Can human cluster data sets manually? For example, consider the Iris data set, depicted below:

enter image description here

Instead of using clustering algorithms like connectivity-based clustering (hierarchical clustering), centroid-based clustering, distribution-based clustering, density-based clustering. etc.

Can a human manually cluster the Iris dataset? For our convenience, let us consider it as a two dimensional dataset. By which means and how a human would cluster the dataset?

  • 1
    $\begingroup$ Yoy might want to read about family resemblance concept. $\endgroup$
    – ttnphns
    Mar 15 '15 at 6:30
  • $\begingroup$ In Weka there is iterative UserClassifier function that is tree based clustering algorithm that asks user to define cluster boundaries. $\endgroup$
    – Tim
    Mar 15 '15 at 9:41
  • $\begingroup$ I see 3 clusters here, probably because of the labels for the 3 different colors & symbols. If they weren't labeled (& w/ different colors & symbols), I suspect I would see 2 clusters in that plot. $\endgroup$ Mar 15 '15 at 15:58
  • $\begingroup$ @gung what are the clustering algorithms that are closest to the human clustering or how humans would cluster the dataset? $\endgroup$
    – Parashara
    Mar 15 '15 at 16:01
  • 1
    $\begingroup$ I'm voting to close this question as off-topic because this is about how humans cluster perceptually. $\endgroup$ Mar 15 '15 at 16:06

Certainly, humans can cluster data sets, but those of limited size and complexity. In fact, we do that all the time in our everyday life. For (a very simple) example, when we stay on a busy street and want to hire a taxi cab, we look around and perform visual clustering of all nearby vehicles that we think are taxi or might be taxi. We can come up with a large number of similar examples, but I'm sure that you get the idea.

How do humans perform manual clustering, for example, of taxi cabs? I think that one, and a rather simplified, answer is this: by comparing certain attributes of [objects] vehicles around (such as special marking or color, sign on the roof, presence on a specially designated location) with our mental associations for a taxi cab (along the above-mentioned attributes). Those attributes are usually referred to as dimensions in statistics, data science, machine learning and other fields. The same principles apply to Iris data or any other data sets that are feasible to be clustered manually. I hope that my explanation is clear enough and close to what you've expected.

UPDATE per comments [my thoughts on relation between human clustering and statistical models and clustering algorithms - note that I'm not a cognitive scientist, so take them with a grain of salt]:

Then what are the clustering algorithms that are closest to the human clustering or how humans would cluster the dataset? – Parashara

I think that it all very much depends on what level and from what perspective you want to analyze the phenomenon (of human clustering). For example, from the application (functional) and/or statistical perspectives, I would say that human clustering resemble decision trees or, more accurately, random forest algorithms and, perhaps, some Bayesian approaches, such as hidden Markov model and its cousin hierarchical hidden Markov model. From a structural physiological perspective, I would place human clustering in a neural networks category, for obvious reasons. From a comprehensive approach perspective, most likely, the best (or one of the best) category for the human clustering is hierarchical temporal memory model, which is based on a combination of Bayesian models, spacial and temporal clustering and neural networks. I hope it makes sense.

  • $\begingroup$ then what are the clustering algorithms that are closest to the human clustering or how humans would cluster the dataset? $\endgroup$
    – Parashara
    Mar 15 '15 at 15:59
  • $\begingroup$ @Parashara: Please see my update, which hopefully addresses your question in the comment above. $\endgroup$ Mar 15 '15 at 23:40

Herman Chernoff invented long ago Chernoff faces which address precisely your problem. He observed that humans are particularly able to tell apart other people looking at their faces, and proposed to code up to 18 variables (as I recall) in a single face.

In his original JASA 1973 paper he shows an example in which a human can do a very good job at classifying fossils just by looking at those faces.

  • 4
    $\begingroup$ I've read numerous discussions of this method over 40 years. Most I've seen credit Chernoff with a clever idea but deny that, or dispute how far, it really works in practice. But either way, this has the question backwards. If the Chernoff method works at all, it's because it exploits how people identify similarities. That doesn't explain how they do it. Also, even if the Chernoff method is thought by its users to work, the question is still open on how far they are identifying genuine clusters. $\endgroup$
    – Nick Cox
    Mar 15 '15 at 13:17

Not the answer you're looking for? Browse other questions tagged or ask your own question.