Uplift Models for better marketing campaigns (Part 1)

Abhay Pawar
Oct 7, 2018 · 7 min read

Recently, Facebook has been showing me ads relating to physiotherapy clinics and the weird thing is I am actually suffering from lower back pain. It’s very likely that Facebook picked this up from the pages that I visited. This can be said to be one of the best instances of targeted marketing and it is opening up a lot of new avenues for businesses that use these platforms.

source

Facebook has been a recent addition to the marketing arsenal of businesses. There has been a traditional way of marketing which companies have been using for a long time and those are Direct to Customer(D2C) channels namely, email, phone calls & SMS. There is one fundamental difference between Facebook-like ads platforms and D2C channels. Facebook is primarily used for brand building and acquiring new customers based on the rich demographics data that Facebook has about its users. On the other hand, D2C campaigns are also used for acquiring new customers but are not very effective due to lack of information about the prospective customers. But, these channels perform like a rock-star in case of existing customers of the business. These channels are very effective for promotional offers, cross-selling and customer service to existing customers. The obvious reason being that the customer is already aware of the business and is more likely to read the mail and take some action on it.

How do D2C campaigns work?

Small businesses typically send out their promotional mails/SMS to all their customers as well as to new email ids/phone numbers that they acquire from other sources. Not much thought is put into it, largely because the number is small and the activity is not very cost intensive. But, things get complicated in case of large businesses.

Lets take an example of one the banks I worked with. This bank has over 10 million active customers and more than 20 product offerings. They run cross-selling D2C campaigns for each of these products. In addition to these, they run campaigns related to service activation (like e-statement activation, Internet Banking activation), promotional offers relating to a specific product (like cash-back on Credit Card) and numerous other campaigns. At this point, it might have been clear that it just doesn’t make sense to send each of these campaigns to all the 10 million customers. And that’s where the art of selecting a target group comes in. Luckily, banks are always in a good position when it comes to target group selection because they have a lot of information about their customers. Now, lets pick a Debit Card(DC) activation campaign for our example here. The bank wants to increase its DC usage because it is cheaper than other transaction channels like ATM. There are two ways in which a campaign manager selects a target group:

  1. Heuristic Approach: This is an approach where the campaign manager takes some subjective calls to select the target group based on his experience. In our example of DC activation, the campaign manager might go after customers who are young, live in urban areas, have good balances in account and use ATM. Such, target selection works well to some extent because it selects customers that a campaign manager already knows will get activated on DC. But, the biggest problem with this approach is that it is not data driven and might leave out other customers who are very likely to get activated on DC.
  2. Propensity Models Approach: Companies have started moving towards the propensity models based approach largely because they are data driven and don’t need much human intervention while selecting the target base. A propensity model tells you based on customer profile, the propensity of a customer to do the intended action. In our case, it is getting activated on DC. These are built using past data by using any standard classification algorithm. The dependent variable would be, if the customer is using DC and the independent variables would be data like customer demographics, transaction pattern, product holding, etc.

Well, lets say we have a propensity model but how to do the target group selection? Before doing that we need to understand how the performance of a D2C campaign is measured.

Campaign Performance

The most widely used methodology to gauge campaign performance is Lift. After selecting a target group, it is divided randomly into two similar looking groups with population distribution of say, 70%:30%. The larger group(70%) is called the Treatment Sample and the other group is called Control Sample. As the name suggests, the treatment sample is targeted with the email/SMS and the control group is not targeted with any email/SMS. We then give the treated customers some time (typically 3–4 weeks) to take the intended action and then compare the performance of the two groups in terms of the activation rate. Lets assume some numbers:

Treated DC Activation rate (Treatment Response rate) : 14%

Control DC Activation rate (Control Response rate) : 12%

Lift is nothing but the difference between the response rates of the two groups i.e. 14%-12% = 2%.

The control response rate tells you how many customers would have gotten activated irrespective of the campaign and the treatment response rate tells you how many additional customers (which is the lift: 2%) got activated due to the campaign. Higher the lift, better the campaign and higher the ROI will be.

Now that we know what lift is, let’s get back to the target group selection. Selecting a target group once we know the propensity of a customer is more of a trial and error activity. It is typically observed that customers who have high propensity also lead to higher lift in a campaign. This is usually true when the penetration of the service which is being campaigned is low. So, based on this information campaign managers typically select the high propensity customers in the target group. Although, this might not be always true!

In the adjoining plot, the propensity scores for DC activation of customers are mapped on X-axis and DC activation rate of corresponding customers are mapped on the Y-axis. DC activation rate is defined as % of customers using debit card. Remember, these are based on past data and not on any campaign data. The DC activation rates just tell you what percentage of these customers are using DC. As can be clearly seen, the customers with higher propensity scores have higher DC usage rates. Customers in these high propensity score buckets but are not active on DC, are selected in the target group.

What are uplift models?

At this point, you have a good idea of how a D2C campaign is run and how the performance/ROI is measured. Using propensity models makes the process of selecting target group automated and easy to manage. But, there is a very fundamental disconnect in the way we select target group using propensity models and the way we gauge the campaign performance. Model just tells you the propensity of a customer to use a product or get activated on a service irrespective of a campaign. It does not tell you anything about the lift that the customer will have when targeted. Whereas, all we care about in a campaign is the final lift. We want customers who will go from response rates of 10% to 30% and not 90% to 92%. Uplift model helps us to predict the lift for each customer and then empowers us to target only the high-lift customers.

Based on the actions that a customer takes, they can be put into 4 groups. The four groups are shown in the adjacent figure. We are interested in finding out customers who fall in the first group: Persuadables. And if we target only these customers, then the campaign will give high lift as they wouldn’t have taken any action without any campaign (Low response rates in Control group and high response rates in Treatment group). The second group: The Sure Things is the one which would have anyway activated and hence should be left out of the campaign. And obviously, group three and four should also not be targeted. Thus, to summarize target Group 1 and leave out all the other groups.

The question is does the propensity model enable us to find these groups. The answer is NO. The customers who have high propensity mostly lie in group 1 and 2 but by using just the propensity model we cannot pin-point which groups they exactly fall into.

And that is the wonder of uplift models! It can segregate the customers into these 4 groups and help you in running a high ROI campaign. Rather than using just the customer-related data which the propensity model does, uplift model also uses previous campaign’s data to predict the lift of each customer in the future campaigns.

Part 2 of this blog series which talks about how to build uplift models is here.

Abhay Pawar

Written by

Machine Learning @ Instacart. Blogger. Loves data. twitter/linkedin/gmail: @abhayspawar

Welcome to a place where words matter. On Medium, smart voices and original ideas take center stage - with no ads in sight. Watch
Follow all the topics you care about, and we’ll deliver the best stories for you to your homepage and inbox. Explore
Get unlimited access to the best stories on Medium — and support writers while you’re at it. Just $5/month. Upgrade