| Summary: | 碩士 === 國立臺灣大學 === 資訊工程學研究所 === 101 === Recommendation system has been an important and well-studied topic in recent years. However, most of the existing studies focus on the recommendation commercial produces such as movies and music. In this thesis, we aim to bring recommendation to another dimension: recipes. The most special characteristic of recipe compared to movie and music is that recipe provides detail information, ingredients and directions to help people reproduce almost the same taste food. We believe a recipe must have quite charming features, which meet people’s preferences perfectly. So people would like to reproduce it by their self, tasted it then rated it.
In this thesis, we process the problem of recipe recommendation in a different aspect. We treat recipes as an aggregation of lots features, which extract from ingredients, categories, directions, profile and nutrition. We use an extension of matrix factorization to module the how people like a feature. Then we add several extra biases to module time-dependence features, and finally we use the ensemble technology to improve our methodology. We used Root Mean Squared Error (RMSE) to evaluate result. RMSE is the most popular metric used in recommendation system to evaluating accuracy of predicted ratings. And our result RMSE is 0.5813, which is improved 0.0202 (3.36%) than MF.
|
|---|
