Active learning for positive examples

I have a colleague who wants to look through large amounts of (text) data for examples of a pretty rare phenomenon (maybe 1% positive class, at most). We have about 20 labeled positive examples and 20 labeled negative examples. The natural thing to do at this point is some sort of active learning.

But here's the thing. We have no need for a classifier. And we don't even care about being good at finding negative examples. All we care about is finding as many positive examples from a fixed corpus as possible.

That is to say: this is really a find-a-needle-in-a-haystack problem. The best discussion I've found of this is Section 4 of Inactive Learning, by Attenberg and Provost. I came across a couple other papers that basically do some sort of balancing to deal with the imbalanced data problem in active learning, but the Attenberg and Provost results suggest that even this is tricky to get right.

But I think this only partially addresses the problem. Here's why.

Let's say that my colleague (call her Alice) is willing to spend one hour of her life looking at examples. Alice estimates that she can look at about 500 examples in that time period (small caveat: some examples are larger than other and therefore slower, but let's ignore this for now). Alice's goal is to find as many positive examples in that one hour as possible. Here are some possible strategies Alice could deploy

1. Look at 500 randomly selected examples, getting 5 (more) positive examples in expectation (likely somewhere between 2 and 12, with 95% confidence). This gives her a total of 25 positive examples
   
2. Train a classifier on her 40 labeled examples, have it rank the entire set (minus those 40). Then look at the top 500 examples. If the learned model has a recall of r% in the top 500, then she should expect to get 5*r more examples, giving her a total of 20+5r examples. (Caveat: this is a bit dangerous, because with very few training examples, results can often be anti-correlated with the desired output and you get better performance by flipping the predicted label. There was a paper on this maybe 5-10 years ago but I can't find it... Maybe somewhat knows what I'm referring to.)
   
3. Train a classifier, spend 30 minutes looking at 250 random examples, getting 2.5r more positive examples, then train another classifier, then spend 30 minutes looking at it's top ranked 250 examples. If the second classifier has a recall of r'% in the top 250, then she should expect to get another 2.5*r', and she'll have a total of 20+2.5r+2.5r' = 20 + 2.5(r+r') so long as r>r' this should be better.
   
4. Taking this to the extreme, Alice could annotate one example at a time (subject to constraints on either the learning being very fast or Alice not actually being a human). As long as the recall-at-one of the classifiers learned is non-decreasing, this should (I think) be the optimal strategy.

So it seems that in a setting like this, what the classifier should optimize for is simply recall-at-one. In other words, it doesn't care about anything except that it's top ranked output is correct.

Okay, so why is this difficult? Basically because we don't have that much data. If the corpus that represents the haystack is huge, then we want to ensure that the top-one example from that haystack is a positive example. So in principle even if all we're trying to do is select between two different hypotheses (say our hypothesis class at each time step has cardinality two) then in principle we should use as much data as possible to evaluate these two hypotheses. In particular, looking at the recall-at-one on the subset of 40 labeled examples that we have is probably not representative, essentially because this loss function doesn't decompose nicely.

So what could we do instead? Well, we could pool the labeled and unlabeled examples and predict on those instead. Certainly if one of the positive labeled examples outranked everything else, that would be a good thing. But if truly 1% of the unlabeled dataset is positive, then this is not a necessary condition. In particular, the top-ranked positive labeled example could be as far as 1% down the list and we could still (in principle) be doing a good job.

Ok I'll admit: I really don't know how to do this. Maybe it's enough in most cases to optimize for zero-one loss (or a weighted version thereof) and just cross your fingers that the recall-at-one is good. But it does feel like there should be a more direct way to go about this problem.