One of the hottest areas for subscription services right now are ingredient and meal-kit delivery services. Two of the companies in this space – Blue Apron and Plated – were both in the news. Blue Apron went public last year, raising $300 million in the process, and Plated was acquired by the grocery chain Albertsons. These companies make weekly deliveries of pre-measured ingredients with recipes and instructions on how to put the meal together. These meal-kit services have distinctive supply chains and difficult forecasting issues. Plated is using machine learning and artificial intelligence to improve their value chain.
Haftan Eckholdt, PhD, the Chief Data Science Officer at Plated, spoke on the research they are using to solve these problems at the Machine Learning Innovation Summit in New York City on December 11th.
First, Mr. Eckholdt was brought in to build a data science team, consisting of data engineers, systems engineers, analysts, and modelers. Companies with longer standing IT and analytics operations often have these groups reporting up to different functions and leaders. According to Mr. Eckholdt, having all these groups report to him allows him to shortcut many budget and resource issues; in this way, his group can begin to tackle projects in weeks that would take other companies months to come to grips with.
Mr. Eckholdt started by looking at Plated’s strategy, business models, and process flows. He begins by looking for decision points. “Where do things move from A to B? These are model opportunities, these are hiring requirements, and these are like catnip to a good candidate.”
Once he has mapped the processes, the diplomacy phase begins. depends on the company’s goals, ROI, and key performance metrics are all things that need to be considered before creating a rank-ordered list of problems to solve. But it is also about the cultural fit with a department, the difficulty of the effort, and the level of interest within his own staff. The last might sound surprising, but less interesting problems take longer to complete, have more errors, and therefore, may demand additional resources.
When Mr. Eckholdt arrived, one of the first business issues his group tackled was the question of how many recipes they should offer? When he arrived, there were twelve recipes per week, with six new ones being swapped in each week. But was this the right number of offers? More offers could lead to increases in customer acquisition and retention, but too many choices could overwhelm customers and drive up supply chain costs.
Currently, a meal for two, delivered twice a week, costs $47.80 plus $7.95 for delivery. Each week there are 20 recipes to select from and two dessert choices. A customer does not have to order every week, they can skip weeks.
Answering that first question required some interesting analyses. An ensemble of logistic regressions was their starting point to leverage previous segmentation research conducted on a small group of customers. Two of five segments were considered target customers based on their attitudes toward food and use of meal kits. Segmenting all customers, provided by the logistic ensemble, allowed the data team to leverage more data. Key insights came from studying the differences in utilization between highly engaged and less engaged target customers. Highly engaged customers are using the product all the time, as it is designed. The comparison groups are using the product less often and they tailor the offering to suit their needs.
Differences in how customers use the product provide product development insights. The distribution of proteins in the meals needs to mirror the distribution of proteins consumed by customers. While 12 proteins satisfied highly engaged customers, they found that by offering 20 proteins they could fulfill the needs of the moderately engaged and tempt them into becoming more engaged. They did a similar analysis of their recipes in terms of fats, salts, and sugars to see what kinds of recipes would have the best ROI for the moderately engaged group.
This lead the team to the next problem: they decided to offer 20 recipes, but this many recipes would make it difficult for customers to choose what to eat; they might easily become overwhelmed and choose nothing. So machine learning was then used to make recommendations for customers.
Mr. Eckholdt spoke of developing new taxonomies of food, multi-arm bandit experimentation, natural language processing, and neural networks to solve this problem. Fortunately, for this listener at least, he did not try to describe these techniques in any detail. But he did describe the data used in these analyses. The machine learning team closely examined the cuisine tags their chefs used to describe what was in the meals, image features – the types of pictures customers react positively to, the molecular nutrition, as well as the words used to title and describe the recipes. Plated is currently testing what amounts to as one recommendation model per customer. This is quite different from the more typical path of training a few models on large groups of customers.
Now that demand is better understood there are some new opportunities for the forecasting team. Better forecasting leads to better inventory management. The first forecasting problem they will solve will be around better demand forecasts for recipes ordered infrequently.
Now that Plated has been acquired by Albertsons, there will undoubtedly be new problems to solve. The Plated network consists of a large network of small facilities, thus reducing shipping times with each additional node. Albertsons has more than 2300 stores, 27 distribution centers, and 18 factories. They are already experimenting with selling meal kits in stores.
Mr. Eckholdt began his talk by admitting that part of the appeal for him in taking the Plated job was that he was a “foodie.” The tagline he displays on his LinkedIn profile perhaps sums his speech up the best: Food + data engineering + data analytics + data modeling = JOY!