Answer ML fundamentals and diagnostics questions

You are taking a timed online assessment with multiple-select and numeric-response questions.

1) Confusion-matrix metrics (multiple select)

A binary classifier is evaluated on 200 examples. For each option below, the confusion matrix is given as (TP, FP, FN, TN).

Select all options where:

• Recall $= \frac{TP}{TP+FN}$ is > 0.90 , and
• False Positive Rate (FPR) $= \frac{FP}{FP+TN}$ is < 0.10 .

Options:

• A: (95, 8, 5, 92)
• B: (92, 12, 8, 88)
• C: (180, 18, 20, 182)
• D: (45, 1, 5, 149)
• E: (91, 9, 9, 91)

2) Ensemble learning for loan-default prediction (multiple select)

You built an initial classifier to predict whether a customer will default on a loan, but its accuracy is only slightly better than chance. You consider using an ensemble method.

Select all statements that are true:

• (1) If the dataset contains both linear and non-linear relationships, ensemble learning can impair performance compared to most approaches.
• (2) Modern ensemble learning techniques can improve overall model interpretability.
• (3) Ensemble learning techniques can be time-intensive to train.
• (4) Ensemble learning techniques typically create overfitted models.
• (5) If the dataset contains both linear and non-linear relationships, ensemble learning can improve performance compared to most approaches.
• (6) None of the above.

3) Decision-tree split criteria (multiple select)

You are choosing an impurity measure to score candidate splits in a classification decision tree.

Select all options that are valid impurity/split criteria:

• Entropy
• Classification Error
• Gini index
• Pruning
• None of the above

4) Training loss increases every epoch (multiple select)

You train a model to detect intrusion attempts. You notice that training loss consistently increases every epoch.

Select all rationales that could plausibly cause this:

• Regularization is too high
• Step size (learning rate) is too large
• Regularization is too low
• Step size is too small
• None of the above

5) Learning-curve diagnosis (multiple select)

You are training a sentiment regressor that predicts a score in $[-1.0, +1.0]$ using 47 features.

You observe this pattern:

• Training error decreases steadily and becomes very low.
• Validation error decreases initially, then starts increasing and stays much higher than training error.

Select all actions that best address the problem:

• Reduce the size of the training data
• Include a regularization component to your model
• Increase the number of features included in your data
• Increase the number of epochs used to train your model
• Choose a more complex modeling technique
• None of the above

6) Random Forest vs Gradient Boosting (training speed focus) (multiple select)

You need a fast-to-train baseline fraud classifier. You are choosing between a Random Forest and a Gradient Boosting Machine (GBM).

Select all statements that are true and relevant to training speed:

• (1) GBM fits trees sequentially rather than independently like Random Forest.
• (2) GBM is harder to overfit than Random Forest.
• (3) Random Forest is slower than GBM for real-time prediction.
• (4) GBMs are typically more accurate than Random Forest for anomaly detection.
• (5) Random Forest has fewer parameters than GBM.
• (6) None of the above.

7) Manual forward pass in a small neural network (numeric)

Compute the output of the following neural network. Assume:

• All biases are 0 .
• Hidden activations $f_1$ and $f_2$ are linear (identity).
• Output activation $f_3$ is sigmoid : $\sigma(z)=\frac{1}{1+e^{-z}}$ .

Inputs:

• $x_1 = 1.2$
• $x_2 = -0.7$

Hidden layer (2 units):

• $h_1 = 0.5\,x_1 + (-1.0)\,x_2$
• $h_2 = (-0.25)\,x_1 + 0.75\,x_2$

Output layer (1 unit):

• $z = 1.0\,h_1 + 0.5\,h_2$
• $\hat{y} = \sigma(z)$

Return $\hat{y}$ rounded to the nearest thousandth (3 decimals).