Original subtitles

Speaker: The first AI feature

that we're gonna dig into

is something called Anomaly Detection

which is used to automatically detect

and explain anomalies in time series data.

So this can be applied to line charts in Power BI.

And the way it works is that this feature will add flags

to existing line charts when its spots anomalies

and it will also produce AI generated explanations

and summaries to help you understand

what may have driven the anomaly.

Now, before we jump into our demo

we want to talk about a few limitations

of anomaly detection.

For one, it's currently only supported

for line charts with time series fields on the X axis.

It doesn't support charts with legends,

multiple values or secondary axes.

It can't be applied with other types

of visual features like forecasts.

It's not compatible with drill up

or drill down functionality.

And you need at least four data points in your time series

in order for this to be available.

So with that, let's jump back

into our AdventureWorks Report,

pull up our exec dashboard line chart

and see if we can spot some anomalies.

All right.

So if you'd like to follow along,

head to the Exec Dashboard tab.

And we're gonna see if we can detect some anomalies

in our revenue trending line chart.

So first things first,

if we drill into our quick format options

you'll see, Find Anomalies here at the bottom of the list

and it's currently deactivated.

And the reason why is because

we have three different fields in our X axis,

year, quarter, and month so that we can enable that drill up

and drill down functionality.

Now, if we want anomaly detection,

we kind of have to make this trade off.

And the simplest way to do that

would be to remove the year and quarter fields.

And now, if we head back to those options,

we can activate anomalies.

And what this is gonna do is create

what's called a sensitivity band

shown in gray surrounding our line chart here.

And if we hover over some of the tool tips,

you'll see three new values appear,

an expected value, an expected minimum,

and an expected maximum.

And that minimum and maximum value

defines the width of that sensitivity band.

Now that's something that we can adjust

and customize in our formatting pain.

If we scroll all the way down to the bottom

to the Find Anomalies Options,

right here we see our sensitivity,

which is defaulted to 70%.

And what would happen if we had a less sensitive model?

We can apply a 50% sensitivity

and we'll see that band expand pretty dramatically.

And what that basically means in plain English

is that it would take a very, very high or low value

in order to exceed the bounds of that band.

Whereas on the other hand,

if we increase the sensitivity to something like 90%,

we'll see that band contract significantly.

So now we're more sensitive to those outliers and anomalies.

It doesn't take much volatility

in order for a data point to get flagged.

Now, we still don't see any anomalies called out here

but keep in mind that this visual

and the anomaly detection feature

responds to filter context just like any normal visual.

So we can click to Accessories or to Bikes and check it out.

Now we see an anomaly,

which is marked by this icon or this flag

that's appeared on December, 2021.

And it's flagged because the revenue for that month,

at 1.511 million, exceeded the expected maximum value

of 1.508 million.

And what's really cool is that you can click that marker

and it's gonna open up a new tab

a new pane here on the right that describes the anomaly

and provide some possible explanations.

So right now it says, we couldn't find any

significant explanations for this anomaly.

And the reason why is because,

if we go back to our format pane,

right here beneath the sensitivity option,

we don't have anything in this Explain By field.

So what we can do here is say,

hey, try to explain the anomaly

by looking at changes in a certain field,

like product subcategory name for instance.

Let's go ahead and click apply.

Now when we head back to that anomaly pane

now we get a possible explanation,

which is subcategory name is Mountain Bikes.

And it says, total revenue for subcategory name,

Mountain Bikes, was unusually high

which may have lifted the total revenue total.

So it's kind of helping us with some root cause analysis

and it's also giving us a measure of strength.

In this case, 13%.

And like this tool tip says,

this gives us the degree to which

each value is correlated with the anomaly.

A higher percentage or higher strength

means that this factor has a greater impact.

And we can add multiple fields here as well.

So let's add another field

for something completely different,

like country for instance,

click apply, head back to anomalies,

and now, it's found an even stronger factor,

which is, country is United States.

That has a strength of 36%.

Now, one feature that I love about anomaly detection

is that these visuals that are produced

to support the explanation

can be added to the report as new visuals.

So if you found a really interesting anomaly

and narrowed down to a significant root cause,

you could add that visual to the report

by clicking this button.

It's gonna drop it right here on the canvas

and you can resize, format, treat it like any other visual.

Really, really cool stuff.

So I'm gonna delete that visual here.

And last thing to call out is that,

we can just format our anomaly options

just like anything else, right.

So, let's select our line chart,

head back into our Anomaly Options.

And this is where we can change things like the color

or the size or shape of the marker itself.

Actually, kind of like this teardrop in gray.

And we can change how that sensitivity band looks as well.

So, maybe we want something like a yellow shade here

with a little bit less transparency.

Totally up to you.

But just note that all this stuff is customizable.

So there you have it.

That's your crash course in anomaly detection.