Perhaps I’m making a mistake asking some actuaries this question, but we’ll see…
I’m dealing with a fish population that can be divided into a northern contingent and a southern contingent. For part of the year they’re spatially segregated and for part of the year they spatially overlap. We’re trying to develop a new test that can help us determine whether the fish is from the northern or southern contingents.
The difficulty is that we can easily get samples from the northern contingent when its isolated, but our sample collection on the southern contingent occurs when the two overlap.
This creates issues for tools like discriminant analysis, KNN, SVMs, etc. because it’s a struggle to get a clear idea of what the variability of the southern contingent metrics look like and try and seperate them by maximizing the difference between the two contingents.
TL;DR answering the question is this a northern or southern contingent fish could be difficult.
Some classification tools are capable of working with a single class. Instead of asking is this a or b, you instead ask is this a or not a. I’m wondering if my better bet might be to fit a single class classifier to my data and then see if I can separate the northern contingent from the southern contingent aka not a.
Can you see any obvious problems with trying that approach?
That would seem like a reasonable approach, assuming you only have two types of fish. Otherwise, would you maybe ask ‘is this not A, but something that has been observed in B?’ as a means to avoid calling a C a B (insert your own joke there).
Also, you might check with Jian-Yang to see how he developed his hot dog identifying app.
The only other fish that could be relevant are based in Europe (but do show up in Iceland, so its plausible they could make it over, but hopefully not anytime soon…). So not too worried about that form of C. Another form of C could be other species of fish. I’m hesitant to stick them in as the level of difference from A to C is vastly bigger than A to B. I’m afraid using them as the outgroup(?) would affect the treatment of A so much that all of B would fit into A.
Eta: the way that this method works, we’d know that the the thing we’re looking at is indeed from the species of interest. So, if we’re seeing Cs, we’re expecting them. The problem is that the use of Cs makes the difference between As and Bs seem small.
I think if the prevalence of C is small, probably not a thing to worry too much about.
I was just thinking that if Cs are present, ‘not A’ does nothing to discriminate B from C. And since the B sample includes some As, then it might be worth sequentially asking:
Is it A?
If not, then is it similar to at least some fish in B? That might be tricky to define, I’m trying to basically say it’s not A, and my B sample includes SOME As, but mostly Bs.
And if it’s not A, and wildly different than any fish in B, I should exclude the fish.
I obviously am not well-versed in what you’re doing, I’m barely qualified to say things about breaker boxes.
I don’t think we’ll see Cs at all. If we did see Cs, they would likely be incredibly sparse. Testing has been done previously and Cs weren’t found to exist on this side of the Atlantic. On the other hand, we are dealing with fish and the population size is likely in the many billions. I’m sure a few individuals have done the implausible, but our odds of actually sampling them are pretty miniscule.
So, i think we can skip the testing for C.
What I’d like to do is maybe start with just testing for not A. If it seems like that’s working, then start sampling towards where we think we’d encounter B. Then as a phase 2 of this project, do the testing on the not As using genetics to verify they’re Bs. Then when I have enough verified Bs, use them to build my population of Bs so that I can go is this A or B? At that point, if I start getting neithers, I could test them genetically for Cs or expand the pool used to identify As and Bs.
I would add to Snik’s question one about just how variable is the Northern population in identifying them?
A relevant analogy I can see here is related to how they train folks to identify a counterfeit bill . . . basically, they train them to know the real thing to a very high degree of specificity.
Perhaps you could “artificially” train a model on the Northern population in the context of a bunch of other species (including samples from the Southern population) for the “A vs. ~A” dichotomy to gain confidence in the model.
What proportion of the commercial catch in the southern continent’s stock area consists of northern contingent fish. Further, in parts of the northern contingent area are we seeing any southern contingent fish.
I’m pretty sure I can avoid southern contingent fish in the northern contingent area by selective sampling of the samples I use to build the model.
The method I’m using is pretty new. I’m using near infrared spectral data collected from their ear stones or otoliths. I then reduce the dimensionality of the data using principal component analysis. When i do that with species level data and to a lesser extent with within species data, you can see clear separation of species and to a lesser extent stocks (e.g. regional groupings of a population). I’m hopeful I’ll see a similar pattern here which would give me confidence, but I’m still working on getting the samples.
We also use the method for fish aging, though the species in question here likely isn’t a good candidate for this methodology of aging due to their growth pattern. The growth of the fish contributes a lot to the variability. When I look at the spectra in PCA space, variability in growth is mostly showing on PC1 and regionality shows up on PCs2+. More simply, the bulk of the variability is due to age while regionality tends to be a smaller contributor to variability.
My difficulty is getting a pure sample of southern contingent fish. I’ll know I have a problem if all the fish from my samples from the southern contingent get counted as northern contingent fish using a model trained only on northern contingent fish. If that’s happening then I’ll have to use genetic testing or stable isotope analysis to identify which fish are which and switch to a two class classification tool.
There’s a climate change aspect where we haven’t looked at this in detail in awhile and there’s some evidence thst there may be more overlap than in the past.
Societally/politically/regulatory it matters because the stock straddles the Canada-US border and it affects how the fishery gets managed in the two jurisdictions. It might also affect quotas where theoretically one country has a greater access to fish from the other contingent than they previously thought.
You could limit it to fish that were only certain ages. With the stable isotope work, in the past they’ve limited it to fish of particular ages.
At this point, I’m just trying to do proof of concept work. If it seems effective then we can go looking at ways of optimizing this. I can get it to work for herring, but its iffy for cod. I’m hoping mackerel, the species in question, are more like herring than cod.
If I’m really lucky, it works well, and perhaps I get to help with European mackerel as well.
