1

I trained Vertex AI forecasting AutoML model one with target column as String and other numeric input features as String then I trained another AutoML model with target column as float and other input features as Integer.

The predictions are different for both the models. The data is same only the datatypes/schema changed.

Google documentation says:

When you train a model with a feature with a numeric transformation, Vertex AI applies the following data transformations to the feature, and uses any that provide signal for training:

  • The value converted to float32.

So both the data should be same even after transformation. Why would results be different? Is it possible?

Wytrzymały Wiktor
  • 11,492
  • 5
  • 29
  • 37
Hitesh Somani
  • 620
  • 4
  • 11
  • 16

1 Answers1

0

I have follow the steps to have a forecasting model as show on Build an AutoML Forecasting Model with Vertex AI and reach the conclusion that vertex AI compress a lot of the steps of the prediction model generation so it can be easily operate by users.

I think the most reasonable answer for your observation among strings and numeric values resides in the way data processing is performed to generate our prediction models. I think you will not find inside vertex AI documentation as it would mean to disclose how vertex AI code works and handles its Feature Engineering and train steps to generate the models, which is protected.

Regardless, Lets speculate a bit, I think the difference among datatypes conversion might occur when datatype is converted and passed to the algorithm for processing. Lets said a linear regression sample, you will find that the slightest variation on data conversion can affect the outcome of your prediction model which could also be what is happening here.

Betjens
  • 1,353
  • 2
  • 4
  • 13