[S2E1] Prescriptive Analytics | 5 Minutes With Ingo - YouTube

Hey Ingo, do you know what time it is?

I guess it's time for Five Minutes with Ingo?

I mean, it's also time for all the rain, look at that.

But this actually gives me an idea.

I would like to discuss something which is called 'prescriptive analytics.'

The idea here is, if you know for example the weather forecast, the prediction for tomorrow,

you could try to figure out what is the best course of action.

For example, should you bring your umbrella or not?

So, let's go over there and discuss this a little bit yes?

So, the idea behind prescriptive analytics really is you take predictions, and then you

combine those predictions with optimization schemes.

So for example let's take the umbrella example.

If you know it's very likely that it's going to rain tomorrow, should you bring the umbrella?

Hmm, well, you could just say yeah, sure, why not?

But I would say, let's have a look at your calendar first, let's figure out if you actually

need to leave the house.

So let's say you have an appointment in the office at 8:00 in the morning, then yeah sure,

take the umbrella because you leave the house.

But wait, is the office actually close enough to your home?

So maybe it's not, and you need to take the car.

So let's say the distance is ten miles.

You figured this out by looking on a map.

So now you looked into the calendar, and into the map, so you already had two different

data sources.

So, since you need to take the car, the next thing that would be, well everybody's taking

the car because it's raining.

So there will be more traffic.

So you take maybe traffic predictions in as well.

So you have the rain prediction, you have the map information, you have all these pieces

of information, bring them together and now the best course of action is actually, well

take the car, but leave early because everyone will be on the road.

So here you see, actually how an easy forecast, a weather forecast, can turn into pretty complex

decision making, well, problem here really.

And that is an easy case.

So for most business cases, things are much more difficult.

So let's think about how analytics in general can help us with that, let's go over to the

whiteboard and discuss this.

So in general there are four different styles of analytics.

Let's start here at the bottom with BI or business intelligence, which really is just

a look into the past.

For example a BI report could deliver the piece of information that it has rained 237

days in the last year.

And while this might be an important piece of information, it doesn't really tell you

anything about what is going to happen tomorrow.

So, this brings me to the next level, if you do this for one year you can do this for multiple

years, and you can look into the data of multiple years.

Let's say you had seen 237 days of rain last year, 242 the year before, and then 250 the

year before that, and based on all those BI reports you can create another prediction.

You are kind of confident that it is going to rain more than 200 days in the next year.

While this is a prediction, it doesn't tell you, again, something about tomorrow.

So, in my opinion also not that useful.

Brings me to the next level, or layer here which is predictive analytics.

This is where we use all our machine learning and data science techniques.

A machine learning model here can actually tell you what is the likelihood of rain for

tomorrow.

So if you know there's 95% of a chance of rain for tomorrow, you can at least take this

into account and do something, let's say bring your umbrella, that is good.

But doing something is really the keyword here, and that brings me to the top level

here which is also the level which provides the biggest value.

Because knowing your future is one thing, but if you are able to change this future,

that is actually even more valuable.

And the way how you are doing this is you take those predictions, and you have different

options for your actions.

And you go through all those options, you predict how those options will effect your

future, and then of course you pick the course of action which delivers you the best future

and let's say the best outcome for tomorrow.

Not getting wet and getting into the office on time.

So this is why Prescriptive Analytics is so important, because it provides so much value

by combining machine learning models and optimization schemes.

Ingo, can you discuss more about the optimization schemes for our course of action?

Sure!

Well in general you can use any optimization scheme.

So here is a little dataset here on the bottom here.

We have a couple of columns we created our predictive models on, and you had to predict

the happiness of people.

Based on, for example, the type of car they are buying.

And this is now interesting, because you could basically go through all the car brands and

all the options just to figure out how would a certain person, how would this car purchase

effect the happiness of this person.

And if the number of options are small, you can literally go through all the options.

And that's kind of a brute force optimization technique.

But if the search space is very large, you need to use more heuristics to actually get

to the perfect outcome in a more efficient way.

One widely used technique, evolutionary algorithms, so they work really for all kinds of different

problem types and on large search bases.

And that's it for today, thank you!