Inference using Tecton

In this section of the notebook, we'll load our fraud detection model from mlflow, layer in the transaction information we'll assume we're passed in (type of transaction and amount), and layer it with features about the user we retrieve from Tecton.

Ordinarily, this would be running behind a REST API so you can do this inference in real time.

# Here, we import the tecton library and use its `get_feature_vector` convenience function, 
# but we can also save the import and issue an HTTP request directly to the Tecton endpoint
import tecton
import mlflow
from dotenv import load_dotenv, find_dotenv
import os
import pandas as pd

# Load the TECTON_API_KEY from a .env file - in production you might
# pass this in as an environment variable or use a secrets manager
load_dotenv(find_dotenv())

True

ws = tecton.get_workspace('prod')
fs = tecton.get_feature_service('fraud_prediction_service')

tecton_api_key = os.environ['TECTON_API_KEY']
tecton.set_credentials(tecton_api_key)

vec = fs.get_feature_vector(join_keys={"user_id": "C1986564990"}).to_pandas()

Here's what we get back from Tecton

vec.T

	0
days_since_last_transaction.days_since_last	82
transaction_aggregates.amount_mean_12h_1h	None
transaction_aggregates.amount_mean_168h_1h	None
transaction_aggregates.amount_mean_1h_1h	None
transaction_aggregates.amount_mean_24h_1h	None
transaction_aggregates.amount_mean_72h_1h	None
transaction_aggregates.amount_mean_960h_1h	None
transaction_aggregates.transaction_sum_12h_1h	None
transaction_aggregates.transaction_sum_168h_1h	None
transaction_aggregates.transaction_sum_1h_1h	None
transaction_aggregates.transaction_sum_24h_1h	None
transaction_aggregates.transaction_sum_72h_1h	None
transaction_aggregates.transaction_sum_960h_1h	None
user_age_days.age	10477
users_credit_score.credit_score	None

req_cols = pd.DataFrame({
  'amount': 123.45,
  'type_cash_in': True,
  'type_cash_out': False,
  'type_debit': False,
  'type_payment': False,
  'type_transfer': False
}, index=[0])

req = vec.merge(req_cols, left_index=True, right_index=True)

req_ordered = req[['amount', 'type_cash_in',
       'type_cash_out', 'type_debit', 'type_payment', 'type_transfer',
       'transaction_aggregates.transaction_sum_1h_1h',
       'transaction_aggregates.transaction_sum_12h_1h',
       'transaction_aggregates.transaction_sum_24h_1h',
       'transaction_aggregates.transaction_sum_72h_1h',
       'transaction_aggregates.transaction_sum_168h_1h',
       'transaction_aggregates.transaction_sum_960h_1h',
       'transaction_aggregates.amount_mean_1h_1h',
       'transaction_aggregates.amount_mean_12h_1h',
       'transaction_aggregates.amount_mean_24h_1h',
       'transaction_aggregates.amount_mean_72h_1h',
       'transaction_aggregates.amount_mean_168h_1h',
       'transaction_aggregates.amount_mean_960h_1h',
       'users_credit_score.credit_score',
       'days_since_last_transaction.days_since_last', 'user_age_days.age']]

And here's the final vector we'll pass to the model, incorporating the additional features we were passed in (transaction type columns) and Tecton features about the user in the correct order

req_ordered.T

	0
amount	123.45
type_cash_in	True
type_cash_out	False
type_debit	False
type_payment	False
type_transfer	False
transaction_aggregates.transaction_sum_1h_1h	None
transaction_aggregates.transaction_sum_12h_1h	None
transaction_aggregates.transaction_sum_24h_1h	None
transaction_aggregates.transaction_sum_72h_1h	None
transaction_aggregates.transaction_sum_168h_1h	None
transaction_aggregates.transaction_sum_960h_1h	None
transaction_aggregates.amount_mean_1h_1h	None
transaction_aggregates.amount_mean_12h_1h	None
transaction_aggregates.amount_mean_24h_1h	None
transaction_aggregates.amount_mean_72h_1h	None
transaction_aggregates.amount_mean_168h_1h	None
transaction_aggregates.amount_mean_960h_1h	None
users_credit_score.credit_score	None
days_since_last_transaction.days_since_last	82
user_age_days.age	10477

As we load the model from databricks, you need to have all of the libraries installed that you use in your model, and using the same versions as well. You can see which versions are captured in your model by inspecting the conda.yaml file in your MLFlow run in your model directory under "Artifacts"

# We load the model from mlflow - in Databricks we transitioned the model trained above to "production" so we can specify that 
# stage below
os.environ['MLFLOW_TRACKING_URI'] = 'databricks'
model = mlflow.sklearn.load_model('models:/fraud_detection/Production')

# Predict will return a dataframe. Since we only passed in one vector to predict, res will be length 1.
res = model.predict(req_ordered)

# This is the result of the prediction (0 = not fraud, 1 = fraud)
res[0]

0