Data-driven Coral Reef Conservation (2026-US-30MP-2890)

Hey everybody, welcome. My name is Anderson Mayfield, and I'm a coral reef researcher and long-term JMP user. Over the next 30, 35 minutes I'm going to talk to you about some recent progress I've made on this idea of decision support for coral reef interventions. I'll define what exactly this means here in a second.

This has been something I really hope that… It might not be the most powerful showcase of JMP, but it's something where I think the biological and environmental implications might actually be more of the selling point. I'll also talk about some more hardcore use cases of JMP later that I may not get to today, but for people that are more advanced users and want to see some maybe more compelling uses of JMP, I'll send you off in that direction. This is a little bit more of a layman's talk or maybe something that will be more useful for newbies.

The idea is basically this: we know coral reefs are in bad shape across the world. It's mainly climate change-related, increases in seawater temperature. But there's local things like disease that we're dealing with now in the Caribbean, so local-scale insults. Most people that are actually doing the work, the good work in the water, are not necessarily scientists. They're just passionate people that want to do something good. That means they might lack knowledge and expertise.

Different reef conditions and different problems might warrant different solutions. Depending on your reef, the corals that live there, the nature of the stressors, you might have an array of what we call interventions or solutions that you might want to try. These are things like giving the corals medicine, improving the seawater quality, cooling it off, establishing a marine park to help the fish.

How do you know which intervention or solution is going to give you the most bang for your buck? This is what I'm calling decision support. We're taking the data from the reef, making a decision about it to try to find the optimal intervention. This is a little bit of a tangent, but I wanted to throw it out there because I think it's become maybe a little bit of a contentious issue amongst statisticians and AI people. They might feel like they're at odds. But actually, if you're familiar with a software package like JMP, and you're maybe not AI savvy, but you're somebody who uses it regularly, you're a smart AI user, let's say, you can actually do a lot of really good stuff.

That's going to be one of the other messages percolating through the background of this talk. But there's the caveat. Would I use exclusively JMP for most of the analytics I'm going to show you? I'd have no qualms doing that. Would I like to incorporate AI stuff to speed up parts of the process that I don't enjoy? Definitely. Where I think people get into trouble is where you don't have any analytical foundation in your models, and you're jumping past the modeling and the statistics to the AI. That's where you're getting into this realm of hallucinations and all these the bad side of AI.

I'm going to hopefully, by showing you what I've done over the last couple of months, try to sell this idea. Unfortunately, I really thought I was going to get to the punch line over the course of this talk, which is, can my JMP models that I've been building over the last few years do better than a scientist? I set it up like this. It was going to be me versus ChatGPT, that might have a wealth of data at its disposal, but it's prone to hallucination, obviously never been in the water.

There's an organization that actually I'm wearing their shirt right now. I have a lot of respect for them. Their approach is different. Their approach is, we don't care what the coral reef looks like, we only have one recommendation: it's always to restore it. That's one end of the spectrum. You don't care about the data, you just recommend one thing. They're at peace with that. They wouldn't be annoyed that I've said this.

Then we've got what I'm hoping will be the gold standard, is a properly trained and validated JMP model. We're going to try to run this live during this talk, but I didn't quite get to where I wanted to be. That's okay. Hopefully, if you check out my talk at JMP Discovery in person in Baltimore or come back to this page as I'll be routinely updating the content, you can actually see the final results of this comparison because I think it's going to be really cool.

Scientists, we like data, but sometimes people go with their gut. Can we do better than going with our gut? Can we have the data dictate the optimal intervention? This is what I mentioned for people that want to hear something totally different using JMP to really get into the weeds of exploring text data.

Huge bodies of molecular data like gene sequences, proteomic data, things like this. I've got a book chapter on this coming out soon. If you want a really light version of what I'm going to be talking about over the next few minutes, I'm writing it up as this JMP Foreword article that will hopefully be out soon. It will focus more on text mining of coral reef data, whereas today I'm going to be talking more about the actual data from the coral reefs themselves, not necessarily the manuscripts.

