Rule Generator (Optimisation algorithm) Example¶
This notebook contains an example of how the Rule Generator (Optimisation algorithm) can be used to create rules based on a labelled dataset. This algorithm generate rules by optimising the thresholds of single features and combining these one condition rules with AND conditions to create more complex rules.
Requirements¶
To run, you’ll need the following:
A labelled, processed dataset (nulls imputed, categorical features encoded).
Import packages¶
[1]:
from iguanas.rule_generation import RuleGeneratorOpt
from iguanas.metrics.classification import FScore
import random
import pandas as pd
import numpy as np
Read in data¶
Let’s read in some labelled, processed dummy data.
[2]:
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 optimisation function (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.
[3]:
fs = FScore(beta=1)
[4]:
params = {
'opt_func': fs.fit,
'n_total_conditions': 4,
'num_rules_keep': 50,
'n_points': 10,
'ratio_window': 2,
'remove_corr_rules': False,
'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.
[5]:
rg = RuleGeneratorOpt(**params)
[6]:
X_rules = rg.fit(
X=X_train,
y=y_train,
sample_weight=None
)
--- Generating one condition rules for numeric features ---
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 24/24 [00:00<00:00, 275.12it/s]
--- Generating one condition rules for OHE categorical features ---
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 8/8 [00:00<00:00, 114520.25it/s]
--- Generating pairwise rules ---
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:01<00:00, 1.82it/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.
[7]:
X_rules.head()
[7]:
RGO_Rule_20211123_1 | RGO_Rule_20211123_0 | RGO_Rule_20211123_2 | RGO_Rule_20211123_5 | RGO_Rule_20211123_3 | RGO_Rule_20211123_95 | RGO_Rule_20211123_37 | RGO_Rule_20211123_124 | RGO_Rule_20211123_123 | RGO_Rule_20211123_36 | ... | RGO_Rule_20211123_303 | RGO_Rule_20211123_302 | RGO_Rule_20211123_224 | RGO_Rule_20211123_353 | RGO_Rule_20211123_221 | RGO_Rule_20211123_220 | RGO_Rule_20211123_219 | RGO_Rule_20211123_199 | RGO_Rule_20211123_196 | RGO_Rule_20211123_195 | |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
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 × 340 columns
[8]:
rg.rule_descriptions.head()
[8]:
Precision | Recall | PercDataFlagged | OptMetric | Logic | nConditions | |
---|---|---|---|---|---|---|
Rule | ||||||
RGO_Rule_20211123_1 | 0.991837 | 1.0 | 0.027547 | 0.995902 | (X['account_number_num_fraud_transactions_per_... | 1 |
RGO_Rule_20211123_0 | 0.991837 | 1.0 | 0.027547 | 0.995902 | (X['account_number_num_fraud_transactions_per_... | 1 |
RGO_Rule_20211123_2 | 0.972000 | 1.0 | 0.028109 | 0.985801 | (X['account_number_num_fraud_transactions_per_... | 1 |
RGO_Rule_20211123_5 | 0.960474 | 1.0 | 0.028446 | 0.979839 | (X['account_number_num_fraud_transactions_per_... | 1 |
RGO_Rule_20211123_3 | 0.960474 | 1.0 | 0.028446 | 0.979839 | (X['account_number_num_fraud_transactions_per_... | 1 |
Apply rules to a separate dataset¶
Use the .transform() method to apply the generated rules to a separate dataset.
[9]:
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.
[10]:
rg.rule_descriptions.head()
[10]:
Precision | Recall | PercDataFlagged | OptMetric | Logic | nConditions | |
---|---|---|---|---|---|---|
Rule | ||||||
RGO_Rule_20211123_1 | 0.991453 | 1.0 | 0.026700 | 0.995708 | (X['account_number_num_fraud_transactions_per_... | 1 |
RGO_Rule_20211123_0 | 0.991453 | 1.0 | 0.026700 | 0.995708 | (X['account_number_num_fraud_transactions_per_... | 1 |
RGO_Rule_20211123_2 | 0.958678 | 1.0 | 0.027613 | 0.978903 | (X['account_number_num_fraud_transactions_per_... | 1 |
RGO_Rule_20211123_5 | 0.943089 | 1.0 | 0.028069 | 0.970711 | (X['account_number_num_fraud_transactions_per_... | 1 |
RGO_Rule_20211123_3 | 0.943089 | 1.0 | 0.028069 | 0.970711 | (X['account_number_num_fraud_transactions_per_... | 1 |
[11]:
X_rules_test.head()
[11]:
Rule | RGO_Rule_20211123_1 | RGO_Rule_20211123_0 | RGO_Rule_20211123_2 | RGO_Rule_20211123_5 | RGO_Rule_20211123_3 | RGO_Rule_20211123_150 | RGO_Rule_20211123_96 | RGO_Rule_20211123_66 | RGO_Rule_20211123_67 | RGO_Rule_20211123_151 | ... | RGO_Rule_20211123_314 | RGO_Rule_20211123_290 | RGO_Rule_20211123_230 | RGO_Rule_20211123_268 | RGO_Rule_20211123_267 | RGO_Rule_20211123_266 | RGO_Rule_20211123_311 | RGO_Rule_20211123_310 | RGO_Rule_20211123_293 | RGO_Rule_20211123_195 |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
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 × 340 columns