An ensemble does not automatically improve performance
On raw data, the recorded RF + SGD ensemble reached 0.8279 validation macro-F1, just below Random Forest at 0.8293.
Keep a simple baseline and measure what the added method actually contributes.
Question
Does combining Random Forest and SGD improve validation macro-F1 over either model alone?
Experiment
A recorded comparison on raw genomic data evaluated Random Forest, SGD, and an RF + SGD ensemble on a validation split.
Recorded results
| Model | Validation macro-F1 |
|---|---|
| Random Forest | 0.8293 |
| SGD | 0.7860 |
| RF + SGD ensemble | 0.8279 |
Limits
These are results from one recorded validation. The ensemble threshold was selected on that validation set; the values do not establish independent test performance.
Interpretation
In this run, the ensemble did not exceed Random Forest. Adding a model by itself did not improve the chosen metric.
Keep a simple baseline and measure what the added method actually contributes.
Next experiment
Rebuild a leakage-safe pipeline, reserve an independent test set, and compare the simple baseline with the ensemble under the same folds and documented threshold selection.
Origin & contribution
Educational genomic-classification project. Values come from an editorial synthesis of saved notebook outputs supplied for this update; the notebooks are not in this checkout.
The original implementation and precise personal contribution still need source review. This note organizes the supplied observation and its limits; it does not report a new run.