I just want to talk a little bit more about the nature of the interventions before I get into JMP. Basically, if you're not a marine biologist, that's fine. You really only have two options to save a coral reef. You either need to improve their environment, so cool the seawater, reduce the light, reduce the harmful bacteria, things like this, could even move the corals. These tend to be harder. What most people do is they say, "Look, a coral as an adult can't move. If you can't clean up its water, improve its conditions, you need to make that coral stronger."

Most people actually focus on the latter because it's something an individual feels that they can do. "I'm one person. I'm not cool down Florida's coral reefs. But there's a couple of corals out here near my house. I might be able to do something to give them a chance." That's totally fine. But what do the data tell us will be the best thing to do?

I have binned all these interventions into five categories. You can do nothing. You can just wait and see and monitor. That's what people have to do if they don't have many resources, unfortunately. You can conserve. That's where you're not actually intervening per se. You're more preventing other people from intervening. That tends to be more of a best-case scenario. That's for places that maybe haven't gone too far down the dark path yet. They're not super degraded.

Mitigation is what I was getting at in the last slide. These are the things where you're actually trying to improve seawater quality. These tend to be hard and expensive, but they don't have to be. Manipulation, this was actually my background. This is where you're actually trying to do something to make the coral stronger, give them better probiotics to help the microbes living in their tissues, giving them little pulses of heat stress to make them stronger.

Restoration is where you're actually trying to go out there and replant corals in places where they've degraded. This is the hot topic right now. This was a little bit of the motivation, as I mentioned with the Coral Restoration Consortium, is restoration we know is important. It's going to be something we're going to be doing for decades. But is it always the intervention that you should do first? That's my question.

Biopreservation: we're not going to talk about too much, but this is basically the last ditch effort. If you failed at everything else, and you need to make sure those corals don't disappear off the face of the earth, then put them under the ice, put them in an aquarium, and hope for the best.

What I did is not… I've been working in this field for 25 years. I have a lot of my own data, but I'm obviously not the only person that ever tried to save a coral reef. With the help of my trusty AI, Claude from Anthropic, as well as literature I've amassed over the years, I basically whittled it down to 5,000 studies, typically peer-reviewed, but not necessarily, of anybody who's done anything to try to help a coral reef.

I then spent considerable time trying to figure out success. Did they meet their goals? Which might be different from did they save a coral, yada, yada, yada. Not going to talk about that too much. Basically, I tried to whittle it down to a binary. Did what they try work? Did it not work? We'll see that in a minute in the table.

Then again, there's a lot of different things that might influence the success of an intervention, and I'll talk a little bit about some of them. What's living on the reef? The seawater quality. How much money do you have? How much stress is on the reef? So many different things might determine the outcome. These are going to be the predictors in our models. Basically, as I was doing this, I threw up all the data up on my website. Most of it will be in the JMP tables that I'm going to provide, but if you want more insight and transparency about everything that I'm using today, it's up here as well.

Now let's finally get in to the actual JMP data table. I'm not going to belabor it too much. Most of the nomenclature is fairly straightforward. We've got the articles. This is one of the most important things to keep your eye on, the action class. This is actually the intervention. This is what the team tried. This is going to be the model Y, the verdict. This is actually something interesting I've never done until this talk is, success is not always cut and dry. There's granularity. What I did was I looked at various metrics of success within a study and tried to see if they were unanimous, in which case you'd have 100% success or not. We're going to actually use these later as model weights.

Just to give you a flavor of some of the other predictors, I started with about—let's see—200 predictors that I thought would be important for dictating the success of an intervention. I whittled it down to 41. Just to give you an idea of some of these, the year. Some of these interventions are new. We're getting better over time. We need to track this progress. Was the intervention done alongside other interventions? You're not limited to one in a study. Where was it done? The country, the income of the country where it was done, things like the study design, was it replicated. Some of the things are more on the statistical side that might actually influence the success. The stressors, was it thermal stress, things like this.

I don't know. We can just… I'm not going to go into this in too much more detail, but just to give you a sense of some of the ones that might make more sense. But you can see it's already going to get really messy. Oh, this paper didn't report where the study was done. Heavy bias towards the United States. Note this is going to be a thorn in my side because the US's coral reefs are actually in terrible shape, and a lot of these interventions are being worked on here. They may not necessarily make sense in other places.

