I'm aware of the Wikipedia article, but it is unnecessary dependent on economics jargon (eg. the explanation assumes that you know what the "determinants of wage offers" are).
Can anyone provide a more intuitive explanation, particularly as it might relate to correcting training data for a machine learning algorithm?
Specifically, can Heckman correction be applied when a supervised learning algorithm is responsible for choosing the samples which are used to train future versions of that supervised learning algorithm?
As a concrete example: We must choose between 100 ads to show people when they visit our website, and our goal is to maximize the number of people who click on the ads.
We can train a logistic model to predict the probability that a user will click on any given ad, and then run this model on all 100 ads for each visitor to our site, showing them the ad that has the highest probability of a click.
But now when we collect the results of these ad recommendations we have a biased sample, the bias being determined by a previous iteration of our supervised learning algorithm.
Can Heckman correction be used to correct this? Are there other techniques that I should investigate?
[link] [17 comments]