News for hackers. Stuff that matters.
Re:Datasets for Machine Learning by bhnmmhmd on Thursday June 14, 2018 @04:48AM (#17310321)
I've heard that Kaggle data sets encourage people to do "supervised" ML only. Is that true?
Re:Datasets for Machine Learning by benhamner on Thursday June 14, 2018 @08:05AM (#17310986)
The competitions we host (https://www.kaggle.com/competitions) are supervised and always have a target we can create a numeric leaderboard on, but the public datasets (https://www.kaggle.com/datasets) are used for everything under the sun.

There's some supervised ML use of those, and a lot more open-ended exploration, visualization, cleaning, clustering, language modeling, etc.

Re:Datasets for Machine Learning by stuartaxelowen on Thursday June 14, 2018 @06:19AM (#17310592)
Not at all - I released a customer support on Twitter dataset there specifically focused on unsupervised tasks! I think the focus on supervision in what people do with the data shows that there are still a lot of people poking around with the easier supervised tasks.

[0]: https://www.kaggle.com/thoughtvector/customer-support-on-twi...

Re:Datasets for Machine Learning by hideo on Thursday June 14, 2018 @05:20AM (#17310428)
(Not Ben, but - ) outside of academia, the main thing that seems to encourage people to do supervised ML is that it's the only thing that seems to work. I haven't really heard of any success stories with using unsupervised techniques for most common ML applications.
Re:Datasets for Machine Learning by laichzeit0 on Thursday June 14, 2018 @07:15AM (#17310778)
I'm not an expert, but I feel that:

Unsupervised techniques work really well for language modelling.

There is also weakly supervised and distant-supervision, where the labels are "noisy" or not exactly what you want.

You're right in that strong supervision, where you basically trust your class label, works really well, because it's probably the easiest case.

Combining unsupervised (e.g. pre-trained language models) with a very small set of strongly labeled data, or a larger set of weakly labeled data, seems to work pretty well too.

Re:Datasets for Machine Learning by raverbashing on Thursday June 14, 2018 @06:19AM (#17310590)
Unsupervised works, but your ability to measure "does it work or not" is much more dependent on a case by case evaluation rather than a score.

(Because if you know a priori what is it that you want to measure - it's supervised)

Re:Datasets for Machine Learning by atupis on Thursday June 14, 2018 @07:56AM (#17310948)
Yeah this my experience too, evaluation ends being almost endless time sink.
Re:Datasets for Machine Learning by dotancohen on Thursday June 14, 2018 @06:05AM (#17310552)
I used a very simple unsupervised ML built in scikit-learn to find good matches on OK Cupid. Worked very well, it found definite boundaries between the clusters of women.

One of the features was a subjective rating of how much I liked some of the women, and scikit-learn then suggested to me other women in the clusters that had my best ratings. It turns out that I like vegetarians, redheads, and left-wingers. Which happens to be true, even though I eat meat and do not identify as left-wing. But those traits correlate with _other_ traits that are more difficult to measure objectively, such as caring about children, liking to hike, and preferring an evening of sex to an evening of television.

Re:Datasets for Machine Learning by taneq on Thursday June 14, 2018 @07:42AM (#17310886)
I think it's more that supervised ML is sufficient for most of the low hanging fruit. It's relatively easy and well-understood, and there are a lot of things out there where we have copious data that we just need to digest into a model to make it useful.
Re:Datasets for Machine Learning by harias on Thursday June 14, 2018 @05:36AM (#17310469)
What about clustering?