Customer Engagement Playbook

Overview

This is a TIBCO Data Science Team Studio hosted demo. For access - please contact askdatascience@tibco.com and someone from the TIBCO Data Science team will assist you.

Telco X, a fictitious telecommunications service provider, is running a marketing campaign for the SmartPhone J7 model. The campaign was not performing well. The marketing manager of Telco X would like to target customers with relevant offers and improve conversion rates. She works with a data scientist to improve the campaign targeting model and have it deployed.

Business goals

• Target customers with relevant offers of SmartPhone J7, improve conversion rates

• Provide visibility into the data science project to business and technical users

• Deploy improved model into operation

The TIBCO Data Science Team Studio workspace

Access to this demo can be requested by emailing askdatascience@tibco.com

Use the rest of this document as a guide to explore the workspace to see how the Telco X team collaborates in TIBCO Data Science Team Studio to achieve its business goals.

Project collaboration

Data Science is a team sport that involves members with business and technical backgrounds. The workspace provides an environment for data science project members to communicate and stay up-to-date on each other's work.

Click on the "Overview" tab of the workspace. Here, project members can read and comment on activities in the workspace in a Slack-like interface.

Team Studio provides access control to workspaces and their data sources. You have been added to this workspace as a member because you are a user of Team Studio's hosted demo environment. In real-life projects, only members working on the project are added to the workspace. "Workspace Membership" on the right shows the members and their roles in the project.

Click on the "Milestones" tab of the workspace. Here, the project team keeps track of phases in the project and their target dates.

The Data

Data Science projects often require blending data from different sources.

Click on the "Data Source" tab of the workspace. This workspace is configured to work with data from a PostgreSQL database named ?DB? and a Hadoop cluster named "CDH". Permissions can be enforced so that project members have access only to data sources they are authorized to use.

Processing of the analytics operations, such as joins, aggregations, and training machine learning models, are all pushed down and run within the data sources. There is no movement of data to bring them to analytics engines. This makes Team Studio a scalable solution for Data Science projects on Big Data.

Click on the "Data" tab of the workspace. This tab shows the datasets from the data sources that have been associated with the workspace. Think of these as "bookmarking" datasets relevant to the project.

This workspace uses the following datasets:

Orders - Transactional records of orders placed at Telco X

Customers - Information on Telco X's customers

Campaigns - Customers' responses to previous campaigns

Demographics - Customers' demographics, such as gender, age, frequent shopper class etc.

SearchLog - A log of the search terms that customers used on Telco X's website

The Analysis

As we will see next, analyses in Team Studio are built in a point-and-click visual workflow editor. This makes it easy for non-technical users to understand the operations performed on the data.

Click on the 'Work File" tab. This is where all files containing the analyses for the project are found. These include workflows, models, documents, Jupyter Notebooks, SQL Files, links to visual dashboards and other external resources.

Let us recall that the goal of this analysis is to target customers with relevant offers of SmartPhone J7 and improve conversion rates. We will use historical data on customer purchases of SmartPhone J7 to build a predictive model to predict if a customer is likely to buy the SmartPhone J7 phone. The model can then generate predictions for other customers and Telco X would then make the SmartPhone J7 promotion offer to customers whom the model predicts are likely to buy.

The analysis for this demo is built out in the following steps.

Data Preparation and Feature Engineering

Click on the item "1. Data preparation and feature engineering?. This brings you to a preview of the workflow. Click on "Open Flow" on the top left to open the workflow in the Workflow Editor.

Click "Run" on the list of commands on the top right to run the workflow. As each operator completes the run, you can click on any of them to inspect the resulting data in the Results Console.

Since our goal is to figure out if a customer is likely to buy the SmartPhone J7 phone model, the objective of this step is to prepare a dataset that has a customer record in each row. Each record contains information about the customer that would help us predict if this customer is likely to buy the SmartPhone J7 phone model. To train a model to make this prediction, each record should also have a column that indicates if the customer has bought the SmartPhone J7 phone model before.

This workflow does this by counting the number of orders placed by each customer for each product in the order history and then filtering down to the SmartPhone J7 model only. It joins together the tables that have customer-level information and adds a column called "HasBought" to indicate whether a customer has bought SmartPhone J7 in the past.

