Segmentation of the trading space

Model of retail space segmentation

We have a system for preparing, managing and balancing training samples, a staff of specially trained markers and a distributed computing infrastructure - this allows us to separate large data volumes of hundreds of millions of lines.

• At the input

  data for a specific month

  

• Segmentation model

  data for a specific month
  

• At the output

  monthly determination of sales outlets segments

  

How do we solve problems?

The “Smart Cashiers” company has a large amount of cash receipts data (hundreds of thousands of cash registers, daily data updates, billions of transactions).

ПCompany sets us the following task: to determine the sales outlets types (grocery store, car service, spa, etc. - 150 types in total) based on sales data from cash receipts in several outlets.

We conducted an EDA and found a number of limitations that we were able to overcome:

It is impossible to understand from sales outlets which segment of the business it belongs to.

To do this, we have developed an ML model that determines the type of outlet based on econometric data and the textual name of goods in receipts.

But the names of the goods in the receipts are indicated by the sellers themselves, the names are not uniform, not structured, in 50% of the positions there is no name at all. How does the model get such high accuracy?

Sales data in receipts are firstly carefully prepared, the semantics are analyzed for various problems (a variety of descriptions of goods and services, incorrect data entry, anomaly detection, etc.), and then the data goes through a text cleaning algorithm (cleaning texts from noise digits, forming "white" list of words, etc.)

And in the obtained data, the sample of outlets is not balanced relative to the general population, the regions’ indicators are shifted relative to the whole country

To increase the data representativeness and accuracy, as well as to eliminate possible errors, we calibrate the data (adaptive RIM-weighting), extrapolate sales on sampling (city, region, country) and apply our dynamic data quality management system before running the model.

As a result, our segmentation model for the “Smart Checkout” company showed the segment determination accuracy from 86% to 97% depending on the sales outlets’ type (150 types of outlets in total).

And now, with the help of sales outlets segmentation, we can build sales analytics in the channel, category trends, analyze the individual markets dynamics (FMCG, Beauty, Fashion, etc.) in order to solve new problems.

We will solve your problem too!