Uplift model

Uplift Modeling , also known as incremental model , true lift model or net model ( English ), a method for modeling behaviors (English predictive modeling ) that the additional effects of treatment (such. As a direct marketing campaign) to a predicts individual behavior.

application

The Uplift Modeling is in customer relationship management ( Customer Relationship Management ) for up-selling , cross-selling ( cross-selling ) and as a model for customer loyalty applied. It has also been used in personalized medicine . The uplift model uses a random control group not only to measure the effectiveness of a marketing campaign, but also to predict which changes in the behavior of a target person (more precisely: target group) can be generated. It is therefore a new data mining technique that is mainly used in the areas of financial services, telecommunications and in retail for direct marketing for up-selling, cross-selling, customer churn and retention .

Measurement of uplift

The uplift of a marketing campaign is usually defined as the difference in reactions between a selected group and a random control group. This allows the marketing team to look at the impact of a single marketing campaign in isolation and measure its effectiveness. Marketing teams can use budgets better for a success-increasing effect of their marketing campaigns, i.e. a result that is higher than that in the control group. The table below provides the details of a hypothetical marketing campaign. The number of responses and the calculated response rate are shown. For this campaign, an uplift of 5 percentage points is achieved in the response rate, which means that 50,000 more responses were achieved through the marketing campaign.

groups	Number of customers	Feedback	Response rate
Treated group	1,000,000	100,000	10%
Control group	1,000,000	50,000	5%

Traditional model for feedback

In the traditional model for feedback, selected customers are usually selected and an attempt is made to create a predictive model that separates the best feedback from non-feedback using a number of predictive modeling techniques . Decision trees or regression analyzes are typically used. Only the selected customers are used to create the model.

Uplift model

In contrast to this, the uplift model includes selected customers and control customers in order to create a prediction model that focuses on the additional feedback. To understand this type of model, a basic segmentation is proposed, which customers divide into the following groups (whose names are based on the designations: Persuadables , Sure Things , Lost Causes , Do Not Disturbs or Sleeping Dogs by N. Radcliffe):

The accessible : Customers who only respond when they are addressed via a marketing campaign .
The safe ones : Customers who respond, regardless of whether they are addressed or not.
The resistant : Customers who never respond, even if they are spoken to.
The silent ones : Customers who are unlikely to respond when they are approached or even give negative feedback (e.g. fail to make a planned purchase)

The only segment that provides additional feedback is the accessible. The uplift model supports an assessment technique which can subdivide customers into the groups described above. Traditional models of feedback mostly relate to the secure and the accessible , as these models cannot distinguish between these two groups. Nor can traditional models differentiate between the resistant and breastfeeding .

Return on investment

The uplift model makes it possible to concentrate exclusively on additional feedback and is therefore able to generate very good returns (English return on investment ) for conventional sales promotion measures and customer loyalty activities . For example, if only the accessible are involved in an outbound marketing campaign, the overarching contact costs and the return on investment per customer can be dramatically improved.

Reduction of negative effects

One of the most effective uses of the uplift model is to avoid negative effects in customer loyalty campaigns. In the telecommunications and financial services industries , customer loyalty campaigns often encourage customers to renew their contract or insurance contract. The uplift model makes it possible to remove those customers who probably would not have quit, the silent ones , from the campaigns. As a result, they are no longer contacted and therefore not encouraged to take an action (e.g. change of provider).

Use for A / B and multivariate tests

It is seldom the case that there is a simple selected group and a control group. Often times the choice is a variety of variations of messages or multidimensional contact strategies that are classified as simple treatment. Using A / B or multivariate tests, the uplift model can help to understand whether the variations in tests result in a significant uplift comparison with other target criteria, such as behavioral indicators or demographic indicators.

History of the uplift model

The first approach to a true response model can be traced back to Radcliffe and Surry. Victor Lo also published content on this, as did Radcliffe, who supported this with a useful FAQ section on his website. Similar approaches are described in personalized medicine . A detailed representation of uplift models, their history, their creation and the differences to classic modeling, as well as meaningful evaluation techniques, along with comparisons of different software solutions and an economic representation of various impact scenarios can be found here.

Web links

Abby Johnson explains how it works in this video broadcast.

Notes and references

^ N. Radcliffe (2007). "Identifying who can be saved and who will be driven away by retention activity". Stochastic Solution Limited
^ NJ Radcliffe & PD Surry. "Differential response analysis: Modeling true response by isolating the effect of a single action." Proceedings of Credit Scoring and Credit Control VI. Credit Research Center, University of Edinburgh Management School (1999)
↑ Lo, BBY (2002). "The true lift model". ACM SIGKDD Exploration Newsletter. Vol. 4 No. 2, 78-86. 1
^ Radcliffe, NJ (2007). "Using Control Groups to Target on Predicted Lift: Building and Assessing Uplift Models," Direct Marketing Analytics Journal, Direct Marketing Association.
↑ The Scientific Marketer FAQ on Uplift Modeling
↑ Cai, T., Tian, L., Wong, PH, Wei, LJ (2009). "Analysis of Randomized Comparative Clinical Trial Data for Personalized Treatment Selections". Harvard University Biostatistics Working Paper Series, Paper 97.
^ R. Michel, I. Schnakenburg, T. von Martens (2019). "Targeting Uplift". Springer, ISBN 978-3-030-22625-1

[1] N. Radcliffe (2007). "Identifying who can be saved and who will be driven away by retention activity". Stochastic Solution Limited

[2] NJ Radcliffe & PD Surry. "Differential response analysis: Modeling true response by isolating the effect of a single action." Proceedings of Credit Scoring and Credit Control VI. Credit Research Center, University of Edinburgh Management School (1999)

[3] Lo, BBY (2002). "The true lift model". ACM SIGKDD Exploration Newsletter. Vol. 4 No. 2, 78-86. 1

[4] Radcliffe, NJ (2007). "Using Control Groups to Target on Predicted Lift: Building and Assessing Uplift Models," Direct Marketing Analytics Journal, Direct Marketing Association.

[5] The Scientific Marketer FAQ on Uplift Modeling

[6] Cai, T., Tian, L., Wong, PH, Wei, LJ (2009). "Analysis of Randomized Comparative Clinical Trial Data for Personalized Treatment Selections". Harvard University Biostatistics Working Paper Series, Paper 97.

[7] R. Michel, I. Schnakenburg, T. von Martens (2019). "Targeting Uplift". Springer, ISBN 978-3-030-22625-1