I actually think this is an interesting one. We'll look at it in a second. The number of people on the project. You might guess that bigger projects probably, they'll probably have more resources and they're more likely to succeed. These were some of the predictors I looked at. Was it actually tested in the field or was it a pie in the sky idea or was it a lab idea? Yada, yada, yada.

Let me just look at my notes, if there's anything I wanted to… Oh, I did want to mention this. I'm a coral reef researcher. Obviously, corals aren't the only thing living on a coral reef. There's fish, there's other inverts. This list of 5,000-ish papers is very coral and fish-centric, so it's not literally everything anyone's tried to help a coral reef. It's things people have done to try to help corals and fish. I did want to make that aside.

Claude, I can download these manuscripts myself. That was not something I needed AI to do. But what I had Claude do was go and mine out a lot of these predictors from the articles. I did not want to sit there and get all 200 potential predictors out of 5,000 articles. That's a good job for AI. Did I spend weeks spot-checking things and making sure there weren't hallucinations, errors? Yes. Are there still errors? Almost certainly. But I can't imagine having done that manually. Obviously, we've skipped that step.

We imported the data into JMP. It did do some cleanup. I think this talk, maybe I can do one quick example. I'm a big fan of using the recode feature. One thing that I actually think is really useful, especially for these messy environmental studies, is a trick I really just hadn't used much until today. Let's find a good example of this. This is getting to the idea of multiple responses.

Right now, country… I think it's going to be one of my most important predictors. If I actually look at the distribution of countries, let's look here in the table. Well, it's binning these multi-country projects together, which might be okay. There's an option. I could actually have this be multiple response. Then it's actually factoring this comma in and saying, "Japan, Taiwan, US," they're clearly different countries, so the distribution changes.

Now we've got more like what we want to see. The issue is most modeling platforms would not accept a multiple response column. But obviously, I don't want to throw out country. It's going to be so important for this study. One little tip is you can actually do this make indicator columns after you've done the multiple selection. It's going to make your data table way messier. But now it basically has a binary of was this country featured or was it not that I could then use if I really wanted to pick out particular countries that might be doing the best interventions or doing the worst interventions. That's a little not so much data cleanup, but getting the most out of your data and something that I had not used much until recently.

Filtering is something I do often in JMP, and there's a lot of different ways to do it. Let's look at an example of something I could filter. Let's see. We've got a lot of missing data here. Maybe these studies, they didn't record a lot of stuff. I might feel like, "Well, they're just going to add noise to the model. They're all from the same place, the same author team." Obviously, I could do a hide and exclude this way. That's not actually the filtering I was talking about.

I actually wanted to see if you got instances where there's a lot of unknowns, and this either could be for a true unknown or it could be missing data. These are things you could filter out in your models, and that's something I find really useful. You can also do a subset. I actually thought about doing this for this talk. Does it make sense to do a big meta-analysis of coral reef interventions where you've mixed in fish and coral? Why don't you want to just look at the fish interventions and just look at the coral interventions? That's probably what I'll do when I go to write the paper. But in this demo, I want to cast a wide net, so I'm actually not going to filter out the fish-specific studies in this talk.

Another thing that I think is a really useful tool when you've got dozens to hundreds of different predictors is before you even do any of the actual fun analysis, is looking at this explore missing values. I know I'm going to have tons of missing data. Let's look at all these. Well, I don't want to look at all these. I don't want to look at these new country code.

Let's look at some of the predictors I already know. These are ones I filtered out because they have too few data. I obviously don't want to do this manually. I can sort so my table has 5,700 rows. Do I really want to keep in some predictors that are only found in… What is that? Five-hundred? A lot of the modeling types I want to use are not very robust with missing data. It's not going to be a hard and fast rule, but it looks like what I did in this case is if you were missing from about 1,000 out of… 15% of the rows, if you were missing, I basically said, "No, I don't want these in my model," even though some of them are actually going to be potentially really important, the type of reef.

Why didn't people note the type of reef that they were working on their study? That's annoying to me, but I can go back and figure this out later. Explore missing values, for me, with these big, messy ecological datasets, is definitely going to be something that's going to be your friend. Let me just do a quick… Just to clean this up a little bit before we get into the more interesting parts. I'm going to group these country columns together. I'm actually going to… We can leave them in for now.

