Exam AWS Certified Machine Learning - Specialty topic 1 question 274 discussion

Correct Answer B why not D ? While introducing a small L2 regularization might help in some cases, it doesn't address the main issue, which is that the L1 regularization is too strong. The primary problem needs to be addressed first.

Decrease L1 will ensure that all features are not zero

From Claude 3: Based on the AWS documentation and industry best practices, introducing a small L2 regularization parameter in addition to the existing L1 regularization (Option D) is a recommended approach to address overfitting while retaining the feature selection capabilities of L1 regularization [3]. The combination of L1 and L2 regularization, known as Elastic Net regularization, can effectively handle collinearity and provide a balance between sparsity and weight shrinkage, potentially improving the model results and addressing the overfitting issue [4]. While decreasing the L1 regularization parameter (Option B) may seem like a logical step, it does not directly address the overfitting problem and may not be effective in the specific scenario where all features have zero weights. Introducing a small L2 regularization parameter (Option D) is a more appropriate solution based on AWS documentation and industry best practices

Seems like a problem of too much regularization, so I would start by decreasing L1

Ans: B Decreasing the L1 regularization parameter would reduce the penalization on the coefficients, allowing some features to contribute to the model without being driven to zero. This adjustment can help in achieving a balance between overfitting and underfitting.

leaning towards C .. confusing though

Letter B is correct, since L1 is very strong and is eliminating all variables. Letters A - C would harm even more and Letter D would not change the current result at all.

you applied an L1 regularization parameter and got all features having zero weights, it means that your parameter value was too high and caused too much shrinkage. This resulted in a model that underfits the data and has poor performance.

Also don’t think any options given improve the model

Can’t apply L2 on zero weights