Discovering Emotion-Inducing Music Features Using EEG Signals

Abstract

Music induces different kinds of emotions in listeners. Previous research on music and emotions discovered that different music features can be used for classifying how certain music can induce emotions in an individual. We propose a method for collecting electroencephalograph (EEG) data from subjects listening to emotion-inducing music. The EEG data is used to continuously label high-level music features with continuous-valued emotion annotations using the emotion spectrum analysis method. The music features are extracted fromMIDI files using a windowing technique. We highlight the results of two emotion models for stress and relaxation which were constructed using C4.5. Evaluations of the models using 10-fold cross validation give promising results with an average relative absolute error of 6.54% using a window length of 38.4 seconds.

An evaluation of feature extraction in EEG-based emotion prediction with support vector machines

“…Furthermore, it performs very well on handling the problem domains where the number of features exceeds the number of training examples. [4]:In these works, the data collection is divided into two phases. In the first phase, participants listened to a set of songs and, then, they gave each song a rating score between 1-5 for joyful, sad, relaxing, and stressful.…”

“…2) 10 channels:it is a subset of channels proposed in [4] including Fp1, Fp2, F3, F4, T7, T8, P3, P4, O1, and O2.…”

“…These bands follow the ranges used in the DEAP dataset [5]. Also, the delta band is removed following prior works [2,5,4,8] since it has proved that the low frequency band is not necessary for the emotion prediction task.…”

“…In this domain, there are many benchmark databases, e.g.,International Affective Picture System (IAPS) and International Affective Digital Sound (IADS) [3]. Many recent trials showed that human emotion can be induced by different kinds of stimulus (picture, music, and video) [2,4]. One of the most widely used benchmark is a database for analysis using physiological signals (DEAP) [5].…”

“…Some of them used 32 probes, while others suggested that 10 probes are sufficient. Also, there are many choices of EGG signal transformations: bandpower and Power Spectral Density (PSD) wavelet [4]. Since EGG signals are continuous,there are many levels of samplings, e.g., seconds, minutes, and statistics of minutes.…”