Skip to main content
There are three numbers you can set. That is the whole surface.

​
Target

The accuracy each training attempt aims for. This is the one that matters. If your dataset came from Feature Refinery, use the value it carries. The refinement measured what the cleaned data can actually reach. The console pre-fills it; over the API, read recommendedThreshold off the dataset and send it as performanceThreshold. If you uploaded your own split, there is no measured value, so you are choosing. Start at something the data plausibly supports and move from there. On hard, imbalanced problems that is often 0.7 to 0.8, not 0.95.
A target far above what the data supports is the most expensive mistake available here. Every attempt runs to its time cap without converging, and the run can settle on a model that predicts a single class for every row. Accuracy will look reasonable, F1 will be zero, and the grid will look busy and confident. Check F1 and recall, not just accuracy.
Lower target, faster run, weaker model. Higher target, slower run, and past a point, no model at all. Move it in steps of 0.01 to 0.02 and re-run; runs are cheap.

​
Buckets

How finely the data is split for the internal cross-validation. More buckets means more thorough validation and a longer run. In Feature Refinery you choose 4, 10 or 20. Start at 10. Go to 20 on a wide or subtle dataset where you want the column decisions to be more reliable, and accept the extra time. Go to 4 on very small data, or to get a first answer quickly. On a dataset the slider goes from 4 to 1000. A dataset built from a refinement inherits the bucket count the refinement ran at, and it cannot be changed: the refined files were produced at that setting. Rough guidance when you are choosing yourself:

​
Scaling

An internal parameter of the training algorithm, between 8 and 499. 19 is the working default and covers nearly every case. Change it only after target and buckets have been tried, and change it in large steps rather than small ones. If a value works better on your data, it will be obvious.

​
When a run disappoints

​
A working method

  1. Refine first. Let the run measure the target.
  2. Train with that target and the defaults.
  3. Look at the candidates table, not just the champion. Pick on the metric your problem cares about.
  4. If you want more, move the target by 0.01 to 0.02 and run again. Compare false decisions.
  5. Only then touch scaling.
Runs are minutes, not days. Two or three of them beats a long argument about the right setting.

See it applied

The German Credit walkthrough, including the target that came out of refinement.