You Ask, I Answer: Multi-Objective Optimization for IBM Watson Studio AutoAI?

In today's episode, our Juna asks, could you please suggest an approach to forecast multiple targets? Is there a way to select multiple columns? And IBM Watson Studio Auto AI. In our use case, we need to develop time series forecasts for multiple products. If we correctly understood auto AI or allow us to select one column at a time to generate such a forecast, is there an alternative to select multiple columns representing multiple targets?

Thank you. Okay, so there's a lot to unpack here. One, uh auto AI does not do time series forecasting. Auto AI does either regression or classification. So it doesn't support that at all.

There are methods for doing time series forecasting. In uh Watson Studio, you'd want to use the uh SPSS model or for some of that. Um Watson Studio Auto AI out of the box does not support multi-objective optimization. In fact, none of the auto AI slash auto ML family of technologies right now support uh multi-objective optimization. And the reason for that is that is it it auto AI itself is fairly costly because uh the the analogy I like to use is if you're baking cookies, these tools are essentially varying every possible oven temperature, every possible ingredient to see what the best overall cookie is.

Uh that is computationally very costly. Multi-objective optimization is also very costly, uh, and it adds crazy immense uh amounts of dimensionality. The current uh uh technical name for it is Pareto multi-objective optimization. And if you think about two people playing tug of war, right? Uh they're playing tug of war, and the little ribbon in the middle of the rope is the objective, right?

And they're pulling back and forth. Um that's a good example of like a single objective optimization. You would you know somebody's got a win. Now imagine uh tug of war with three people, three people holding on ropes, and there's that's still, you know, there's things in the middle, and each one has uh a thing, and now they're four or five or ten people playing Tog of War, all holding different ropes. You can see how very, very complex this gets.

Um multi-objective optimization gives you many, many different scenarios to to uh to plan for. And then auto AI has many scenarios of each scenario. So you can see how it just stacks up and becomes computationally unfeasible. The way we handle multi-objective optimization most of the time is doing uh what's called constraint-based uh multi-objective optimization, where you say there's guardrails. So in the marketing world, we have in when we're doing SEO, we have keywords, right?

And we have the volume of searches for a keyword, we have the number of likely clicks on that keyword, we have the cost per click if it's paid, and we have the difficulty it uh we have to rank for a certain keyword. Trying to do a four-way or five-way um algorithm to to create the best balance of all those possible outcomes is really difficult because you have to compute every possible edge case. You know, sometimes you want difficulty 100, you'll never rank for this keyword. Well, that that doesn't that's not very sensible, right? Uh sometimes you want uh zero dollar cost.

Well, again, not necessarily all that realistic. So what as data scientists will do is apply constraints first into the data set before we do auto AI on it. We'll say, you know what? I'm not willing to pay more than seven bucks a click. Right.

So that immediately knocks off a certain part of the the table. Um I'm not interested in keywords that are, you know, uh above uh difficulty score 50, because I know my content's not that good, so I'm not going to be able to really rank for stuff above that. So let's chop off that part of the table. Uh I'm not really interested in keywords that have no search volume. We'll chop off that part of the table, and you can see we're starting to apply constraints to our data set first, so that when we stick it into something like auto AI, we already have uh a much more slimmed down data set where a single objective now makes sense, right?

We'll manually look at the table and say, you know what, I want to optimize for clicks. Clicks is what I care about. Traffic to my website, but I'm going to apply constraints manually on those other columns. I don't want a below a certain volume or above a certain cost or too tough to rank for. And then that goes into auto AI, and auto AI it actually makes auto AI much more efficient because uh it has much less data to crawl through.

So you would apply those constraints in advance. Um you can do this with multi objective optimization as well. You'd apply your constraints first, and then uh in Watson Studio, there's the facility to use R or Python notebooks right within the interface, and so you can write your own code to apply using the the auto the multi objective optimization uh library of your choice to do it there. So you could do that. That would not get you the auto AI capability, but it would let you do multi objective optimization.

You can also use um the decision optimization or the the C Plex facilities also within uh Watson Studio to do some of that if you're not comfortable coding. Again, it doesn't get you the auto AI capability, but it does get you the decision making capability. Finally, on the topic of time series forecasting. Um series forecasting is tricky in the sense that um you need to do the constraints first, then you need to do the auto AI first, next to um uh probably regression, uh either regression or or uh classification, mostly regression, to figure out what you want to forecast, what is worth forecasting, and then you do the time series forecasting on that. So that's a three-step process.

There's you go from uh constraint to uh regression to forecast, and that's the process for that. That is not automated either. This actually, this whole question, this discussion is really good because it highlights the immense difficulty uh the data science and AI community is having with a lot of these automated AI solutions. They are good at very narrow tasks, they're good at uh one thing, but the number of techniques that you can combine that your your human data scientist will uh know to combine and in what order is very difficult to put together in a machine just to have a you know a push the button and and and let the machine do its thing. Um it will come in time, but it's gonna be a while.

Uh it's it's not gonna be uh in the next quarter's release, let's let's put it that way. So, to answer your question, do your constraints, do auto AI uh to determine which uh which feature selectors are the most relevant to your outcome, and then do time series forecasting. And and again, you can do that um in the SPSS modeler in uh Watson Studio, or probably you'll use uh um uh a fancier library uh like uh any number of the the Python or R libraries to to really kick it up a notch after that. The good news is within Watson Studio, all of that, even though those are separate pieces, pieces of that can then be pushed to Watson Machine Learning for production use cases. But it is it's this is not an easy project, but it is an interesting one because you're really talking about the heart of making great decisions using machine learning.

Um so good question. If you have follow-up questions, please leave them in the comments below. Uh please subscribe to the YouTube channel and to the newsletter. I'll talk to you soon. Take care.

Want help solving your company's data analytics and digital marketing problems? Visit TrustInsights.ai today and let us know how we can help you.