Currently, one model is trained and evaluated, then used downstream to generate predictions. This approach is tried and true, but even if the model is slightly off in "understanding" the training data, it could generate inaccurate results. One approach to minimize the chances of this happening is to implement a model ensemble, where we have 2 or more models trained on subsets of the overall training dataset, then evaluated on the evaluation data. These models would be smaller by nature, and each could learn some specific, reproduceable niche of the overall dataset. Combining the outputs of these models in downstream prediction theoretically could yield more accurate results.
Done when:
- model training is refactored to generate N models each with varying focus on input training data
Currently, one model is trained and evaluated, then used downstream to generate predictions. This approach is tried and true, but even if the model is slightly off in "understanding" the training data, it could generate inaccurate results. One approach to minimize the chances of this happening is to implement a model ensemble, where we have 2 or more models trained on subsets of the overall training dataset, then evaluated on the evaluation data. These models would be smaller by nature, and each could learn some specific, reproduceable niche of the overall dataset. Combining the outputs of these models in downstream prediction theoretically could yield more accurate results.
Done when: