EAMR: An Emotion-aware Music Recommender Method via Mel Spectrogram and Arousal-Valence Model

For music platforms, the cold start of new songs is a big challenge for recommendation systems. Most existing methods adopt music content (i.e., the audio signal) to alleviate this problem. But these approaches typically overlook the emotional characteristics embedded in the music, which is particularly important for modeling the characteristics of the songs. Considering the existence of negative and positive classifications of songs and the existence of alternating processes of calmness and excitement in song melodies. We propose an emotion-aware music recommender method (EAMR) via Mel spectrogram and the arousal-valence model. First, each audio clip is transformed into a log-Mel spectrogram, and a deep neural network extracts the emotional features. Meanwhile, each audio clip is dimensionally reduced to a 16-dimensional feature vector normalized to obtain each sentiment feature vector of songs. Second, the Euclidean distance of the song feature vector is calculated to recommend the song with the highest similarity to the user's current favorite. Finally, taking into account the impact of features extracted by different convolutional kernels on the model, the ability of model feature extraction is further improved by introducing an attention mechanism called Att-EAMR. The proposed methods are tested on the public Spotify Million Playlist dataset and the free music archive dataset FMA. Experimental results demonstrate that the proposed methods outperform the other comparison approaches when the system faces a cold start.