Building a Roadmap of Human Emotion: THE BIG 5

  • August 31, 2012
  • Blog
  • 7 Comments
Download PDF

Emotions are complicated, unpredictable and deeply personal. We feel them; we usually don’t think we can quantify them. Some people even claim that emotions are completely outside the sphere of what can be understood. After all, emotions are not perceived with the mind but instead felt with the heart, right?

But if we hope for technology that is more sensitive to human needs and interactions, we will need a more structured understanding of emotions. At Kanjoya, we’ve built a dataset that can finally take us there. For the first time, we’re uncovering the structure that underlies our disparate emotional experiences, and we’re doing it democratically and empirically — by learning from what real people really feel.

In a series of posts starting now we will walk through several insights we discovered while analyzing our dataset of emotion updates. In this first post we start by showing you how we discovered the big five human mindsets (not to be confused with the big five personality traits) and how your current emotion can predict your future.

The dataset

The dataset consists of emotion updates since October of 2006. When users sign in to the Experience Project they can set their mood periodically. The updates include an emotion selected from a list of 175 emotions, an intensity level ranging from 1 to 5 and a brief description explaining why they feel this way. Users use this feature to show their friends how they feel. To give a sense of the magnitude of the dataset, the number of updates used in this analysis is over 1.3 million and the total number of users that updated their emotion at least once is over 290,000. This specific dataset is great for this kind of statistical analysis and machine learning. However, the complete dataset we work with is thousands of times larger.

For simplicity we will refer to all of the options in this dataset as emotions. Keep in mind, however, that the options belong to the broader category of human states of consciousness. [1]

Grouping Emotions:

Discovering the Big Five Human Mindsets

We all know intuitively that “joyful” and “happy” are more similar to one another than, say, “joyful” and “sad”. But how does our dataset reflect this intuition? When an user posts a “happy” update briefly after posting a “cheerful” update, we know that in one instance “cheerful” transitioned to “happy”. When this happens tens of thousands of times, we can build a pretty good understanding that “cheerful” and “happy” are related. A given transition is more telling when it is fast, so by weighting the transitions according to the time elapsed, we are able to find how likely any given transition is within any interval of time. When we do this for every pair of emotions, we get a roadmap of human emotions — a network that shows the empirical likelihood of transitioning from any emotion to any other emotion.

Think of this graph as an actual blueprint or roadmap: If you are feeling “doomed” it is not very realistic to expect feeling “blessed” a few minutes later. However, the road from horror to bliss does exist –it just passes through other emotions first. To get from “doomed” to “blessed” you generally need a few stepping stones such as “okay”, “calm”, “tired”, “happy” and “cheerful”, in that order. An entire process of emotional recovery might take time, but the roads that can take you there are indeed very real.

With this roadmap of human emotions, we can do cool things like grouping them meaningfully. We used an algorithm [2] that finds communities in networks like this one and we discovered that, time and again, five clearly defined clusters emerged. The nice thing about encoding emotion transitions in this way is that a cluster of tightly connected emotions has a very concrete interpretation: These groups reflect regions of the emotion roadmap where one is likely to stay for a while.

To grasp what a cluster means in this context, think of emotions as cities and emotion clusters as countries. Traveling between cities in a country is common and easy; traveling between countries is way more difficult and it generally happens only along specific routes.

This is the roadmap of human emotion. It visualizes the common transitions between emotions. The size of the nodes is proportional to the overall popularity of the emotions and their colors are given by one run of the community detection algorithm. For the final coloring we used in the rest of our analysis we ran the algorithm several times and used the most common groups found. The edges are weighted according to the probability that a transition will happen, and their colors are determined by the color of the emotion of origin (e.g. if the edge is from a blue node to a green node the edge will be blue).

These clusters can be described with five color labels. This is what we call the big five human mindsets:

  • Green → Positive/Blissful/Energetic
  • Blue → Low energy/Depressed/Sad
  • Red → Anxious/Angry/Scared
  • Orange → Physical discomfort/Neutral/Circumstantial
  • Yellow → Sexual/Aroused

While these five clusters we found are surprisingly intuitive in hindsight, the grouping is done with a blind mathematical technique. The algorithm doesn’t know anything about humans, how emotions feel nor the words used to describe them. And in spite of these limitations, it still recovers our intuitions. Results like this are what make us stop and wonder whether our assumption that emotions cannot be understood formally is correct.

Another interesting result we found was that the big five human mindsets are universal to all subpopulations. Repeating the analysis with demographic-based subsets of the users delivered almost identical results. From teenage boys to old women, the big five human mindsets always emerge as distinct regions of the human emotion roadmap. If there is anything we all have in common, this might be it; we all share the same underlying emotion blueprint.

Fortune-telling using your current emotion

Detecting groups of emotions is one thing, but can we tell how representative each emotion is of the group it belongs to? After all, not all emotions are created equally. If we look at all of the generally positive (green) emotions, we can see that experiencing some of them usually indicates that there are solid positive things going on in your life (e.g., blessed) while others reflect a more circumstantial or transient state (e.g., chipper). It is not the same when your friend tells you he/she is feeling “chipper” versus “blissful”: While both emotions are positive, they tell different stories and anticipate different futures.

