Turning Art Into Science
Capturing Associate Experience
Photo by Girl with red hat on UnsplashOne of Walmart’s great strengths is the institutional knowledge of its associates. With almost sixty years of experience in retail, this knowledge is vast and powerful. While we are increasingly applying data science, analytics, and machine learning (ML) to retail problems, a portion of it is truly “The Art of Merchandising.”
How can we tap into the expert knowledge of our merchants and understand the artistic sense and feel that these experts have for merchandising? One traditional way is through user studies and interviews. This is an essential first step, but it often misses deep-seated, obvious knowledge. Using data science, ML, and operations research, we can augment and enhance the abilities of our expert associates.
Science Projects
Two projects that are under development now are Category Space Placement (CSP) and Assisted Draw. These are components of the Assortment Excellence (AEX) Platform, which is designed to provide a unified merchant tools experience.
Category Space Placement uses store floor plan layout data and a category strategy document to place modulars for a category onto a floor plan. We use operations research techniques to optimize the layout while also adhering to the constraints specified in a layout strategy document.
Assisted Draw is given an assortment of items and draws an optimal placement of the items on a store modular, also using a strategy with constraints such as brand or color adjacencies. This is known in the industry as a planogram. A related component is the Assortment Selection Engine, which can produce a space-aware assortment that can be passed to Assisted Draw.
Closing the Gap Between Art and Science
In one study of 42 store layouts, manually-drawn layouts were compared with layouts generated by the CSP engine. With the given strategy, the CSP engine could draw acceptable layouts for 80% of the stores. This is already a win because it saves weeks of drawing for those stores. Manual drawings were produced for all 42 stores, and it seemed like it might be better at first glance.
However, only 9% of the manual drawings were conformant based on the specified constraints. As these experts have been successfully drawing store layouts for years, which of the non-conformant choices that they made were not hard requirements? If we identify those, they can be marked as soft constraints, and the CSP engine can draw a higher percentage of layouts.
Art School
As part of a series of Subject Matter Expert (SME) sessions for Category Space Placement, one of the category space experts said, “I always move the scooters next to the bike rack.” This was not something that was explicitly specified in the strategy document. It was too obvious or automatic for this associate to even state explicitly.
We could find this out through more extensive user interviews, but by using an iterative approach, we can use ML to fill in the “obvious” gaps in the strategy. If we were able to study the changes these associates made over time, we could uncover this pattern using ML and add it to the strategy. To enable this iterative approach, we can design our systems with a few things in mind.
Three Steps Towards Enabling ML
Given Walmart’s long history, there are still many systems that are not automated or integrated. As we design new systems, there are three things to keep in mind that will enable improvement through ML as time goes on:
- Instrumentation
- Feedback
- History
Instrumentation refers to keeping track of user decisions or inputs. This is the real key to turning Art into Science. Each time that an expert associate makes a decision, capturing and saving it allows us to use it as input data for a ML problem.
One of the most expensive tasks in supervised learning is labelling data. For example, if a merchant or a customer identifies an item as a substitute for another item, recording that decision gives us a labeled data point that we can use for training associates about substitutions or demand transfer models.
This ties in with the second point: feedback. If we present a list of items to add to an assortment, we should record which of these was chosen and which wasn’t. If we feed that back into the assortment engine, we can learn the Art of assortment selection through the expertise of the associate. This is combined with the Science of choosing an optimal assortment. By the same token, the assortment engine can provide reasons why particular items were chosen to be added to the selection as feedback to the Artist.
The final point is history, which means to keep older copies or versions of documents. Developers using a versioning system for code, such as git, will have an idea how this works. As one example, in the current system for category floor plans, writing a finalized floor plan so that it can move to the execution phase overwrites the existing floor plan.
With a versioning system, users could work on multiple versions of a floor plan and then commit final versions while retaining previous versions. As a history of these versions builds up, we can use ML to spot differences and trends that can be used as input (feedback again!) to create a more refined layout strategy. A versioning system like this is currently being implemented as a part of AEX.
Finishing Touches
As we develop these systems, we don’t need to be perfect right out of the gate. We can get 80% there and learn from associate feedback as the system is used. This way, we get short-term wins with increasing progress over time.
To make this possible, we need to enable our future systems for access by ML systems using instrumentation, feedback, and history.
Turning Art into Science was originally published in Walmart Global Tech Blog on Medium, where people are continuing the conversation by highlighting and responding to this story.