Abstract:
Most of the conventional artist similarity models adopted the artist's label-based similarity to represent the artist's similarity, which is not accurate and comprehensive. In order to cope with the above problem and enhance the performance of artist similarity model, this paper proposes an artist characteristic representation model by means of deep semantic mining of the artist's profile. The proposed model performs word embedding on the artist's profile with Word2Vec and GloVe, respectively. The contextual semantic dependency of the obtained word vectors is further extracted by using convolutional neural networks(CNNs). Finally, fully connected network is utilized to judge the effectiveness of the representation of the artist's characteristic. Specifically, we crawl artists' profiles from the website Pandora. The genre tags of corresponding artists are obtained from Last.fm API. All artists' profiles are used to construct a training set of word embedding, by which Word2Vec embedding model and GloVe embedding model are trained, respectively. Each word in the artist's web profile is mapped to the word vector by Word2Vec and GloVe embedding techniques. Thus, the complementary between Word2Vec and GloVe are full considered in the proposed model. We use convolutional neural networks with different convolution kernel heights (3,4,5) to compress each type of word vector matrix. In this way, the obtained word vector contains contextual semantic information and its dimension is also reduced. Lastly, the two compressed word vectors are concatenated into one vector, which is just the artist's characteristics representation. To verify the effectiveness of the proposed model, a fully connected network is added to evaluate the classification accuracy achieved by the artist's characteristics representation. Adam optimizer is utilized to train the proposed model. In the experiment, 5 101 artists' profiles provided by Pandora are adopted as dataset, based on which the performance of the proposed scheme is tested and compared with conventional schemes. It is verified that the proposed scheme outperforms conventional ones in classification accuracy.