Due to Uplift modeling, companies can discover how to improve communications, according to Valeriy Babushkin, the head of Data Development at X5 Retail Group. On November 22, the expert will speak at AI Conference in Moscow on the topic: Uplift modeling: nature and mechanisms.

In an interview with Moscow AI Conference, the specialist explains why Uplift modeling is useful, what kind of businesses should apply it, and shares a case study of his company.

Interviewer: AI Conference (AIC).
Respondent: Valeriy Babushkin, Head of Data Development at X5 Retail Group (VB).

AIC: Companies in what industries do you think should use Uplift modeling?

VB: Uplift modeling is a search for the best possible communication strategy. Consequently, this method can be applied by any company working with the broad audience, offering discounts to customers, as well as preparing promotions and special offers.

The more often and longer people use any service, the more money they bring to the company. In 90%-95% of the cases, communication analytics begins with the churn model. It provides data on how likely a certain user will stop utilizing services and buying goods.

If the probability of customer churn is high, the organization offers them a discount: a customer will use the offer and keep purchasing products and using services. In turn, the company will earn on this sale more than lose on the discount.

It is important to note four scenarios when one offers customers a discount:

  1. A client uses the discount and continues to use the service. However, they would not stop purchasing company’s goods even without a discount, so money was poured down the drain.
  2. A client uses the offer and pays for the product, which they would not buy without a discount: it is a perfect scenario for the company.
  3. A discount does not lead to the purchase, as a client would not buy the product any way. In this case, the company waste money and time for communication. If it refers to SMS distribution, a major company with the multi-million audience has wasted a big sum.
  4. A message about a discount reminds a user of the long-forgotten service and they unsubscribe to company offers. Another option: a customer is tired of the endless message flow from the organization and stops contacting with it.

AIC: Is such a ML method appropriate for beginning businesses?

VB: When a company is just entering the market, it does not have a customer base or it is small. Therefore, this kind of business will hardly offer discounts to users in order to retain or return them. In general, this method is not for newcomers.

Theoretically, one can implement Uplift modeling at the initial stage, but there is nothing to build it on without a large client base.

AIC: What exactly does Uplift modeling allow to discover about users?

VB: A company does not need user data on the part of the Uplift model. The main thing is to understand how the behavior of a certain client will change after offering a discount.

If the probability that a person will buy a product without a discount is 50% and with a discount is 70%, Uplift in such a case will be 20%. Afterwards, knowing the cost of communication with a client, the company can calculate profit from this 20%. As the result, it can decide whether the discount is profitable or not.

AIC: What data does Uplift modeling use to analyze the behavior of users?

VB: Data on purchases and customer account. The model selects maximally similar clients who are different in the way of communication. Thus, it is desirable to collect as much information about users as possible: history of purchases, age, gender, etc.

AIC: How to determine what users are ready to buy goods only with a discount?

VB: It is significant to collect data that can be used for training in Uplift modeling. For this purpose, one collects a training set: alike customers. For example, a group unites 1000 people with similar behavior in terms of product preferences. A company selects 500 random people among them, sends a newsletter about a discount to them, while the rest customers do not receive offers.

Thereafter, one should monitor changes in any metrics: an average check or conversion. Changes in this case are caused by communication, for instance a newsletter.

If clients buy goods in 70% of the cases and in 80% with a discount, Uplift is 10%. Based on these data, discounts and offers are distributed goal-oriented.

AIC: What method of Uplift modeling is the most efficient in retail?

VB: There is an approach, whereby one builds two models and conduct calculations. A single model can be also enough.

The most efficient approach is to predict and model Uplift directly.

AIC: What are the further steps after receiving the results of Uplift modeling?

VB: Obtaining the results, the company knows what to do next: whom and when to offer discounts and promotions.

For example, if a company has dozens of millions of users and only one million clients require a discount, SMS distributions should cover only these million people. Distributions through the whole database will be the waste of money. Given that discounts are offered more than once a year, the sum spent for senseless newsletters can reach billions of rubles per year.

AIC: Tell us about the application of Uplift modeling based on your company. What results have you managed to achieve using this ML method?

VB: As a piloting project, we have applied Uplift modeling to the audience of 500 000 people. Subsequently, the special offer where X5 Retail Group first applied this machine learning method resulted in the growth of the average check and conversion.

The impact even on 1% of the users brings lots of money in large businesses with a huge turnover. Therefore, the bigger business, the more significant to think of its optimization using Big Data.

AIC: What will you talk about at AI Conference in Moscow?

VB: At the conference, I will talk about the ways of building Uplift models, as well as thoroughly reveal and compare three of them. Besides, I will explain what options of model building exist within each method, what their purposes are, and how to apply all of this.

Discover more about Uplift modeling and its advantages from X5 Retail Group’s expert at AI Conference!