0

I have used the ML.net modelbuilder to apply multi-classification.

My label (prediction) in this classification is the product class code. To make this prediction, I used the following features:

  • manufacturer id
  • manufacturer
  • product description
  • Brand
  • product series
  • weight

This works to predict the product class when I enter the product description. But if I leave the product description empty and enter only the brand and the manufacturer, the model does not predict anything.

Could this be because, for example, the brand and manufacturer are often the same and do not have as much variety as the product description in the training data?

Patrick
  • 331
  • 3
  • 18

1 Answers1

0

I would guess that the model 'latched' onto the description text features, because they are richer and allow for more accurate prediction.

Therefore, for the specific model you trained, your other features hold less value (or maybe no value?), and the description is strictly essential.

If you want your model to work well for products with no description, try injecting some description-free examples into your training data (or take it to the limit, and strip off all descriptions), and see if that helps.

Overall, answering the question of 'how/why did the model predict this outcome?' falls under 'model understanding', which is more of an art these days.

Zruty
  • 8,377
  • 1
  • 25
  • 31
  • Thanks for your answer sir. I agree with the point on "description text features, because they are richer and allow for more accurate prediction". The other features are indeed sometimes empty or null etc. I focus now on description. I have a question left, how many labels(to predict) could a single model do? – Patrick Dec 04 '19 at 07:38