SO1 holds guest lecture on targeting AI for Sloan’s MBA program
Guest lectures are frequently used to connect the two worlds of science and the industry. Often, both worlds are far apart from each other as scientists are way ahead with their developments compared to the ones from the industry. But sometimes, both worlds join forces to create something bigger…
In November 2018, I was invited to the Massachusetts Institute of Technology to present the current status of our latest joint research project covering AI-powered target marketing in Grocery Retailing. For more than a year I have worked with MIT’s top marketing scientists to further push the edge of current AI usage in the retail industry.
Why is this important?
Recent research shows that “the personalized customer” was the top strategic priority for retailers in 2018. Yet, the real world of grocery retail still looks way different. Customers are flooded with irrelevant mass promotions, such that retailers waste billions of dollars on wrong customers and wrong discounts.
Customers might get a discount on a product they would have purchased anyway – called cannibalization. Or they learned about regular patterns of promotions which attracts them to stockpiling once a product is on sale. These are just two examples of how wrong discounts instantly lead to a decrease in revenue.
Targeted marketing is a solution to overcome these weaknesses. Over time, the tools and possibilities for such targeted approaches have evolved and became more and more precise. Palmatier has defined three different eras of marketing (based on Palmatier (2017). Marketing Strategy. Palgrave. ISBN: 9781137526236) where one major differences is the size of the market being targeted:
Classical retail promotions like current circulars are a tool existing since the mass marketing era. They target a whole market at once with a single message. Most likely, this will not exactly fit all potential customers being targeted. The goal of personalization is to individualize these messages according to the customers’ individual needs and preferences.
Increasing the level of personalization automatically leads to smaller segments or niches of a market being targeted. With simple rule-based issuance of coupons and other promotions, a well-defined segment of the population can be treated better compared to mass marketing. However, true personalization requires a real one-to-one communication towards customers.
Truly leveraging customer level data is key to excel in this discipline. One cannot rely on mostly qualitative and ad hoc data and simply trusting in the effectiveness of marketing resources any longer. Having a well developed IT infrastructure and demanding mathematical proof of the effectiveness of used models in communication is necessary instead. (source: Palmatier)
In this spirit SO1 (Segment of One) has delivered a concept of personalized messaging in grocery retail. The so-called “feed” of individualized promotions – which exactly matches the customer’s personal preferences and needs. It might contain content of any kind, e.g. classical price-offs, information about issued industry coupons or even recipe suggestions. SO1’s personalization also includes each individual’s actual willingness-to-pay for each product. Therefore, SO1 can determine the optimal discount level – as well as in which cases simply showing an ad is a sufficient trigger for additional purchases without any discount at all. This leads to true incremental revenue for the retailer.
SO1< achieves this high level of personalization by scoring each available offer or campaign for each customer, based on the customer’s purchase likelihood per product. This allows us to rank and order the campaigns and only send out those, that increases the likelihood of incremental sales or profit or customer satisfaction.
This capability of predicting the purchase likelihood is the core of our AI. What distincts the SO1 AI component from other engines is the combination of several models at the same time, so-called “model stacking”. Instead of using one model or type of algorithm for all decisions along the process, we add different models for each decision type in the process. The type and complexity of models changes with the complexity of decisions to be made.
Base of the model stack is a rather simple Boosted Tree Model which as well can be referred to as an Expert Model. Expert Features like the purchase history or shopping regularity are added to make first assumptions. On a secondary level, additional models are integrated to account for impact factors like purchase timing or price sensitivity for each single product. The most complex layer contains Deep Features like autoencoders, representation learning, or deep recommender systems.
Its continuous improvement of this model stack with the newest algorithms and developments is part of the research cooperation between SO1 and MIT. We currently investigate the cross-category purchase choice and the capabilities of different model types to influence the components of revenue.
As findings show, deep neural networks have the biggest potential to increase new purchases.
These findings will support us in developing scalable deep-neural networks that use rich market basket data and loyalty card data to predict consumers’ purchases across all categories in a retailer’s product assortment.
SO1 and MIT will continue to work on the next generation of AI for retail. During the next weeks, we will post insights into a new tool developed by SO1 that improves timing for such developments drastically through detailed simulation: The SO1 Supermarket Gym