1
00:00:00,660 --> 00:00:03,720
Now these are just a few examples
of the types of things that can be

2
00:00:03,720 --> 00:00:05,800
analyzed using time series.

3
00:00:05,800 --> 00:00:10,750
And just about anything that has a time
factor in it can be analyzed in this way.

4
00:00:10,750 --> 00:00:14,490
So what types of things can we do with
machine learning over time series?

5
00:00:15,730 --> 00:00:20,410
The first and most obvious is prediction
of forecasting based on the data.

6
00:00:20,410 --> 00:00:22,940
So for example with the birth and
death rate chart for

7
00:00:22,940 --> 00:00:24,870
Japan that we showed earlier.

8
00:00:24,870 --> 00:00:29,180
It would be very useful to predict future
values so that government agencies can

9
00:00:29,180 --> 00:00:33,880
plan for retirement, immigration and
other societal impacts of these trends.

10
00:00:35,070 --> 00:00:39,430
In some cases, you might also want
to project back into the past

11
00:00:39,430 --> 00:00:41,260
to see how we got to where we are now.

12
00:00:42,440 --> 00:00:45,100
This process is called imputation.

13
00:00:45,100 --> 00:00:48,810
Now maybe you want to get an idea for what
the data would have been like had you been

14
00:00:48,810 --> 00:00:51,540
able to collect it before
the data you already have.

15
00:00:52,760 --> 00:00:56,360
Or you might simply want to
fill in holes in your data for

16
00:00:56,360 --> 00:00:58,700
what data doesn't already exist.

17
00:00:58,700 --> 00:01:01,490
For example,
in our Moore's law chart from earlier.

18
00:01:01,490 --> 00:01:04,980
There was no data for some years because
there were no chips released in those

19
00:01:04,980 --> 00:01:06,710
years, and you can see the gaps here.

20
00:01:08,290 --> 00:01:10,210
But with imputation, we can fill them up.

21
00:01:11,780 --> 00:01:16,130
Additionally, time series prediction
can be used to detect anomalies.

22
00:01:16,130 --> 00:01:19,970
For example, in website logs so
that you could see potential denial of

23
00:01:19,970 --> 00:01:24,110
service attacks showing up as a spike
on the time series like this.

24
00:01:25,210 --> 00:01:29,280
The other option is to analyze the time
series to spot patterns in them

25
00:01:29,280 --> 00:01:32,540
that determine what
generated the series itself.

26
00:01:32,540 --> 00:01:37,210
A classic example of this is to analyze
sound waves to spot words in them

27
00:01:37,210 --> 00:01:40,810
which can be used as a neural network for
speech recognition.

28
00:01:40,810 --> 00:01:45,620
Here for example, you can see how
a sound wave is split into words.

29
00:01:45,620 --> 00:01:49,900
Using machine learning, it becomes
possible to train a neural network

30
00:01:49,900 --> 00:01:54,160
based on the time series to
recognize words or sub words.