I've done a little bit of cleanup. I didn't actually do any filtering because I decided to just leave everything in. Let's get back to the question at hand of the subset of predictors that I actually want to include. I'm going to take these binary countries out for redundancy for now. I've got these 41. Let's see, before we even build a model, let's go to one of my favorite tools and see what these 41 predictors are or are not telling me about my process.

It's giving me this error because I think I left some multiple response. The Multiple Response Country is still in here, so it didn't let me keep it. That's okay. We'll put country in later. Remember, my Y is the verdict. Was it a success or not? This will take a couple of seconds to run. Basically, what's going to happen is it's going to do a bunch of bootstrap forest models and then see how many times a particular predictor was in the model. I think it ran 100. In 100 of these random forest models, the meow_ecoregion, which is an acronym or abbreviation for Marine Ecosystems of the World, so basically the reef region, was found in 20%. This is already interesting.

We know that there's different parts of the world that are doing these coral reef interventions, and their success might vary. We can just quickly, I think I've still got it selected here, go over to Graph Builder. Actually, I can just type it in. I'm just going to say, meow_ecoregion. Then, rather than have the verdict, which is going to be binary, I think it's going to be too hard to interpret, I'll use the continuous metric of success that I calculated, which is this percent favorable.

You know what I could do is… This is still going to be very difficult to see. Even this, I think, is not going to help too much, but we're getting a little bit better. But one thing I really like to do when you've got a lot of different categories, countries, ecoregions, what have you, this is going to be too difficult to read. Yes, I can say Lord Howe Island, they're batting 100%. That's in Australia. I know the team. They're really good. That doesn't surprise me. Anyway, this is a messy figure, so I don't want to get into an interpretation.

What I can do is I can actually go to the local data filter, find the meow_ecoregion again. What you can do, this is something else I learned that's new, is I can order by count. A lot of them are unknowns. I don't care about those for now. Let me say I just want to focus on regions that have been in at least 100 studies. Now we're getting into something that's much easier on the eyes, and the data are probably going to be much more robust. Honestly, you don't see a huge discrepancy here.

Northern and Central Red Sea. This is where I was working prior to moving back to Miami recently, Saudi Arabia, Egypt. There's a lot of money for coral reefs there now, which is awesome. Their percent success is actually the highest of the places that have done more studies. Fifty-five percent success for intervention, it's actually pretty good.

Where I live now in the Northern Caribbean, the Caribbean's reefs are in the worst shape in the world. Even if you're the best scientist, your odds of success are lower. This is not speaking to the caliber of the work. This is the nature of the environment. This is interesting. This is, honestly, the first time I'm seeing this. I like to do these talks organically. I'd be really excited to share this finding in a paper or LinkedIn or my preferred means. That's just to me, an example of, "Hey, I went from this somewhat dense analysis here into something that would be, I think, even a layperson hopefully would get something out of."

This is interesting. My whole point is optimizing the intervention. Is there differing success rates across the interventions? That's a no-brainer for something we'd want to explore here. Let's look at the interventions and then just look at this probability of success again. You know what? Messed up there. I need to go back. I think I dragged in two things by accident. Let's see, we could set this up like this, or maybe we can do a wrap. Still going to be a little bit messy. No. I see what I've done. Let me just drag it over here from the get-go and then pull this in. No, I think it actually was okay having them all on the same scale.

Let's go back to the original, put the intervention in. Lots of different things you could do, even though I'm going to bin them into six. I'm going to do that same little trick. I'm going to add my data filter so I can reduce the complexity, sort by count. Let's look at the interventions. Let me look at the ones that have been done at least 100 times. Let's sort by… Let's do this. This still may be a little bit hard to read. Let's look at top 10-ish.

We're seeing a similar spread as we saw across the regions. We're seeing in the best-case scenario, which is where you're making an artificial reef. Your probability of success is 55-ish, so again, that's pretty good. Maybe the lowest-probability thing you could do in terms of this subset would be fisheries policy planning, low 30s.

This is probably because most coral reef regions are in developing countries, they don't have a lot of resources. It's hard for them to police things like against poachers and things like that. Again, I don't want to interpret too much. That was more just to demo how easy it is to go from some of these more esoteric data-driven graphics into something that might actually be more visually compelling and easier to understand.

