Detecting whether a song is favorite for a user is an important but also challenging task in music recommendation. One of critical steps to do this task is to select important features for the detection. This paper presents two methods to evaluate feature importance, in which we compared nine available features based on a large user log in the real world. The set of features includes song metadata, acoustic feature, and user preference used by Collaborative Filtering techniques. The evaluation methods are designed from two views: i) the correlation between the estimated scores by song similarity in respect of a feature and the scores estimated by real play count, ii) feature selection methods over a binary classification problem, i.e., “like” or “dislike”. The experimental results show the user preference is the most important feature and artist similarity is of the second importance among these nine features.