inverse document frequency” and I have assigned the calculated similarity as weights of the features. 2.3 Knowledge Acquisition Technique & User Profile Representation The type of the user profile derived by a content-based recommender depends on the learning method employed. Decision trees, neural nets, and vector-based representations can be all used. In this project I have used decision vector based representations constructed with the help of user ratings. At this step, my system uses explicit data collection. Specifically, after each recommendation, user can explicitly state whether the recommendation is satisfying or not. The next recommendations will be mostly based on this user feedback.