Campaign Optimization - Predicting How and When to Contact Leads

Typical Workflow

  • Step 1: Server Setup and Configuration
  • Step 2: Data Prep and Modeling
  • Step 3: Operationalize
  • Step 4: Deploy and Visualize
    • When a business launches a marketing campaign to interest customers in some new or existing product(s), they will typically use a set of business rules to select leads for their campaign. Machine learning can be used to help increase the response rate from these leads.

    This solution package shows how to use a prediction model to increase the response rate to a campaign by recommending how to contact (for example, Email, SMS, or Cold Call) as well as when to contact (day of week and time of day) each lead identified for use in a new campaign.

    To demonstrate a typical workflow, we’ll introduce you to a few personas. You can follow along by performing the same steps for each persona.

    Step 1: Server Setup and Configuration with Ivan the IT Administrator

    Let me introduce you to Ivan, the IT Administrator. Ivan is responsible for implementation as well as ongoing administration of the Hadoop infrastructure at his company, which uses Hadoop in the Azure Cloud from Microsoft. Ivan created the HDInsight cluster with ML Server for Debra. He also uploaded the data onto the storage account associated with the cluster.

    The cluster has been created and data loaded for you when you used the Deploy to Azure button on the Quick Start page. Once you complete the walkthrough, you will want to delete this cluster as it incurs expense whether it is in use or not - see HDInsight Cluster Maintenance for more details.

    Step 2: Data Prep and Modeling with Debra the Data Scientist

    Now let’s meet Debra, the Data Scientist. Debra’s job is to use historical data to predict a model for future campaigns. Debra will develop these models using HDInsight, the managed cloud Hadoop solution with integration to Microsoft ML Server. 

    After analyzing the data she opted to create multiple models and choose the best one.  She will create two machine learning models and compare them, then use the one she likes best to compute a prediction for each combination of day, time, and channel for each lead, and then select the combination with the highest probability of conversion - this will be the recommendation for that lead.

    Debra will develop her R scripts in the Open Source Edition of RStudio Server, installed on her cluster's edge node. You can follow along on your own cluster deployed by to Azure automatically. Access RStudio by using the url of the form:
    http://CLUSTERNAME.azurehdinsight.net/rstudio.

    Now you’re ready to follow along with Debra as she creates the scripts needed for this solution.

    The steps described below each create a function to perform their task. The individual steps are described in more detail below. The following scripts are then used to execute the steps.
    • campaign_main.R is used to define the data and directories and then run all of the steps to process data, perform feature engineering, training, and scoring.

      The default input for this script uses 100,000 leads for training models, and will split this into train and test data. After running this script you will see data files in the /Campaign/dev/temp directory on your storage account. Models are stored in the /Campaign/dev/model directory on your storage account. The Hive table recommendations contains the 100,000 records with recommendations (recommended_day, recommended_time and recommended_channel) created from the best model.
    • Copy_Dev2Prod.R copies the model information from the dev folder to the prod folder to be used for production. This script must be executed once after campaign_main.R completes, before running campaign_scoring.R. It can then be used again as desired to update the production model.

      After running this script models created during campaign_main.R are copied into the /var/RevoShare/sshuser/Campaign/prod/model directory.
    • campaign_scoring.R uses the previously trained model and invokes the steps to process data, perform feature engineering and scoring. Use this script after first executing campaign_main.R and Copy_Dev2Prod.R.

      The input to this script defaults to 30,000 leads to be scored with the model in the prod directory. After running this script the Hive table recommendations now contains the recommended values for these 30,000 leads.

    Below is a summary of the individual steps used for this solution.

    1. The first few steps prepare the data for training.
      • step0_data_generation.R: This file was used to generate data for the current solution - in a real setting it would not be present. It is left here in case you'd like to generate additional data. For now simply ignore this file.
      • step1_data_processing.R: Uploads data and performs preprocessing steps -- merging of the input data sets and missing value treatment.
      • step2_feature_engineering.R: Performs Feature Engineering and creates the Analytical Dataset. Feature Engineering consists of creating new variables in the cleaned dataset.  SMS_Count, Email_Count and Call_Count are computed: they correspond to the number of times every customer was contacted through these three channels. It also computes Previous_Channel: for each communication with the Lead, it corresponds to the Channel that was used in the communication that preceded it (a NULL value is attributed to the first record of each Lead). Finally, an aggregation is performed at the Lead Level by keeping the latest record for each one.
    2. Now she is ready for training the models, using step3_training_evaluation.R. This step will train two different models and evaluate each.

      The R script draws the ROC or Receiver Operating Characteristic for each prediction model. It shows the performance of the model in terms of true positive rate and false positive rate, when the decision threshold varies. </p>

      The AUC is a number between 0 and 1. It corresponds to the area under the ROC curve. It is a performance metric related to how good the model is at separating the two classes (converted clients vs. not converted), with a good choice of decision threshold separating between the predicted probabilities.  The closer the AUC is to 1, and the better the model is. Given that we are not looking for that optimal decision threshold, the AUC is more representative of the prediction performance than the Accuracy (which depends on the threshold).

      Debra will use the AUC to select the champion model to use in the next step.

    3. Finally step4_campaign_recommendations.R scores data for leads to be used in a new campaign. The code uses the champion model to score each lead multiple times - for each combination of day of week, time of day, and channel - and selects the combination with the highest probability to convert for each lead. This becomes the recommendation for that lead. The scored datatable shows the best way to contact each lead for the next campaign. The recommendations table (Recommendations) is used for the next campaign the company wants to deploy.

    4. After creating the model, Debra runs Copy_Dev2Prod.R to copy the model information from the dev folder to the prod folder, then runs campaign_scoring.R to create recommendations for her new data.
    5. Once all the above code has been executed, Debra will use PowerBI to visualize the recommendations created from her model.

      You can access this dashboard in either of the following ways:

      She uses an ODBC connection to connect to the data, so that it will always show the most recently modeled and scored data.
    6. A summary of this process and all the files involved is described in more detail here.

    Step 3: Operationalize with Debra

    Debra has completed her tasks. She has executed code from RStudio that pushed (in part) execution to Hadoop to clean the data, create new features, train two models and select the champion model She has scored data, created recommendations, and also created a summary report which she will hand off to Bernie - see below.

    While this task is complete for the current set of leads, our company will want to perform these actions for each new campaign that they deploy.

    In the steps above, we saw the first way of scoring new data, using campaign_scoring.R script. Debra now creates an analytic web service with ML Server Operationalization that incorporates these same steps: data processing, feature engineering, and scoring.

    campaign_deployment.R will create a web service and test it on the edge node. If you wish, you can also download the file campaign_web_scoring.R and access the web service on any computer with Microsoft ML Server 9.0.1 or later installed.

    The service can also be used by application developers, which is not shown here.

    Step 4: Deploy and Visualize with Bernie the Business Analyst

    Now that the predictions are created and the recommendations have been saved, we will meet our last persona - Bernie, the Business Analyst. Bernie will use the Power BI Dashboard to learn more about the recommendations (second tab). He will also review summaries of the data used to create the model (first tab). While both tabs contain information about Day of Week, Time of Day, and Channel, it is important to understand that on the Recommendations tab this refers to predicted recommendations to use in the future, while on the Summaries tab these values refer to historical data used to create the prediction model that provided the recommendations.

    You can access this dashboard in either of the following ways:

    Bernie will then let the Campaign Team know that they are ready for their next campaign rollout - the data in the Recommendations table contains the recommended time and channel for each lead in the campaign. The team uses these recommendations to contact leads in the new campaign.