| Abstract: | Classification techniques have been applied to real world problems such as fish classification and email sorting. In this work, we introduce a new application called ANIMAL. ANIMAL is a Machine Learning Disc Jockey. A model for beat (from music) and bop (from head motion) is proposed. Using this model, we treat a listener's musical enjoyment as a classification problem. We define a beat/bop similarity metric based on harmonic matching over frequencies in the Fourier domain across the raw inputs (windowed proportionally to heterogeneous sampling rates, uncovered empirically). From our similarity metric, we define features which we use in a number of classification methods. We use both generative and discriminative methods such as Logistic Regression, Naive Bayes, Stochastic Gradient Linear Regression and Support Vector Machines (SVMs). We have results for a test set comprised of a 33% set aside from our corpus of data. We compare the performance of these methods among different sets of features extracted from the harmonic match. Our results show that the Naive Bayes Classifier outperforms the aforementioned classification techniques. |
