In this work, we investigated the automatic prediction of user ratings for recipes. Information including the ingredients, the instructions, and the reviews from Epicurious were fed into a machine learner, a multi-class support vector machine, to examine how reliable they are when predicting recipe ratings. Our results show that information from the reviews results in the most reliable predictions: we reached an accuracy of 62%. The problem is difficult, partly because of the skewing of the ratings: most recipes are rated with 3 or 4 out of 4 forks.
展开▼
