Rule Generator (Decision Tree algorithm) Example¶
This notebook contains an example of how the Rule Generator (Decision Tree algorithm) can be used to create rules based on a labelled dataset. This algorithm generate rules by extracting the highest performing branches from a tree ensemble model.
Requirements¶
To run, you’ll need the following:
A labelled, processed dataset (nulls imputed, categorical features encoded).
Import packages¶
[6]:
from iguanas.rule_generation import RuleGeneratorDT
from iguanas.metrics.classification import FScore
import pandas as pd
import numpy as np
from sklearn.ensemble import RandomForestClassifier
Read in data¶
Let’s read in some labelled, processed dummy data.
[7]:
X_train = pd.read_csv(
'dummy_data/X_train.csv',
index_col='eid'
)
y_train = pd.read_csv(
'dummy_data/y_train.csv',
index_col='eid'
).squeeze()
X_test = pd.read_csv(
'dummy_data/X_test.csv',
index_col='eid'
)
y_test = pd.read_csv(
'dummy_data/y_test.csv',
index_col='eid'
).squeeze()
Generate rules¶
Set up class parameters¶
Now we can set our class parameters for the Rule Generator. Here, we’re using the F1 score as the main rule performance metric (you can choose a different function from the metrics.classification module or create your own).
Note that if you’re using the FScore, Precision or Recall score as the optimisation function, use the FScore, Precision or Recall classes in the metrics.classification module rather than the same functions from Sklearn’s metrics module, as the former are ~100 times faster on larger datasets.
Please see the class docstring for more information on each parameter.
[8]:
fs = FScore(beta=1)
[9]:
params = {
'n_total_conditions': 4,
'opt_func': fs.fit,
'tree_ensemble': RandomForestClassifier(n_estimators=100, random_state=0),
'precision_threshold': 0.5,
'num_cores': 1,
'target_feat_corr_types': 'Infer',
'verbose': 1
}
Instantiate class and run fit method¶
Once the parameters have been set, we can run the .fit() method to generate rules.
[10]:
rg = RuleGeneratorDT(**params)
[11]:
X_rules = rg.fit(
X=X_train,
y=y_train,
sample_weight=None
)
--- Calculating correlation of features with respect to the target ---
--- Returning column datatypes ---
--- Training tree ensemble ---
[Parallel(n_jobs=1)]: Using backend SequentialBackend with 1 concurrent workers.
[Parallel(n_jobs=1)]: Done 100 out of 100 | elapsed: 0.2s finished
--- Extracting rules from tree ensemble ---
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 100/100 [00:00<00:00, 1865.53it/s]
Outputs¶
The .fit() method returns a dataframe giving the binary columns of the generated rules as applied to the training dataset.
Useful attributes created by running the .fit() method are:
rule_strings: The generated rules, defined using the standard Iguanas string format (values) and their names (keys).
rule_descriptions: A dataframe showing the logic of the generated rules and their performance metrics as applied to the training dataset.
