RapidMiner vs. Human estimation?

eldenoso

Hello altogether,

for an upcoming presentation regarding the topics machine learning/AI I want to give the audience a practical example of how powerful tools like RM are.

For that I want to do a survey/estimation whereas rapidminer should produce the same results as the average of the audience. My first guess was using ensemble models. But since I have never used them until now I dont really know how to achieve that and if this kind of demo is even possible.

Thank you :-)

Philipp

MartinLiebig

Hey,

 

Analyze Sentiment is part of either Aylien or Rosette extension which are freemium extensions. Aylien got 1000Examples/day for free.

 

And LHC: Yep, i've done my PhD in a similar environment so I now a bit of it.

MartinLiebig

Hi Phillip,

 

to speak in more general i think you should have two principal ideas in mind.

 

The more complex it gets the weaker is a human

Humans are used to work in a 1,2,3 or 4D world. This is how we work. If you take a high-dimensional data set where the pattern to extract is complex, the humans will loose. Simply because we can not grasp a 10 level deep tree like structure

 

The more transfer knowledge is usable, the better the human

Humans excel in text and image mining. A reason for this is that a human can use his fast knowledge on visuals or the language he is using and put this into his decision function. We are very used to extract features from faces to remember someone. On the other hand you can see that this might dramatically fail if you are european and the face to recognize is asian. In this case your "built-in feature generation" breaks. You need to adapt.

 

Best,

Martin

eldenoso

Thank you for your explanation, Martin! :-)

I understand what you were saying. But if it is like you said, there must be a point where the machine is equally good as the human and I think that would be the most interesting point to show to an audience. So is there a way to model a task (for example classification, estimation) that can be solved by an audience and e.g. an ensemble that produces nearly the same result? I think that would somehow "shock" the audience, that a machine learning process can reproduce nearly the same result. 

Thank you

MartinLiebig

Dear Philipp,

 

did you think about sentiment analysis? The prebuilt Aylien/Rosette operators might be a good starting point. You can pick ~10 people from the audience and give them red and green papers. Then they should vote intependendly on the sentiment and you can compare this with the result of a Sentiment model.

 

Another story to tell would be Scanner girls in particle physics. Thats a thing where we moved from recognition by eye to multivariate approaches in the last 50 years.

 

Best,

Martin

eldenoso

Sentiment Analyzis sounds really good to show. I have read some websites few minutes ago, but in former version of rapidminer the operator "Analyze Sentiment" was available, in the version 7.4 it is not. Where do I find the operators needed for this?

Yes, thats a good example! :-) The LHC is doing that constantly, since there are millions of particle collisions every second, which data has to be analyzed.

Best regards

Philipp

MartinLiebig

Hey,

 

Analyze Sentiment is part of either Aylien or Rosette extension which are freemium extensions. Aylien got 1000Examples/day for free.

 

And LHC: Yep, i've done my PhD in a similar environment so I now a bit of it.

eldenoso

Thank you, Martin! 

I played a little bit with the sentiment Aylien analysis and it is just amazing. I already tried myself on analyzing twitter data of different politicans. Interesing who is playing with fear and who is the hopeful one .

Philipp

Thomas_Ott

HA!