This paper aims to solve the problem of song classification by decade, so that online music apps who use this method can try different categories to catch users’ interests then gain advantages in this competitive field. The data is from Million Song Dataset which involves time-series features and time independent features. To achieve the best result, an ensemble learning architecture with LSTM Neural Network and Feed-forward Neural Network is implemented to make predictions. This model works for this problem and outputs high accuracy. The main find ing from the project is: motivated by how different parts of human brain co-operates a new ensemble Deep Learning architecture is adopted.

[1] C. Doll, "Million Song Dataset. Columbia University. http://labrosa.ee.columbia.edu/millionsong," Journal of the Society for American Music, vol. 8, (1), pp. 121, 2014.