To capture this distinction, we looked at how people felt throughout the two weeks following the day they reported experiencing a given emotion. By adding up the average volume of each emotion color in these two weeks, we identified the moods that are the most powerful and long-lasting –e.g. the happy moods that predict that you’ll be on more green emotions in the future and the sad moods that predict that you will be experiencing more blue emotions. The following list summarizes the top ten emotions representative of each cluster.

The top ten stickiest emotions for each color cluster.

We call these emotions “Sticky Emotions” because experiencing them predicts that the current general mindset will be longer lived than it would be otherwise. To see what this actually looks like for a specific emotion, let’s look at the effect that feeling “blessed”, “enraged” and “depressed” has on your future emotions. In these graphs, the size of the nodes is proportional to the probability of experiencing the given emotion the day after feeling the central emotion. Note that the size of the center node itself varies; this is because the probability of experiencing the same emotion two days in a row is different for each emotion.

The size of the nodes is proportional to the probability of experiencing the emotion one day after feeling blessed. The colors represent the cluster the emotion belongs to.

Probability of experiencing each emotion one day after feeling depressed.

Probability of experiencing each emotion one day after feeling enraged.

The previous analysis shows how current emotions predict the overall amount of future emotions of each color. But what about overall wellbeing? We can also find the overall best and worse emotions by adding up the results for each color. By assigning to each color a general emotional valence (how positive or negative emotions in the cluster are) we can find the future average valence a person will experience in the next two weeks after experiencing a given emotion. In this way, we can rank all the emotions according to how much emotional wellbeing they predict in the near future. Here are the top ten and bottom ten emotions:

Best predictors of emotional wellbeing in the next two weeks: blessed, delighted, blissful, festive, horny, jubilant, courageous, cheerful, confident, grateful.

Best predictors of low emotional valence in the next two weeks: depressed, pessimistic, lonely, numb, negative, tearful, upset, rejected, doomed, okay.

We are barely scratching the surface

This analysis and other work is slowly but surely helping us build a computational model of how emotions work. This understanding will allow us to create technologies that are more responsive to human wants and needs. In addition, by understanding the underlying structure of emotion transitions, we will get better at navigating the emotion roadmap ourselves. While lows and highs are both part of life, with a better understanding of the laws that govern emotion transitions, we can at least make sure the lows never have to last a lifetime.

Next week we will explore what causes people to change their current mindset. Just to give you a taste of next week’s post, we will uncover which emotions predict mindset changes and what are the specific events, actions and circumstances that trigger them. If the big five human mindsets are countries, next week we will tell you how the passport looks like (and hopefully how to get one).

Credits:

The entire Kanjoya team
Graphs generated with Gephi.

[1]  Some of the options are better described as moods or bodily feelings such as “tired”, “hungry” and “sleepy.”
[2]  Vincent D Blondel, Jean-Loup Guillaume, Renaud Lambiotte, Etienne Lefebvre, Fast unfolding of communities in large networks, in Journal of Statistical Mechanics: Theory and Experiment 2008 (10), P1000
7 Comments
Sounds like a very worthwhile project. Emotions are of great importance--their function, after all, is to motivate--but until recently, they have received hardly any scientific attention.
Gordon Cornwall
Fascinating work! Look forward to next weeks post.
Nifty. However, the significant appearance of 'blessed' suggests to me that your sample is likely ethnocentrically biased by being completely US based. As a dual Canadian/US citizen who's also lived in Mexico, the only place I've *ever* heard people regularly describe themselves/their mood as "blessed" is in the US.
Tyler Style
And what about months and years of long-term equalization, where the story-telling that we are so happy becomes a counter-productive confabulation?
Pierre Rousseau
I love this research, and as a pastoral student and psych grad this *will* be of IMMENSE value to the field of counseling of all types. I have a big critique, however. Your conclusion of five categories appears to be both arbitrary and highly subject to cultural and linguistic variation, as well as systematic bias on the part of the researchers. I would like to see the data processed across cultural and linguistic groups, and expanded to be phrased and parsed by a multi-lingual team to help mitigate the internal biases toward linguistic "pre-grouping".
WarrenSensei
I have not enjoyed a study on emotions as much as I have enjoyed this one. Surprisingly, I wrote several presentations on this subject unaware of this great research. For example, almost three years ago I wrote a presentation on Social Network Analysis Of Intangibles. The synopsis reads as follows: This presentation covers the application of Social Network Analysis in the study of intangibles such as emotions and motivations as these are the real drivers of organizational performance http://www.slideshare.net/hudali15/social-network-analysis-of-intangibles
ali anani
I suspect religious bias, and would like to see some more rigorous data and info to establish credibility.
secret agent girl

Leave a reply

Your email address will not be published. Required fields are marked *

You may use these HTML tags and attributes: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <strike> <strong>