This project is a research on how to extract rules from the existing data. The dataset used to train the model and extract rules is Boston Housing Dataset. The goal is to extract the rules based on which the taxes are formed.
Rule extraction works as follows:
- Decision Tree Regressor is trained on data provided in data/apartments_sale.csv
- Decision tree is traversed and rules are extracted into data/extracted_rules.json file
- Since some rules are redundant they are simplified by simplify_rules.py script (Eg: greater than 2 and greater than 3 becomes greater than 3).
- Simplified rules are then refactored - for all the pairs of conditions that exist in more than 5% of generated rules refactoring is performed: Each pair becomes a subrule that is then used as a 'variable' in the rules where the pair is found.
NOTE: Make sure you have python 3.7 installed with all the requirements listed in requirements.txt
- Run main.py - trains a model and extracts raw rule json into data/extracted_rules.jsonj
- Run simplify_rules.py - loads data/extracted_rules.json and eliminates redundant rules.
Eg: greater than 2 and greater than 3 becomes greater than 3 - Run refactor_rules.py - simplifies rules by introducing variable-like clauses
- Run human_readable_rules.py - extracts human readable rules and writes them to data/human_readable_rules.csv
After rules and subrules are extracted subrules can be renamed by experts in order to better match the domain.
There are 14 attributes in each case of the dataset. They are:
- CRIM - per capita crime rate by town
- ZN - proportion of residential land zoned for lots over 25,000 sq.ft.
- INDUS - proportion of non-retail business acres per town.
- CHAS - Charles River dummy variable (1 if tract bounds river; 0 otherwise)
- NOX - nitric oxides concentration (parts per 10 million)
- RM - average number of rooms per dwelling
- AGE - proportion of owner-occupied units built prior to 1940
- DIS - weighted distances to five Boston employment centres
- RAD - index of accessibility to radial highways
- TAX - full-value property-tax rate per $10,000
- PTRATIO - pupil-teacher ratio by town
- B - 1000(Bk - 0.63)^2 where Bk is the proportion of blacks by town
- LSTAT - % lower status of the population
- MEDV - Median value of owner-occupied homes in $1000's
| Boston Housing Dataset tax rules | |
|---|---|
| 0 | SUBRULE_0: IF AND indus <= 26.695 AND nox <= 0.583 THEN SUBRULE_0 |
| 1 | SUBRULE_1: IF AND rad <= 16.0 AND nox > 0.583 THEN SUBRULE_1 |
| 2 | SUBRULE_2: IF AND rad IS IN RANGE (3.5, 16.0] AND dis <= 2.8589 THEN SUBRULE_2 |
| 3 | SUBRULE_3: IF AND ptratio > 13.85 AND indus <= 20.735 THEN SUBRULE_3 |
| 4 | SUBRULE_4: IF AND ptratio > 13.85 AND indus IS IN RANGE (20.735, 26.695] THEN SUBRULE_4 |
| 5 | SUBRULE_5: IF AND dis <= 6.2433 AND rad <= 3.5 THEN SUBRULE_5 |
| 6 | SUBRULE_6: IF AND dis > 6.2433 AND rad <= 3.5 THEN SUBRULE_6 |
| 7 | SUBRULE_7: IF AND rad <= 16.0 AND indus > 26.695 THEN SUBRULE_7 |
| 8 | SUBRULE_8: IF AND dis > 2.8589 AND rad IS IN RANGE (3.5, 16.0] THEN SUBRULE_8 |
| 9 | SUBRULE_9: IF AND indus <= 26.695 AND ptratio <= 13.85 THEN SUBRULE_9 |
| 10 | IF SUBRULE_5 AND SUBRULE_0 THEN tax=233.8448275862069 |
| 11 | IF SUBRULE_6 AND SUBRULE_0 THEN tax=328.0416666666667 |
| 12 | IF SUBRULE_2 AND SUBRULE_0 THEN tax=376.8888888888889 |
| 13 | IF SUBRULE_8 AND SUBRULE_0 THEN tax=299.8781725888325 |
| 14 | IF SUBRULE_9 AND SUBRULE_1 THEN tax=264.0 |
| 15 | IF SUBRULE_3 AND SUBRULE_1 THEN tax=400.4736842105263 |
| 16 | IF SUBRULE_4 AND SUBRULE_1 THEN tax=437.0 |
| 17 | IF SUBRULE_7 THEN tax=711.0 |
| 18 | IF rad > 16.0 THEN tax=666.0 " |