Let's just do one more. Let's see if my hypothesis is correct that we're getting better over time in our interventions. Let's see what I've done. I dragged in country by accident, so I can just remove. Maybe we put a… It may not be too useful. This is going to be a little bit misleading because we didn't have too many. This is being driven by maybe a couple of these old interventions, which will be interesting, not to write them off, but let's look.

This is a new science. Let me look more at what's happened since about the time I was born, so early '80s. Just to make the graph honest, we'll again use the local data filter, year. When you've got a continuous variable, you can actually directly specify where you want it to begin. It's going to actually change the slope if need be.

Honestly, again this is the first time I'm seeing this result. This is an interesting one. On the one hand, our technological capacity to help coral has certainly improved since the early '80s. We only went from 47% to not even 49%. The reason, probably, is because as our ability to save coral reefs has increased, their stress loads and existential crisis that they're facing has also increased. We're getting better at what we're doing, but at the same time, coral reefs are getting, of course, in worse shape.

That's a couple of the more of the simpler things I wanted to demo. I do want to revisit this one more time. Because historically, if I was to build a really complex model using, for instance, genes or proteins where I'm running in the thousands, even then, I typically wouldn't prune at this stage. I'd say, "You know what? Let me just leave everything in from the get-go." Now, I know how much missing data I have in this table. I'm more inclined to say, "Look, do I really want to include these terms in that are only found in 1 out of 100 bootstrap forest models or less? Should I just focus on these?"

That's what we're going to do. I'm going to go from this subset of 41 predictors, I'm going to do another subset, and I'm going to call these just the best predictors. Question mark, because some of them, we haven't even talked about the covariation elephant in the room, which I'm not going to address too much. You can see right off the bat that we're going to have things like country. I'm actually going to pull this out because it's multiple response, but we've got things like number of authors and number of countries, things that are likely going to covary.

Most of the modeling types I'm actually going to show you are fairly robust to this, but for many instances, this would be something you'd want to deliberate on is, actually, we can just do it quickly. I can go over to multivariate, pull these in. It's going to weed out the categorical ones, but we can at least see amongst the continuous variables. Oops! We've got some correlation here in the two different, so one of these is probably, I think, the study year. Yes, study year, start and end. When they start, once it started in the '80s, it's probably going to end later. They're always going to progress forward.

The year of the publication. A lot of these things are basically going to just scale with one another. Really, we probably could just pick one that's redundant, but we can see if that haunts us in a second. That's a good… I just looked at hundreds of correlations amongst each other in a couple of seconds, so the multivariate platform really can be your friend there.

Another thing I like to do, so the predictor screen is using this random forest method to help you get a better subset of predictors. The response screen is more familiar to most biologists, I would say, because this is what you're doing, it's the opposite side of the coin. It's, did these predictors, in and of themselves affect my process? We looked at a few of these already, year and country, but I like to do this predictor screen, response screen as a corroboration. They tell you different things in different ways, but you would think that something like the intervention itself, which we found varied a lot in terms of its success rate, is indeed the thing that number 1 had the largest influence on the percent success.

If I really wanted to whittle down this list of 20 or 21 for my model, I might actually subset it again based on the results of this response screen. This is an interesting one. When more than one intervention… These are basically redundant. This was the intervention done by itself or with another intervention? Both of them seem to be important, and that makes sense. Why do one thing to help a reef when you might actually be able to do two at the same time? That's cool.

Nature-based solutions, this is something I'm real big on. This is stuff like cleaning up the water. The type of research that was done. These are all things that are not surprising to me, but it's cool to see the numbers here and say, "Wow, even at the false discovery rate-adjusted p-value, these factors really had an influence on the success of the intervention." Now we have a handle on what might be important in driving the intervention success.

What I like to do, there's so many different models you can build in JMP. How do you know where to begin? I would say my recommendation here is to go to the model screen. This is actually going to simultaneously test a lot of different models. I'm going to do something a little bit different that I haven't done previously is I'm going to go back to my binary Y, success or not a success, but I'm actually going to weight it by the percent favorability. Did multiple methods corroborate the finding?

