Classifying Music & Emotion through different activities
AFFECTIVE CLASSIFICATION METHOD
USER
Music listeners
DURATION
4 months
Role
In order to understand how users classified music, I designed a hypothesis and produced a detailed research study to prove my hypothesis. I published this paper at HWWE (Humanising Work & Work Environment) 2016.
RESEARCH Objective
The goal of this paper was to explore the correlation between music and emotion in users defined through activities.
WHY?
As we are speeding into the future, definitions and relationships are constantly updating. Along with this shift, it's inevitable to see that change reflected in human lifestyle.
Using activities, to define the relationship gave more context to users' classifications. It also presented the potential to convert the findings into a unique recommendation/retrieval algorithm.
RESEARCH METHODOLOGY
33 unique users, all aged between 21-27, and students at MIT - ADT University. The sample size had a potential 15% margin of error at 90% confidence level. You can read more here.In a controlled testing environment, users completed the audio survey that entailed assigning an activity to the respective audio clip. Following this, users completed a SAM (Self Assessment Manikin) that defined the valence and arousal of the activity. I chose to collect this data with the following combination of research tools & data models;
Using activities, to define the relationship gave more context to users' classifications. It also presented the potential to convert the findings into a unique recommendation/retrieval algorithm.
SAM (Self Assessment Manikin)
Audio Survey
Circumplex Model
Radial classification charts
A self assessment mainikin that was used in the user survey. The top part of the mainikin is to measure arousal and the bottom for valence.
INSIGHTS
Users' responses reflected that they have activity driven intent when classifying music.
Users often added the word 'mood' in their survey responses to describe an emotion, when they were in fact describing an activity.
A majority of users assigned a combination of emotions to a single piece. The same applied to emotions assigned to activities.
Valence and affective arousal were important factors in sorting the audio logically. This helped in understanding the intensity of the user's emotion in relation to music.
Defining this relationship showed more specific and unique classification types of music.
This is a compiled radial classification chart. This was a collation of all prior radial charts that show collective trend and affinity.
RESULTS
There were two key metrics recorded that dictated the results. The mood of music and the mood of the activity; a scatter plot that proved a crucial detail. The connection between the users’ inferred moods from music, and selected activities.Using this data, correlating the song mood and the activity mood was effective. It even aided in determining the collective mood, as defined by the circumplex model. This proved that users could define their moods as activities coherently.
In conclusion, this research method and data interpretation has multiple real-world applications. This would apply to music classification and recommendation systems. It could prove useful in aiding user music discovery based on specific activities. It could even help to curate logic based playlists, defined by several unique attributes.
A selection of radial charts of the music genre classification charts. This is a plot of mood & genre for all users grouped by track.
Circumplex Models with the final results plotted to represent the correlation between music and emotion.
LEARNING & IMPROVEMENT
With classification such as this,one could document moods in correspondence to different unique categories. It's also possible to have a system that will store feedback and create a tailor made experiences. This model could potentially predict user’s mood habits. This could aid temporal scheduled music recommendation.
Given the opportunity, I'd make the following improvements to this research to produce a better data set;
Barring physical limitations, it is important to replicate this experiment across varied demographics.
The selection of music surveyed would be more vast and varied to cover all bases.
The audio list would contain control pieces as well as more challenging pieces to widen user definition.
