Hacker News Reader

Comment by mkrfox | Hacker News Reader

That data goofiness issue is pretty much use case #1 for Mechanical Turk. Pay a bunch of people a few cents to do a sensibility check on the data, collect 10 for each item, and put any that are less than 100% through another round.

Not even all that expensive. 10 cents each, 75,000 items, 10 checks each. $75,000 is quite a bargain for that kind of data validation.

It'd take a day or two.

(There's probably some horrible flaw with this idea)

View on HN

There's no real way a turker would know if the data is bad. Really only the merchandiser at HQ who was checking with the supplier would have had the source data (dimensions, units for those dimensions, case quantities, and so on) or warehouse workers who actually receive the product and can physically look at it. Suppliers sometimes change product attributes, quantity per case or pallet, or packaging without giving advance notice so the warehouse often ends up making these updates as they notice them, if they do. Data quality in supply chains is a constant struggle for many companies. It sounds like Target got hit bad with this due to a perfect storm of poorly trained merchandisers and SAP having dummy data that probably looked somewhat reasonable to the untrained eye.

Some of their issues were with the measurement system (English vs metric) and order of dimensions (length * width * height or width * length * height).

I can see a turk system where they show a picture of the product and ask if the unit of measurement looks correct (being 2.5x off is pretty easy to guess) and if the order looks correct. They might not get the width and the length correct but the height could probably be guessed. Course that can probably be done with computer vision.

So for $75k they could get it so at least the product would fit on the shelving units. That could easily pay for itself in a day. Products that failed these two simple tests should be looked at more closely for other issues.

True, but again, that assumes they have a picture of the data. If the product was sold online through target.com, a picture may exist somewhere. Otherwise, most company's item master data tables (that I've seen) don't include a picture. Also, they chose not to use data out of the old systems, because they wanted to avoid getting it out of the old and into the new system. This seems like it should've been the easier route but for whatever reason their legacy software probably made it much more difficult than needed.

Reply on news.ycombinator.com