"Time series"

SharTeel

Hello,
I have just started with Rapid miner and have a very specific dataset. Three columns, representing distance between pairs of points over time. It is about locations of facial expression points, so the three signals may or may not depend on each other. The time is the same for all three, but can differ between samples - sample1 can represent an expression with duration 60 (so 60 rows), sample2 can have 400.

Question: is there a function in RM that deals with this type of time series? I could not find any.

Also, my classes are special. They are not all distinct. They correspond to different emotions, but some are closer than others - I have 'anger', but also 'pain seen' and 'pain felt'. There is a certain level of hierarchy, but not quite - for example, it is unclear whether to file 'considerate agree' under category 'agree' or 'thinking'.

Question: are there tools that deal with such ambiguous hierarchies? A normal hierarchy induction will not work for me.

Currently, I have decomposed each of the obscure labels into a fingerprint on 4 different dimensions. The dimensions are of different importance, for example, D1 is way more necessarry for distinction than D3. 

Question: can you set priorities so with the tool?

I thank you in advance,
ST

land

Hi,
did you take a look at the time series extension? There are all operators that deal with this kind of data.

RapidMiner itself is restricted to deal with one label at a time. Of course you can switch through all 4 labels and deal with them separately. It's then completely under your control how you are going to set the priority. For example if you have numerical dimensions (don't know from your description) you can scale them differently and use for example absolute error as an performance measure. If you average it over all 4 labels, it reflects the importance.

And yes, I know that is quite impossible to do if you are just a beginner. So let's draw the bottom line: It's possible but it's a quite advanced task. You will need some time to learn how this is possible...

Greetings,
Sebastian

SharTeel

Hallo Sebastian,
thank you for your answer. Yes, I looked at the time series, but the applications I saw did not reflect the case that I have three signals over a common time line. So each row of mine is a frame, and the columns are three...

I first want to try with the raw signals. At first I performed discretization on my data, converting it to linear representation of distinct integers [total alphabet size 27], but this introduces even more noise in my sparse, poor quality data. Alternatively, I will, unless I find a way to analyse dependencies between the three distinct signals. Otherwise I will just concatenate their discrete states, ending up with a single string.

So is there something specific for n signals over a common timeline?

Many thanks,
SchT

land

Hi,
actually I can't see a difference between your setting and a multivariate time series? And this can be handled using Value Series Extension very well.

Greetings,
Sebastian

SharTeel

Hi Sebastian,
the time series extension is really great, but I have trouble with the startup. Namely:
How do I express a time sequence if I have other attributes. For example, if 'label' is 'anger', there is an attribute for 'pleasantness' with possible values -2 to neutral 0 to 2 (how do I indicate that those are in gradation?), but there is also the actual content of frames and value for each frame. I have a hard time bringing this in from CSV. Also, to a noob, what are those attribute names 'batch and polynomial' Online search did not get me good answers for batch.
Thank you in advance,
ST

land

Hi,

well, explaining all this would somehow exceed the frame of this forum. But we have a Webinar on that topic tomorrow the first of april. There all this is shown.

Greetings,
  Sebastian