I'm going to take my 20 best predictors, or what I'm thinking are my best predictors that were based on that predictor screen, of which maybe half were corroborated by the response screen. We've got a mix of categorical variables. I'm going to just do a quick look because I already saw some that we just saw a second ago that were clearly redundant, even though I said I wasn't too worried about it. Let's see.

Number of institutions, number of other countries, they might say different things because you might have multiple institutions within a country. Research approach is honestly going to have a lot of overlap with intervention. Here it is, combined with other interventions versus interventions in a paper. Let's just get rid of this one, basically telling the same thing.

Definitely want to have XGBoost on the table. It's for these messy, complex environmental datasets. It tends to sometimes be the only one that can give you a good solution. I have a big enough data table, I think, to where I could do both validation and test columns to really try to reduce the chance of overfitting. I do apologize. I'm running this live, and this will probably take, I don't know, a minute or two. We can either edit this out or, while it runs, let me just go back to the slide deck and see if there was anything I wanted to talk about that I forgot.

Barring this last live demo going perfectly, I want to re-emphasize that hopefully, almost surely in my in-person talk in Baltimore, I'm going to be able to fill out this table. This is where I mentioned at the beginning, I'm comparing my guess as to what I think the best solution for a reef is with the Coral Restoration Consortium, again, wearing their shirt, who always recommend one of interventions, ChatGPT, which I wouldn't trust at all.

The JMP models that I'm currently building and are currently running right now that have a human component with what's likely going to be a machine learning model. Then what I've also been doing is incorporating my best JMP models into an LLM that's got all the text data from these same articles that have been mining out stuff to see if we can make an even better prediction. The model screen still didn't run, so I'm going to talk a little bit more about this. Basically, the idea is getting back to what I said at the very beginning.

If you're exclusively using AI to help you make your decisions, I worry you're going down a dark analytical path. I think I actually have a slide that I skipped that might… Let's do this. I think this highlights it better. This is the basic idea. It's very dense and complex by design, but the idea is, look, we're using something like Claude, ChatGPT to essentially look at the semantic data.

Now, do I do a lot of text mining in JMP? Yes, but it's typically for other stuff. That's basically looking at the PDF text. Then I've got my machine learning, neural network, XGBoost, other models that I build in JMP. These basically build the analytical foundation that works with what we call these CNNs, these convoluted neural networks or the large language models.

I wouldn't trust having only this top blue layer. I think that would be prone to hallucinations. It wouldn't actually be looking at rich coral reef data. But if I were able to integrate with these models that I'm building in JMP, I think it's two different approaches that, when integrated, are going to give you an even more powerful prediction. Let's go back to here. Basically, I did a comparison. Versus the authors of these 5,000 papers, how often did I agree with them? That doesn't mean anything. We're not at accuracy, or who was right yet. It just means, did I agree with them? Half the time I did, half the time I didn't. CRC, again, they only recommend restoration, and not every person always recommends restoration, so they only agreed 10% of the time.

ChatGPT, which is funny, it's more prone to agreeing with the norm or things like this. So it actually agreed with the original authors the highest percentage of the time. This one is, I don't think we're going to get there today, so I'm going to put a question mark here, but we'll know this by October. I did a more cursory JMP model that I fed into Claude and said, "Hey, how often did the authors in those 5,000 papers pick what you think was the best option?" It was actually only 19% of the time.

That might actually be the most useful finding in this talk. That basically means that 81% of the time, barring this number changing things egregiously, that people could have actually done something better than what they chose. That's the whole reason I did this analysis is because I thought people were just choosing what was trendy or what was in their gut, not actually what was right. Annoyingly, this still didn't finish. So let me try to pause and do something else.

Okay, now we're back. Sorry about that. It's a pretty big model, and I was testing quadratics and lots of interactions, and I think it was just going to keep hanging. So I simplified it a little bit. You may not be wowed by some of these metrics, but let's say, okay, maybe we're okay with this misclassification rate. That's going to be about 80% accuracy. Let's see. We'll put that in here for now. Agreement with authors, I don't want to waste too much of your time doing this now. I'm going to bet it's going to be similar to this. I'm thinking probably people chose the right thing one in five times. But let's just do a little quick look into what this very complicated model is going to tell us.

