The modern start to consumer analytics began in 1916 when Clarence Saunders founded a chain of grocery stores in the U.S. called Piggly Wiggly. These stores offered self-service; take a basket and pick what you want from the shelves. Saunders then began to redesign food shopping, methodically arranging things in order to appeal to how customers shopped, for example putting candy and other impulse items at the checkout.
This simple understanding of customer preference and impulses revolutionalized retail strategy and the modern Planogram is the current end point of this ceaseless investing in analytics around customer preferences. We now know that customers are more likely to buy items at eye level or at the ends of rows. This type of aggregate thinking was carried over into the planning of eCommerce businesses.
When commerce moved to the web, there were suddenly many ways to connect to customers and potential customers at scale. You did not need to have customers physically go to your store, you send the store to them. Retailers have been quick to adapt in-store analytics to online analytics.
In-store, we know that people will respond positively to promotional messages and discounts. Online, onsite messaging and email campaigns were developed to take advantage of this type of behavior, and they work! Even simple texting and email messaging works well when sent to a large undifferentiated group of people who have visited your store. This is because among your customers there will be a group that responds positively to these messages and any site that aims for a good CX will include positive and engaging messaging.
However, management wanted to understand more. By simply linking the products that people have purchased (basket analysis) to the time of day they purchased (pattern analytics) and their geographical location, many teams can place a monetary value on each household or person in a specific area. This is why you might see several convenience stores next to one another. Even with a small number of clients visiting, if the basket value of each is projected to be valuable enough, a store will be set up.
When data scientists got to work, that’s when the analytics got serious. Loyal customers did not need attention with this type of messaging and the customers who had already left were not worth pursuing. They noticed that there will always be a group of customers that are immune to the blare of a trumpet and will need more care and attention.
Even so, most companies try to identify the customers most at risk of churn, and put them all in one group – even the ones with no hope of becoming a returning customer and blast an email to them in hopes that they will return. Eva Ascarza, an Associate Professor of Business at Columbia University has outlined wonderfully that she “consistently finds that customers identified as being at the highest risk of churning are not necessarily the best targets for proactive churn programs.“
In fact, she finds these simple promotional efforts to be no better than random! It is to these customers that the scientists turned their attention to using a technique called Uplift Modelling or more formally, Casual Inferencing. At its very basic, it is a technique used to estimate a customer’s desired action, the “uplift” is based on some action – such as a promotion. Where casual inferencing diverges from the email blast is in its laser approach on customer modelling.
To put it in a simple way, we can show how this might work in a non-technical way. Imagine you are a fruit seller and you want to sell durians to your customers. You wouldn’t want to target every one so you try to find customers that like a slightly sweet taste, like large size fruits and are not concerned about smell. A Venn diagram might look like this:
As you can see, if you were not concerned about smell, you might choose a fruit with no smell like a ramban. If you wanted a large sized fruit you might choose a mango, and if you wanted something slightly sweet it might be a tart apple, but encompassing all three would be the shaded area. So, if you are sending promotions it should be to the users in the sweet spot right in the center. Your incremental uplift will have a much higher probability of being successful and you won’t anger the users that are looking for an apple by sending them a durian.
Delvify provides recommendation tools that layer several filters of experience and customer data into a unified personalized approach. You start with Visual Search but end with happiness. The right product delivered to the right person. Customers have to feel that you care for them as individuals and recommendations in particular will be able to showcase products that users are finding by analyzing the specific features that they are searching for. Delvify’s Visual Search and Recommendations are able to enable that seamlessly for you so that your customers will be able to navigate through your platform and check out with a smile on their face!
There are many libraries that look into uplift modeling. Our personal favorite is Pylift because it is worked around scikit-learn which, not only looks into implementation of the uplift model but also into understanding its evaluations.
We now know that the motivation behind uplift models is its ability to foresee the incremental value with respect to a treatment. We were enlightened with outcome models which will be able to tell us the chances/outcomes of an action given a treatment. However, due to the approach of uplift modeling we can now additionally understand the lift caused by subtracting the chances/outcomes of an action given no treatment. Hence, this lift adjustment is the most accurate way to implement the uplift models but these kinds of loss functions are quite difficult to optimize. Moreover, you would need to additionally custom build it for every objective you set out to achieve. Hence, Pylift is the answer to all of this as it takes a “transformed outcome” approach, since we can transform the outcome label and use the scikit-learn algorithms in hand.