Overview

Our Attribution Model

The successful training and predicting of the attribution (propensity) model is absolutley key to our platform. Our attribution model is trained bi-weekly as part of the crons. It is initated from update_agg_pct_batch on backend via the command fabric_train_attribution_model. This command spins up an EC2 instance of relevant size (based on client size) to train the model, once done and the model is saved to S3, the instance terminates.

Once the EC2 instance is initated, we use fabric to connect to the rsrv, passing in the relevent paramaters and running the attribution training script located at /srv/rsrv/attribution/train.R. This script on the rsrv will train a model with the most up to date data for each client, for each product/goal. The output of the script will push a model to the S3 bucket, which is then later used for predictions (visscore) on the DPS.