Full Data Set vs Sample Set Random Forest Workflow

KNIME Analytics Platform
1

Below are the same two H2O Random Forest Regression Workflows. The only difference is the top workflow is partitioning the data 80/20 Learner/Predictor, and the workflow below it feeds 100% of data to the Learner and the Predictor.

Why? it’s a small data set of 630 home sales and constitutes the full data set. There’s no sample–this is all there is.

Two questions: 1) is the second WF a correct set up (100% of data to Learner and Predictor, no partitioning) when using a complete data set, and 2) are the RMSE, MAPE, and R2 scores legitimate when using 100% of data at the Learner and Predictor nodes, or am I creating a self-fulfilling prophesy series of scores?

1
11677×751 155 KB

I realize this may be more data science philosophy than a question of structural KNIME workflow design, but I ask for your patience and kindness.

Thanks

2

If you use 100% for training you have a high risk that the nodel will not work as good on unknown data. This could somewhat be mitigated if you were to use cross validation.

Using 80/20 split is the better way to go.

1 Like
3

@mlauber71 Thanks for the advice and link :grinning:

1 Like