It looks like you're new here. Sign in or register to get started.
Can you please suggest the link to download the Prescriptive analytics operator.
hello @bhushan_kadu - welcome to the community. I'm not really sure what you're asking for here. To my knowledge there is no operator called "Prescriptive Analytics", and in any case the RapidMiner operators work internally within the software. You would never download an operator separately like that. Perhaps this helps?
Thank you for your response. I was looking for the below operators.
yes indeed - looks like an old version!
This might've been an extension that was presented at RCOMM/RapidMiner World. I don't remember seeing this on the marketplace but it might be around somewhere on the web.
I have found in version 8.1 the Prescriptive Analytics operator.
The operator uses an evolutionary optimization method. Do you know which method? It seems a black box.
Could be part of a price optimization process?
It uses a simple evolution strategy approach for mixed numerical and nominal inputs like the one described here: https://en.wikipedia.org/wiki/Evolution_strategy
If you want to see more details about how we implemented cross-over and and mutations you can check out the code on Github, for example of the evolutionary parameter optimization and the evolutionary feature selection. We use the same techniques for the prescriptive analytics operator.
Hope that helps,
Thank you, @IngoRM
If I have not misunderstood, the prescriptive operator calculates an evolutionary strategy based on:
. the model obtained with the existing relationships between inputs and predictor
. and the new data that we put in the prescriptive's training port.
With this, the prescriptive operator calculates an optimal scenario. But the optimal scenario is not existing into the new data, it's a new case. Then, the data (training port) is the starting point to begin the evolutionary strategy?
Is that correct?
Well, not exactly. The training data is actually not really used during the optimization run. The only part of the training data which is used are the statistics, such as what are the possible nominal values and what are the minimum and maximum values for the numerical attributes. This information is used to constrain the optimization to avoid constructing optimal solutions which are outside of the boundaries of what is possible / has been observed in the past. You can further restrict the optimization in the parameters by the way. Other than that, the training data is not used.
What happens is that the optimization creates candidates for optimal attribute values ("individuals" in evolutionary lingo) and applies the models on those candidates and creates the prediction. It then measures how close the prediction is to the desired goal. This is the fitness function in the evolutionary approach. The individuals are changed until the optimization finishes. The delivered example set are the most optimal attribute values, i.e. a model input which delivers the desired prediction (typically either a fixed value in regression, or a specific confidence for a class).
Hope this helps,
I am new to data science and have found RapidMiner to be a good choice to learn about data science. I am currently exploring the Prescriptive Analytics operator through the sample tutorial.
I imagine myself analysing a dataset and comparing different modelling techniques/operators. Therefore, what I did was I ran several modelling operators for the prescriptive analytics and also compare their ROC curves.
The ROC comparison (rapid-pic1.gif) shows Naive Bayes having the lowest ROC and Random Forest having the highest. When prescriptive analytics is run, it shows different confidence levels (rapid-pic2.gif and rapid-pic3.gif). We know from widely-accepted facts that the result produced by Naive Bayes is the most logical.
ROCPrescriptive analytics - NaivePrescriptive analytics - Random Forest
My question is if in the case that I am conducting a real analytics task, should I trust the ROC comparison more than the Prescriptive Analysis operator?
the ROC Curve gives you a overview over a whole set example. Every algorithm will be wrong here and then. That's part of the nature. The picture of the prescr. analytics operator on the other hand show only one example.
My question for you would be; Do you rather want to take the one which is in average better or the one which is in average worse, but in your specific case good?
Hi, I guess you may want to consider what certainty threshold you are looking for and then assess the ROC curve according to your threshold. Check out this video (later part min 6 and later) and the this paper for more details.