|
Predictive modeling covers a set of techniques and algorithms allowing to predict some targeted behavior based on past observations. It is like predicting rain for tomorrow when it rained for the last 2 days... Predictive modeling techniques make some assumptions on the continuity of behaviors, like for wheater prediction. However, we all know that after the rain comes the sun: the problem is to know when! Predictive algorithms identify in data the signs that appear before the rain stops, and will use them to predict when the sun will be back. However, we must hope that our data contains variables describing those signs, like the pressure of the air, which is a very good predictor, as we all know!
This small example just shows that despite the task is easily formulated, a lot of assumptions, and possible traps exists! In order to build reliable predictive models we must have the following:
-
Right data - Having the data allowing to predict the behaviour is obviously a key point;
-
Good data quality - Data that lie damage a lot the performance!
-
Good domain knowledge - Data does not come by itself, you must understand the domain in order to extract what is deemed relevant!
-
Good modeling expertise - There are so many traps... Do you know that the best model of the KDD-98-CUP contest was able to make 14.000$ profit with its recommendations, whilst the worst one's recommendations led to a loss of 50$? (on the same data, just a difference of people!)
-
Good tools - When working on very large data set (such as 18 millions records, with 2.500 variables) the tools starts to be a masterpiece in the puzzle.
Our R&D is very active on this field. We build software components that support the implementation of best practices of modeling methodology. This R&D effort is now packaged in our predictive software RANK.
On client projects we use the technology available of requested by our clients: SAS, SAS/Eminer or any other tool such as SPSS, SPSS/Clementine, SEE5, R (Open Source), MatLab, etc.
|
Technology 
Vadis launches Rank, a turnkey solution for building predictive models (a winner of international KDD contests...)
