It’s controversial because there are only a limited number of phases that we think are sensible (otherwise we end up with 34 steps that no one can remember), and because although it intuitively makes sense to have this phase there isn’t a precise definition of what it actually should be.  To give you an idea of the controversy, in our consortium meetings we were pretty evenly split between people who wanted the phase and people who said “over my dead body”.

 So here’s my take on why we need it…

 A major part of data-mining involves our interpretation of data before we subject it to our data mining algorithms.  We’ve already understood the raw data (Data Understanding) and massaged it into an acceptable shape (Data Preparation), but now we start to REALLY look at it.  Of course we use statistical measures, we use graphical tools, we use our heads!  Sometimes the project stops there – we discover the fault in the data that prevents us doing anything further, or sometimes we spot the trend or linkage that means we don’t even need to build the predictive models.  More often we will go back to the raw data looking for some support, or we will take the plunge and hit it with our model of choice.  But whatever we do we can’t ignore this intermediate exploration step…

Firstly I’ve been playing in Manufacturing.  Despite living in a city that has been home to Jaguar, Rover and is just up the road from Aston Martin, I’ve never really been involved in that type of industry – until now.

 Which leads to the second (and perhaps more relevant) part of this post.  I’ve been looking into Root Cause Analysis.  That’s the art of trying to find out why your auto has gone wrong.  And it seems that there is a lot of scope for using predictive techniques within RCA, but with an interesting twist.  I’m not using the algorithms to predict if your auto is going to fail, I’m using it simply to understand the important relationships that are associated with the failure.

So the question that I want to pose is this – is the CRISP-DM process model as relevant for this type of activity?  Could modifications make it more relvant? 

At least we don’t as far as CRISP-DM is concerned.  It stops at Deployment (and lets be honest, that’s an area that is reserved for the IT Department anyway).

 Well its time we recalimed Montoring for data mining.  We, of all people, should be aware of the value of measurement and – dare I say it – statistical evaluation of our own output.  And we should be honest and admit that not everything we predict either occurs, or if it’s acted on has the effects that we intended.

So lets get out there and evaluate our results.  Lets stop pretending that our job ends at the edge of our cube, and lets add Monitoring to the Process Model.