You can access a significantly larger sample of the platform's content for free by logging in with your Gmail account. Sign in now to explore.

Assume that we are running a simple logit to solve a problem but find the results to be unsatisfactory. What are some ways we might improve the model?

To improve a model, it will help understanding what is currently not working. Here are some ideas:

Learning curves will help isolate the problem.

If this is a high-variance issue (higher error in test than training), then we can use RF, remove features, use regularization, get more data, etc.

If this is a high-bias issue (training and testing error high), we can add features, gradient boosting, add complexity.

We can also try the following:

Run optimization longer (more iterations) \(\rightarrow\) Fixes optimization algorithm

Change optimization method \(\rightarrow\) Fixes optimization algorithm

Hyperparameter tuning \(\rightarrow\) Fixes optimization objective

Alternative models (e.g., non linear models) \(\rightarrow\) Fixes optimization objective

Ablative analysis could also be useful if this was a multistep framework.

Prediction error, Bias-variance tradeoff, Diagnostics, Learning curves

- Bias-variance equation Easy (Bias-variance tradeoff, Formula derivation)
- Multiclass evaluation metrics Medium (Multiclass, Diagnostics)
- Example of high-bias and high variance Medium (Bias-variance tradeoff)
- Multicollinearity Medium (Multicollinearity, Linear regression)
- Bias-variance biased estimator Medium (Bias-variance tradeoff)