Jan Motl (jan@motl.us)

Need

Whenever we want to mine a database (which contains several tables), we have to first transform all the tables into a single table, that is processable by conventional classifiers (KNN, SVM...).

​

This preprocessing is a manual, tedious, boring and time consuming task.

​

But as it’s tedious and boring, it is also automatable!

Approach

• Automatic “propositionalization” of database with a set of ~30 proved patterns.
• The prescription how to transform several tables into a single table will be delivered in SQL to make the solution easily deployable.
• Generation of the SQL itself will be done in a programming language like C# or Java.

​

Benefits

• Fast, painless & correct data preprocessing
• Fast: what takes 1-4 weeks will take 2 days
• Painless: automatization of boring and tedious work
• Correct: machines are not prone to fatigue
• Accurate: much more exhaustive search space
• Documented: Achilles heel of each ad-hoc work

​

​