Collecting music emotion data is a very subjective process, since people have different understandings of one piece of music under different contexts and associate music with their memories and life experience. In the research field of music emotion, researchers often separate extra-musical factors with factors inherent in music that cause emotions to be perceived by listeners.
T. Eerola, in order to study the latter aspect, conducted an optimized fractional factorial study on 7 primary cues and collected emotional ratings by 46 participants on a total of 200 musical samples according to 4 percieved emotional characers (happy, sad, scary and peaceful). In his paper “Emotional expression in music: contribution, linearity, and additivity of primary musical cues”, Eerola argued that many primary cues contributed to emotion in a linear and non-interactive fashion. However, my exploration of this dataset yields a different result, contradicting his linearity of cue levels argument.
Fractional factorial design: A full factorial experiment is an experiment whose design consists of two or more factors, each with discrete possible values or “levels”, and whose experimental units take on all possible combinations of these levels across all such factors (wiki). If T.Erola is going to conduct a full factorial experiment, he needs to manipulate 14400 musical samples. Instead, he did an optimized fractional factorial study with 200 musical samples that he claimed could “allow the research resources to be concentrated on particular questions, thereby minimizing redundancy and maximizing the statistical power”. However, I doubt if some interaction effects are effaced in this reduction process.
- How do primary cues contibute to different emotional expressions in music?
- Can we predict both happiness and sadness rating for music given single set of primary cues?
- Limitation of T. Eerola’s original model
See detailed implementations, models and R code in Rmarkdown file/Poster.