Testing Your AIWhich K Should You Pick?
Small k and large k each have a weakness.
Small k: too pessimistic
With , every model trains on only half the farms, so it is weaker than the final model trained on all 20 will be. Its score is too gloomy: Harpreet's 2-fold error was 0.56, against 0.38 for 5 folds — you saw it as the highest point on the chart. To make the models more like the final one, you push k
Large k: costly and jumpy
With a large , every model trains on almost the same farms, so the models are near-copies of each other, and the score hangs on single farms in tiny validation folds. And every extra fold is one more model to train.
What Does It Cost?
Harpreet's model goes national: the agriculture department has records of 10 lakh farms, and training one model on them takes 3 minutes. Check all three plans.
The middle way wins. In practice almost everyone uses k = 5 or k = 10: the models see 80–90 % of the data, the score is steady, and the cost is only 5 or 10 models.