[12]:
X_rules.head()
[12]:
Rule | RGDT_Rule_20211123_100 | RGDT_Rule_20211123_104 | RGDT_Rule_20211123_130 | RGDT_Rule_20211123_158 | RGDT_Rule_20211123_157 | RGDT_Rule_20211123_156 | RGDT_Rule_20211123_94 | RGDT_Rule_20211123_95 | RGDT_Rule_20211123_173 | RGDT_Rule_20211123_124 | ... | RGDT_Rule_20211123_112 | RGDT_Rule_20211123_13 | RGDT_Rule_20211123_18 | RGDT_Rule_20211123_76 | RGDT_Rule_20211123_208 | RGDT_Rule_20211123_111 | RGDT_Rule_20211123_68 | RGDT_Rule_20211123_106 | RGDT_Rule_20211123_107 | RGDT_Rule_20211123_39 |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
eid | |||||||||||||||||||||
867-8837095-9305559 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ... | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
974-5306287-3527394 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ... | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
584-0112844-9158928 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ... | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
956-4190732-7014837 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ... | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
349-7005645-8862067 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ... | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
5 rows × 218 columns
[13]:
rg.rule_descriptions.head()
[13]:
Precision | Recall | PercDataFlagged | OptMetric | Logic | nConditions | |
---|---|---|---|---|---|---|
Rule | ||||||
RGDT_Rule_20211123_100 | 0.991837 | 1.0 | 0.027547 | 0.995902 | (X['account_number_num_fraud_transactions_per_... | 2 |
RGDT_Rule_20211123_104 | 0.991837 | 1.0 | 0.027547 | 0.995902 | (X['account_number_num_fraud_transactions_per_... | 2 |
RGDT_Rule_20211123_130 | 0.991837 | 1.0 | 0.027547 | 0.995902 | (X['account_number_num_fraud_transactions_per_... | 2 |
RGDT_Rule_20211123_158 | 0.991837 | 1.0 | 0.027547 | 0.995902 | (X['account_number_num_fraud_transactions_per_... | 2 |
RGDT_Rule_20211123_157 | 0.991837 | 1.0 | 0.027547 | 0.995902 | (X['account_number_num_fraud_transactions_per_... | 2 |
Apply rules to a separate dataset¶
Use the .transform() method to apply the generated rules to a separate dataset.
[14]:
X_rules_test = rg.transform(
X=X_test,
y=y_test,
sample_weight=None
)
Outputs¶
The .transform() method returns a dataframe giving the binary columns of the rules as applied to the given dataset.
A useful attribute created by running the .transform() method is:
rule_descriptions: A dataframe showing the logic of the generated rules and their performance metrics as applied to the given dataset.
[15]:
rg.rule_descriptions.head()
[15]:
Precision | Recall | PercDataFlagged | OptMetric | Logic | nConditions | |
---|---|---|---|---|---|---|
Rule | ||||||
RGDT_Rule_20211123_100 | 0.991453 | 1.0 | 0.0267 | 0.995708 | (X['account_number_num_fraud_transactions_per_... | 2 |
RGDT_Rule_20211123_156 | 0.991453 | 1.0 | 0.0267 | 0.995708 | (X['account_number_num_fraud_transactions_per_... | 1 |
RGDT_Rule_20211123_95 | 0.991453 | 1.0 | 0.0267 | 0.995708 | (X['account_number_num_fraud_transactions_per_... | 2 |
RGDT_Rule_20211123_94 | 0.991453 | 1.0 | 0.0267 | 0.995708 | (X['account_number_num_fraud_transactions_per_... | 1 |
RGDT_Rule_20211123_104 | 0.991453 | 1.0 | 0.0267 | 0.995708 | (X['account_number_num_fraud_transactions_per_... | 2 |
[16]:
X_rules_test.head()
[16]:
Rule | RGDT_Rule_20211123_100 | RGDT_Rule_20211123_156 | RGDT_Rule_20211123_95 | RGDT_Rule_20211123_94 | RGDT_Rule_20211123_104 | RGDT_Rule_20211123_157 | RGDT_Rule_20211123_130 | RGDT_Rule_20211123_158 | RGDT_Rule_20211123_173 | RGDT_Rule_20211123_124 | ... | RGDT_Rule_20211123_107 | RGDT_Rule_20211123_77 | RGDT_Rule_20211123_13 | RGDT_Rule_20211123_83 | RGDT_Rule_20211123_76 | RGDT_Rule_20211123_111 | RGDT_Rule_20211123_68 | RGDT_Rule_20211123_106 | RGDT_Rule_20211123_208 | RGDT_Rule_20211123_39 |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
eid | |||||||||||||||||||||
975-8351797-7122581 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ... | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
785-6259585-7858053 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ... | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
057-4039373-1790681 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ... | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
095-5263240-3834186 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ... | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
980-3802574-0009480 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ... | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
5 rows × 218 columns