This is really one of my favorite tools. Validation accuracy. It looks too good to be true, and it is. If you look at this MCC, which you want to be closer to 1 for high predictive power, it's doing a little bit better than random, but not too much. But it will get better. Let's look into a little bit more detail about some of the factors that this XGBoost is relying heavily on. The design rigor: was the study done well? Some of these are not properties of the reef. It was properties of how the scientific team worked together. Upper-middle-income countries probably were doing a little bit better.

This might be a little bit prash-crone, but one of the things I like to do is, this is going to be, this might be sluggish. But I like to use this variable importance because what it's going to do, it's going to say, "Hey, look, we know this is a big, complicated, extreme gradient boosting model. It's hard to interpret. It's almost like a neural network. It's a black box." But actually using this variable importance analysis within the profiler, what it will do is it will rerank the predictors from top, leftmost being the most important, to bottom right being the least. That's going to allow me to then hone in on maybe a few that are explaining my process.

Let's see, I'm probably, in the interest of time, going to stop it early. We already got a snapshot of that. If you saw these importances, that's literally exactly what it sounds like is, what were the most important predictors in this model? The rank order is going to be very similar to what's going to be output here. What I've been doing, which I think… I don't think I have the table set up. Maybe I do.

Basically, what I would do is, once I'm happy with the model, I've got a little script. You can always export the model to the data table or to Formula Depot or whatever. That's always been possible. But instead, I did something a little bit weird, and that's what I was hoping to show you today, but it's going to have to wait till Baltimore, is I basically took the model, put it in a new data table that was similar to the original one, and I basically listed all the different interventions a team could have done, and the model put the probability of success.

Then I could actually see, and that's how I derived this number, is I could actually see what percentage of the time did the team choose the option that gave what the model thinks gave them the highest probability of success. There might be some people that are patting themselves on the back and say, "Hey, look, it said my best odds were to restore this reef with adult corals, and that's exactly what happened." There's going to be other people that are going to say, "Wow, I could have gotten a lot more bang for my buck using another approach." That's the general idea. Once we were happy with the predictions and the model outputs, it would look something like this.

This is a little bit similar. Oh, here we go. This is giving us a similar list as the predictor screen did, which is what you'd want. Number of interventions in a paper. I'm sure more is better. Where it was done, the countries involved, the type of intervention, all these things make sense. What I do, I really like optimizing the desirability, but in the interest of time, what I'll do is instead I will save the model.

I need to go up here. I would save the prediction formula to the data table, and then that's going to allow me to explore it more. I can run simulations within this platform. I can type in values of my study reefs and get an output. If I only had to check a few studies to see if they chose the best intervention, I could just manually type them in here. I could do a simulation experiment where I'm looking at different combinations of all the different parameters. Not just to see who did things well in the past, but to try to see optimizing what would be the best approach for a place maybe you haven't looked at yet.

Hopefully, we'll be able to do a little bit more concrete details once I've really looked at more of these models, once I'm happier with the outputs. I think the moral of the story is, over the course of 30, 45 minutes here, I went from this 6,000-row data table with lots of different potential predictors, and I really honed in on a few that I think are really dictating the success of these interventions. That's going to be something that's going to be really useful to coral reef scientists.

I could integrate the models into something snazzier, like something that's going to look more like a chatbot where you're not having to sit and type in numbers. You can have a conversation, but it's important that the models are fueling this conversation, not simply generative AI. I didn't get to do exactly the comparison that I wanted, where I got to do a really formal comparison against AI, humans, consortiums of scientists, but I got a little bit closer, hopefully.

Didn't factor dollar values in as much as I wanted to because most people didn't actually price out the cost of their interventions. It's very back-of-the-envelope at the moment. That's something that's been a thorn in my side. Really, what I hope in terms of the scientific product in the end is this last statement. As much as I like doing the data analysis, I don't think every impassioned conservationist should have to do it. If they want to, that's great. I want them to be able to leverage the power of these JMP models to make the best decision in a simpler way for them as possible.

I'm not quite there yet. I'm getting there, but I've made a lot of progress. I think if you're familiar with JMP Pro in particular, it's just amazing the kinds of things you can do with a really rich dataset. I'll wrap it up there and look forward to seeing some of you in Baltimore. Thanks.