Association Rule Mining and Audio Signal Processing for Music Discovery and Recommendation

2. Related Works

The authors in (Kim, Lee, Yoon & Lee, 2008) designed a music recommendation system based on personal preference analysis. First they built a music model using Hidden Markov Models (HMM) with Mel Frequency Cepstral Coefficients (MFCC). They calculated HMM for each song, representing one model for each song. They check the similarity between models and calculate and store vectors for similarity. They also use an improved K-Means algorithm which limits the radius of a cluster. For analyzing preference and assign weights to clusters, they set higher weights to clusters which are further away from each other and fewer weights to clusters which are near. Their recommendation system uses these weights to model similarity to recommend music to users. Their system proves to provide very accurate recommendations for some particular generes.

Mandel and Ellis, in their paper (Mandel & Ellis, 2006), showed that using support vector machine (SVM) classifier with song level feature are better than using only SVM classifier. They also showed that use of Kullback-Leibler (KL) divergence for measuring distance is better than Mahalanobis distance over Gaussian model. They used 20 –coefficient MFCCs as suggested by Aucouturier and Pachet (Pachet & Aucouturier, 2004) and a “bag of frames” model for modeling the MFCCs. After extracting MFCCs, the mean and covariance of them are described in a Gaussian model. For distance measurement, they used both Mahalonobis and closed form of KL (Penny, 2001) divergence. They conducted experiments with the combinations of single Gaussian, GMM with KL divergence. They also compared artist-level versus song-level feature and SVM versus non-SVM classifiers. They showed that the use of single Gaussian is an effective way of capturing and comparing song-level feature (MFCC).