And now we are to the last section of today's lecture

today's first lecture of the seminar.

And we want to very briefly touch base on common machine learning tasks for time series.

So machine learning can be regarded as a collection of methods that enables us to solve tasks

which would be too difficult to solve with a fixed classical way of writing programs

that are fully designed by human beings.

So from a philosophical point of view

it's interesting because machine learning can be

seen as the attempt of formalizing the concept of human intelligence.

Machine learning usually describes how machines should process examples

so data collected for a given problem.

So examples are described formally as a collection of features.

In the case of time series

features are observations and also have a temporal dimension.

So they are sorted in time.

And we said that in the beginning

the dependency

the temporal dependencies between those samples

are of interest for us.

We have a number of problems that are specific of time series.

We can talk about time series classification.

In this case

we have a data set that is made of time series S with associated class label C.

And the task is now to learn a function F that is able to learn the mapping from S to C

meaning it's able to process the time series and then associate to hit the correct class.

A concrete example of classification we already saw in the beginning of the lecture,

the monitoring of player actions and the classification of this temporal data into

the 10 classes or the 10 different actions that a professional beach volleyball player can do.

Another task with time series is that of time series forecasting.

In this case

we have a time series S that is made of observation as 1 to S T

capital T 1, as capital T 1 plus 1, as capital T 2.

And in this case

what we want to do is to forecast new values of the time series.

So in particular

in this example

all values from S 1 to S capital T 1

they are the input of our function F and we want the output to match the following observation.

So we want to predict with the function F future values of our time series here depicted in yellow.

And then example of this could be the forecasting monthly sales for a retail store.

Let's read together the scenario.

We have a retail store and the retail store wants to predict future sales to optimize the

inventory management, staffing and promotions.

The data available consists of monthly sales

figures for the last three years.

And we want to train a model that is able to predict sales

for the next six months.

And therefore, this is a classical forecasting problem.

We have a

three years time span of data.