In this paper, we describe the methods designed for extracting the a ective features from the given music and predicting the dynamic emotion ratings along the arousal and valence dimensions. The algorithm called Arousal-Valence Similarity Preserving Embedding (AV-SPE) is presented to extract the intrinsic features embedded in music signal that essentially evoke human emotions. A standard support vector regressor is then employed to predict the emotion ratings of the music along the arousal and valence dimensions. The experimental results demonstrate that the performance of the proposed method along the arousal dimension is signi cantly better than the baseline.
The Emotion in Music task in MediaEval 2015 Workshop aims to detect the emotional dynamics of music using its content. Speci cally, given a set of songs, participants are asked to automatically generate continuous emotional representations in arousal and valence. More details of the task as well as the dataset can be found in [1].
Feature extraction, which aims to discover the intrinsic factors while capturing essentials of original data according to some criteria, plays an important role in music emotion analysis. Some algorithms have been proposed to learn the genuine correlates of music signal evoking emotional responses. You et al. presented a multi-label embedded feature selection (MEFS) method for the task of music emotion classi cation [12]. Liu et al. introduced an algorithm called multi-emotion similarity preserving embedding (ME-SPE) by considering the correlation between di erent music emotions, and then analyzed the relationship between the lowdimensional features and the music emotions [6]. In this paper, we propose a feature extraction algorithm, arousalvalence similarity preserving embedding (AV-SPE), which inherits the basic idea from ME-SPE. The di erence is that the emotion labels in ME-SPE are in the binary form, i.e., 0 or 1, while those in AV-SPE could be any real number between [ 1; 1].
In order to learn the relationship between the feature space and the dimensional emotion space, which is composed of the arousal dimension and the valence dimension, many popular machine learning approaches have already been employed to train the model, such as k-Nearest Neighbor [11], Support Vector Regression [9], Boosting [7], Conditional Random Fields [2], and Gaussian Process [8]. In this task, we employ the -Support Vector Regression ( -SVR) [10] to predict the arousal and valence labels of the music.
In the remaining part, we rst introduce the methods used for feature extraction and label prediction in Section 2. In Section 3, we report the evaluation results. Finally, we conclude the paper in Section 4. 2. 2.1
In order to discover the intrinsic factors in music signals that convey or evoke emotions along the arousal and valence dimensions, we propose a supervised feature extraction algorithm dubbed AV-SPE to map the original high-dimensional representations into a low-dimensional feature subspace, in which we hope that a clearer linkage between the features and emotions could be discovered.
Let x 2 RD be the high-dimensional feature vector of the
music at a certain time point (in our speci c task, D = 260),
and y = [y(
U
n n = arg min X X kUT xi
U i=1 j=1
UT xjk2 Sij; where Sij = hy^i; y^ji = hyi=jjyijj; yj=jjyjjji denotes emotional similarity between xi and xj (i; j = 1; :::; n).
Following some standard operations in linear algebra, above optimization problem could be reduced to a trace minimization problem:
U = arg min tr(UT XLXT U);
U
where X = [x1; x2; :::; xn] 2 RD n is the data matrix, L =
D S^ is the n n Laplacian matrix [3], D is a diagonal
(
j=1 Sij (i = 1; :::; n), and tr( )
denotes the matrix trace operator. Obviously, L is positive
semi-de nite and D is positive de nite. By transforming
(
XLXT u =
XDXT u:
Similarly, u2; :::; ud are the eigenvectors corresponding to
the 2-nd, ..., d-th smallest non-zero eigenvalues of (
After feature extraction, we can obtain the reduced
features by zi = UT xi. We then use the reduced features
as the input to predict the emotion labels of the music via
the -Support Vector Regression ( -SVR) [10]. Given the
training set f(z1; y1); :::; (zn; yn)g, where zi is the extracted
feature vector and yi = [yi(
( 2 s:t: eT ( )T K( ) + (y(m))T ( ) ) = 0; eT ( + )
C
(
C=n; i = 1; :::; n; where i; i are the Lagrange multipliers, K is an n n positive semide nite matrix, in which Kij = K(zi; zj) = (zi)T (zj) is the kernel function, m = 1 or 2, e = [1; :::; 1]T is the n-dimensional vector of all ones, and C > 0 is the regularization parameter. The prediction label of a new coming vector z is:
n
y = X( i
i=1
i)K(zi; z) + b:
(
In this section, we report the experimental settings and the evaluation results. The features used in our experiments are extracted via the openSMILE toolbox [5]. The original dimension of the feature space, i.e., D, is 260. We set the reduced dimension d = 10 for AV-SPE. The SVR toolbox we have used is the LIBSVM [4]. In the training process, we use the radial basis function (RBF) as the kernel function. The ten-fold cross-validation is employed to select the best parameters and C. We nally select = 0:125, C = 2, and = 0:5 for our model.
Table 1 and Table 2 list the averaged Root-Mean-Square Error (RMSE) and the averaged correlation, respectively. From the tables, we can observe that the arousal result of SVR and that of AV-SPE + -SVR are signi cantly better than the baseline. Moreover, the results on the reduced feature space (d = 10), i.e., the results of AV-SPE + SVR, are comparable to the results on the original feature space (D = 260), which indicates that the extracted features play an important role in representing the music emotions. 4.
In this working notes paper, we have introduced our system for music emotional dynamics prediction. The system is composed of a feature extraction algorithm and a support vector regressor. The evaluation results shown that the features extracted by the proposed AV-SPE are informative, and the system worked well in predicting the arousal values. Our future work will focus on extending the proposed algorithm by considering the dynamic nature of the music data. 5.
The authors would like to thank the reviewer for the helpful comments. This work was supported in part by the National Natural Science Foundation of China under Grants 61373122. 6.