This code implements the transformation learning described in Section 3.2 of the paper by Zhu, He, and Chaudhuri, "Auto-Join: Joining Tables by Leveraging Transformations" that appears in the Proceedings of the VLDB Endowment.

I make no claims of completeness of this implementation, it was basically a small enough problem to demonstrate various approaches to program synthesis on a real problem. It is unclear what approach the original authors used since they did not release their source code.

I acknowledge my colleague, Fatemeh Nargesian, who pointed the paper out to me.

This repository contains three approaches to solve the underlying program synthesis problem.

• tl.py, uses a brute-force CSP-style approach.
• tlsolv.py (and the code in solver.py) which implements a "deductive" program synthesis approach
• cvcsolver.py, a prototype that uses a SMT solver (cvc5) to implement a solver, though it is slower than tlsolv.py and does not yet produce all possible solutions.

To run tl.py or tlsolv.py, run:

python3 ./tl.py tests/pres1.csv tests/pres2.csv
...
FUNCTION Concat(SplitSplitSubstr(ROW, 0, ' ', 1, '(', 0, 0, -1, Case.UNCHANGED), Concat(Constant(' '), SplitSubstr(ROW, 0, ',', 0, 0, -1, Case.UNCHANGED)))
...

For the other example, run:

python3 ./tlsolv.py tests/school1.csv tests/school2.csv --sc 1