Company selling consumer goods on e-commerce channel want to reduce customer returns

Question

I have a use case as below

One company selling products online would like to understand how to reduce returns from customers. They on an average have 25% returns. They have data for orders which have been returned and which have not been returned. The data includes all features of product and customer and where they were shipped.
My thinking is that certain item or customer attributes may be linked for them to understand returns from customers better.
What I would like to understand is in this situation what kind of ML models or statistical methods could be used and basically how to proceed as I have only basic understanding of ML

Some additional details we have Data for past 8-9 months (around 8000 order lines) of all orders placed and retuned on e commerce platforms. example data for one of their sales category Shoes Shoe attributes like

• brand size (Around 10-Sizes)
• colour (around 15-20 colours)
• shoe description (around 20 - 100 characters)
• shoe type (training, running , around 10 such classification)
• Gender Customer attributes like
• Customer pin code
• Gender - Order Attributes Order Attributes like
• Order date (DD/MM/YYY) I wanted to classify this as ‘day in month’ and ‘Day in week’ to see if timing of order has any effect on returns
• Payment type (COD, Prepayed)

Answer 1

You could try logistic regression, assuming your sample size is sufficiently large for that (you did not tell us). One advantage of that is that you wil get an estimated probability for a consumer return, as a function of the covariates used. Covariates could be both describing properties of buyers, and properties of products.

After question got edited with more details:

You do have many features, some categorical with many distinct values. So, I would try regularized logistic regression, maybe group lasso, see Why use group lasso instead of lasso?.

• variable shoe description might need special treatment, since it is a natural language description. Maybe try some kind of embedding? See this list of posts with answers.

• date, as you proposed, but to see for changes with time you could also represent it via splines.