Talk:Winnow (algorithm)

Removed the "insufficient context" tag

The tag was placed on 2007-10-29T16:33:23 by user Kallerdis, but the article is much improved since then. I think it does provide enough context. -Pgan002 (talk) 18:31, 22 October 2010 (UTC)

Is that update procedure correct?

The text specifies the update procedure as:

• If an example is correctly classified, do nothing.
• If an example is predicted to be 1 but the correct result was 0, all of the weights involved in the mistake are set to zero (demotion step).
• If an example is predicted to be 0 but the correct result was 1, all of the weights involved in the mistake are multiplied by $\alpha$ (promotion step).

Step 2 moves weights to zero, and neither of the other steps changes any weight that is zero. Because of that, I fear that this will too easily migrate towards an all-zero weights vector.

So, either explain why that will not happen, or adjust the description. —Preceding unsigned comment added by 145.36.235.3 (talk) 11:30, 19 April 2011 (UTC)

That is indeed the original (Winnow1) algorithm. Roughly, any feature that is implicated in a false positive should thereafter be ignored. I think the weights cannot all go to zero if the data is linearly separable: there will always be some instance that on the other side of the current hyperplane. And if the data is not linearly separable, all bets are off, just as with most linear classifiers. —johndburger 18:09, 6 October 2011 (UTC)

Name of the algorithm

…that allows it to perform much better when many dimensions are irrelevant (hence its name).

I don't get the reference. If it's not just me being dumb, it might be good for someone who knows to add a short explanation on how this property leads to the algorithm's name. 2A02:120B:2C61:19F0:BC77:22BD:D7D6:59E8 (talk) 22:08, 24 November 2016 (UTC)

I added a link to the Wikipedia page on winnowing, as I was just as confused as you initially. 176.10.165.46 (talk) 11:24, 26 November 2018 (UTC)