I have time series data of 60 different emotional states that I’ve recorded daily for about 6 years. I’ve explored a few methods for classifying these states, most recently with PCA, or principal component analysis. My goal is to simplify the 60 states into a smaller number of components or dimensions that describe my emotional state. In this case, 25 components described 80% of the variance in the data. There was no labeling of the data, or subjective rule-making, only the PCA determining the most important vectors that describe the data; starting with the first two components, PC1 and PC2.
If I had to subjectively describe what these components are, I would say PC1 is a vector of positive/negative, pleasurable/unpleasurable, or simply happiness/sadness. PC2 looks like activity/inactivity, arousal/calmness, or agitation/steadiness. There’s a cluster of physical states in the middle that are mostly neutral. This is compatible with a few modern dimensional models of emotion. Wilhelm Max Wundt described three dimensions in 1897: "pleasurable versus unpleasurable", "arousing or subduing", and "strain or relaxation". In 1954, Harold Schlosberg named three dimensions: "pleasantness–unpleasantness", "attention–rejection" and "level of activation". The first two dimensions of my PCA seem straightforward. Let’s see PC3 and PC4:
This one is not so obvious. Maybe strain/relaxation fits PC3, maybe agency/disinhibition. PC4 might be something like narrow focus vs. openness/observation/discovery.
Plotting all 25 principal components over time gives a nice heatmap. A lot of highlights are familiar to me as I think back to what was happening at the time of each event. Many of the components around PC5-PC25 are kind of mysterious. They’re either a real cluster of emotion that can’t be succinctly described, or maybe a lot of noise.