The SearchLog dataset contains information on the words customers used for search on Telco X's website. Product searches is an indication of a customer's interest in the product. This makes the SearchLog a valuable source of information that can help us identify customers who are likely to buy SmartPhone J7. The workflow operates on the SearchLog to count the number of times a customer performed "SmartPhone J7"-related searches. It then adds this as a column called ?SmartPhoneJ7_search_count? to the table of customer records prepared by this workflow.

In this step, we have calculated values that fill new columns for the customer records. These new columns are meant to help us build models that make better predictions. Deriving new columns to be used as predictors is known as feature engineering in data science projects.

Click on "Close" on the list of commands on the top right to close the workflow.

Click on the ?Work Files? tab to return to the list of work files in the workspace.

Model Training

Now that the dataset is prepared, we can use it to build a predictive model to predict if a customer is likely to buy the SmartPhone J7 phone.

Be sure that you are under the "Work Files" tab and can see the list of work files in the workspace. Click on the item '2. Model training". Click on "Open Flow" on the top left of the workflow's preview to open it in the Workflow Editor. Click "Run" on the list of commands on the top right to run the workflow. As each operator completes the run, you can click on any of them to inspect the resulting data in the Results Console.

This workflow starts its operations on the dataset that was prepared in the previous step. It splits the data into a training set and a validation set. The training set is used to train models; the validation set is used to compare the models' predictions against the known results in the validation set to assess model quality.

The models are trained to predict the column "HasBought" (called the target variable) using a number of predictors. The workflow trains and compares 2 Logistic Regression models. The model trained in "Logistic Regression (without search)" is the model that is performing poorly in the SmartPhone J7 campaign. The model trained in 'Logistic Regression (with search)" is the improved model proposed by the data scientist to the marketing manager.

The difference between the 2 models is that the latter uses one additional predictor, "SmartPhoneJ7_search_count", derived from the search logs. The ROC operator and the Confusion Matrix show that this model performs better and is able to better target customers who are likely to buy the SmartPhoneJ7 model.

The Export operators in the workflow export this improved model into the workspace so that it can be deployed into operation.

Click on "Close' on the list of commands on the top right to close the workflow. Click on the "Work Files" tab to return to the list of work files in the workspace.

Model Scoring and Deployment

Having trained an improved model, the model can then be used to predict if customers are likely to buy the SmartPhone J7 model. Telco X can then use the predictions to determine if the SmartPhone J7 offer is to be made to its customers.

Be sure that you are under the "Work Files" tab and can see the list of work files in the workspace. You should find the model files exported from the previous workflow. These are the files with 'Targeting_SmartPhoneJ7" in their names.. The models exported from workflows can be used to generate predictions on new customer data in batches (batch scoring) or in real-time applications (real-time scoring). Team Studio exports models in different formats, indicated by the file extensions, for them to be scored in different engines.

Batch scoring

Under the 'Work Files' tab, click on the item ?3. Batch Scoring?. Click on 'Open Flow' on the top left of the workflow's preview to open it in the Workflow Editor. Click "Run" on the list of commands on the top right to run the workflow. As each operator completes the run, you can click on any of them to inspect the resulting data in the Results Console.

This workflow loads the previously exported model and applies it to a customer dataset to make predictions. Predictions on whether a customer would by the SmartPhone J7 phone is given in the PRED_LOR column of the Prediction operator's results.

This workflow can be scheduled to run regularly on new batches of customer data. This is done by adding it to a scheduled Job. Any scheduled jobs for the workspace can be seen by clicking on the "Jobs" tab.

Real-time scoring

Team Studio comes with a light-weight real-time scoring server, called ScORE Server, where models in the PFA format can be deployed. The scoring service is exposed as RESTful API calls that applications can use to get predictions on new customer records.

Click on the "Engines" tab. The previously exported model has been deployed in a ScORE Server engine.

Telco X's application developers can use the scoring service APIs in their applications to get predictions from the model. These APIs may be called, for instance, from a web store or a customer service center app to determine if the SmartPhone J7 offer is to be made to a customer.

Summary and Feedback

You have just explored the workspace where the Telco X team collaborates to target customers with relevant offers of SmartPhone J7 and improve conversion rates. The targeting models created through the course of this project are deployed into operation.

Related Materials

Learn how Telco X's marketing manager can monitor on live dashboards, in real-time, the campaign offers made to customers who visit Telco X's website and the customers' responses that drive profit:

• Read more and download the live dashboards - Live Campaign Monitoring with TIBCO Spotfire and TIBCO Spotfire Data Streams