Global Convenience Store Chain Makes Major Inroads in New Market

Industries: Retail, Supply Chain

Use Case: Demand Forecasting

Architecture: Knowledge-First AI (AI-CALM)

One of Japan’s largest supermarkets is undergoing unprecedented growth in China. Accurate forecasts are vital to deciding on location, size, and timing for new store openings. Engineers struggled to build ML forecasting models, and accuracy was severely limited. This is due to local factors only local managers “knew” in their heads, such as the proximity and ease of access to nearby subways, types of foot traffic, etc., that were difficult or impossible to model without sufficient data.

By using Aitomatic’s Knowledge-First ML, they are able to generate forecasts in minutes with great accuracy. Managers can now focus on more business-critical tasks. The supermarket chain plans to triple its presence in China from approximately 2000 stores to 6000 by 2025, making them one of the top three supermarkets in China.

Before, it took a whole month to make decisions. With Knowledge-First ML, we can generate hundreds of accurate forecasts with one click.

Costs Savings

$10M

Increased
New Stores Openings

20X

Establishing
Market Dominance

Priceless

The Problem

Data-Only ML Falls Short

Eyeing an opportunity to establish a dominant presence in the gigantic China market, one of Japan’s largest supermarket operators is rapidly opening new stores. However, the geography is huge and dense, with various hyper-local influences, their own buying patterns, and preferences. In addition, other operators race to capture market share, creating fierce competition.

The challenge is to run fast while exercising precise control to avoid multiplying errors. For example, accurate revenue and traffic forecasts are needed to determine where, when, and how much to set up and invest in a particular store location.

The sheer volume of analyses to support decision-making is overwhelming. The hyper-locality has made it challenging, even impossible, to collect sufficient stratified data to build accurate ML models.

The management team’s data-only ML approach automated a significant forecasting workload but with low accuracy. Naturally, no one wants to build an automated system to make erroneous decisions faster.

The Solution

Knowledge-First ML Comes To The Rescue

1

Translate
domain-specific knowledge

2

Combine it with
Machine Learning

3

Deploy Production-ready
ML Solution

The solution came naturally when the AI team realized that better forecasts always required local managers’ heuristics.

They applied Aitomatic’s Knowledge-First ML to combine expert knowledge with data. The experts have extensive domain knowledge in specifying heuristics for estimating the effects of age groups, working hours, nearby facilities (schools, hospitals, subway stations), etc. The new forecasting engine met both the speed and accuracy targets while handling the diversity in the data within and across regions.

k-SWE

Knowledge can be applied to different segments

Using the Knowledge-First k-SWE (“Segmented-World Ensemble,” similar to Stacked Ensemble) architecture, the engineers first modeled the experts’ knowledge in individual features segments, such as age group between 18 to 44, residential and office locations, morning and evening business hours, among others, in the data.

Different segments of data were used to train specific ML models that captured the unique behaviors of the population represented by each segment. The models were then ensembled together to make accurate predictions under a broad range of conditions.

This approach has proven to work very effectively for different regions, generating forecasts with variances well less than the required margin of error.

Ultimately, the forecasting engine empowered experts to make quick decisions, with high confidence, on hundreds of new store openings per month. This has enabled the supermarket operator to plan on tripling the number of stores in China over the next three years.