AI0-001 Machine Learning and Deep Learning Practice Test 8 — 15 Questions

Question 1

A data scientist is training a binary classification model to detect fraudulent transactions. The dataset contains 99.9% legitimate transactions and 0.1% fraudulent transactions. After training a logistic regression model, the accuracy is 99.9%, but the recall for the fraud class is 0%. Which of the following is the MOST likely cause?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

The dataset is highly imbalanced, and the model predicts the majority class for all instances.. The dataset has a severe class imbalance (99.9% legitimate, 0.1% fraudulent). A logistic regression model that predicts the majority class (legitimate) for every instance will achieve 99.9% accuracy but 0% recall for the fraud class, because it never identifies any positive fraud cases. This is the classic 'accuracy paradox' in imbalanced classification.

Answer

The regularization parameter is too large, causing underfitting.

Answer

The model is overfitting due to too many features.

Answer

The learning rate was too high.

Question 2

A machine learning engineer is preparing to train a deep neural network for image classification. To avoid overfitting, which TWO techniques should the engineer apply? (Select TWO.)

Accepted Answer

Use dropout regularization.. Dropout regularization is a technique that randomly drops a fraction of neurons during training, which prevents the network from relying too heavily on any single neuron and reduces co-adaptation. This acts as a form of ensemble learning and significantly reduces overfitting by improving generalization.

Answer

Increase the number of layers.

Answer

Remove all non-linear activation functions.

Answer

Reduce the training dataset size.

Question 3

A self-driving car company uses a reinforcement learning agent to navigate. The agent was trained in a simulated environment and achieved high rewards. When deployed in the real world, the agent fails to avoid obstacles. The team collects real-world driving data and uses it to fine-tune the model. However, fine-tuning leads to catastrophic forgetting of the simulated knowledge. Which technique should the team use to mitigate this? A. Increase the learning rate during fine-tuning. B. Use elastic weight consolidation (EWC) to regularize important weights. C. Train the model from scratch using only real-world data. D. Increase the number of layers in the network.

Accepted Answer

Use elastic weight consolidation (EWC) to regularize important weights.. Elastic Weight Consolidation (EWC) is a regularization technique specifically designed to prevent catastrophic forgetting when fine-tuning a neural network on a new task. It identifies the weights that are most important for the original task (simulated driving) and penalizes large changes to those weights during fine-tuning on real-world data, thereby preserving the learned knowledge while adapting to the new domain.

Answer

Increase the number of layers in the network.

Answer

Train the model from scratch using only real-world data.

Answer

Increase the learning rate during fine-tuning.

Question 4

A media company uses a natural language processing (NLP) model to classify news articles into topics. The model was trained on articles from 2015-2018. In 2023, the model's F1 score drops significantly. The data scientists find that the word embeddings no longer capture the meaning of some terms (e.g., 'covid', 'metaverse'). The model uses static word embeddings (Word2Vec) trained on the original corpus. Which solution BEST addresses the observed degradation? A. Replace static embeddings with contextual embeddings from a transformer model like BERT, then fine-tune the classifier. B. Retrain the static Word2Vec embeddings on a larger corpus from 2023. C. Apply data augmentation to the original training data by replacing words with synonyms. D. Increase the dimensionality of the static embeddings.

Accepted Answer

Replace static embeddings with contextual embeddings from a transformer model like BERT, then fine-tune the classifier.. Contextual embeddings (e.g., BERT) capture meaning based on context, adapting to new uses of words like 'covid' meaning pandemic. Fine-tuning the classifier on new data would update the model. Option A (retraining static embeddings) might capture new word senses but still assigns a single vector per word, missing context. Option B (increasing dimensionality) does not address the semantic shift. Option D (data augmentation) does not introduce new word meanings.

Answer

Retrain the static Word2Vec embeddings on a larger corpus from 2023.

Answer

Increase the dimensionality of the static embeddings.

Answer

Apply data augmentation to the original training data by replacing words with synonyms.

Question 5

Based on the exhibit, what is the most likely issue with the trained model?

Accepted Answer

Overfitting because training accuracy is much higher than validation accuracy. The exhibit shows a significant gap between high training accuracy and lower validation accuracy, which is the classic symptom of overfitting. The model has memorized the training data rather than learning generalizable patterns, leading to poor performance on unseen validation data.

Answer

Data leakage artificially inflating training accuracy

Answer

Vanishing gradients causing no learning

Answer

Underfitting due to insufficient epochs

Question 6

A financial institution is developing a fraud detection model using historical transaction data. The dataset contains over 10 million records, but only 0.01% of transactions are fraudulent. The current model uses a neural network trained with standard cross-entropy loss, and the team applies random undersampling of the majority class to create a balanced training set. However, the model still produces a high number of false positives (legitimate transactions flagged as fraud) and misses approximately 30% of actual fraud cases. The business requires that at least 95% of frauds be caught, and the false positive rate must be below 1% to avoid overwhelming fraud analysts. The team has limited resources to collect additional data and cannot change the model architecture significantly. Which approach should the team take to best meet the business requirements?

Accepted Answer

Use cost-sensitive learning by assigning a higher misclassification cost to the fraud class.. Cost-sensitive learning directly addresses the business requirement by penalizing fraud misclassifications more heavily, which shifts the decision threshold to favor recall on the fraud class while still allowing the team to tune the trade-off between false positives and false negatives. Because the team cannot collect more data or change the architecture significantly, adjusting the loss function's class weights is the most practical lever to hit the 95% recall and <1% FPR targets. It also avoids the information loss caused by random undersampling, which discards 99.99% of legitimate transactions and distorts the true class distribution.

Answer

Apply feature selection to remove noisy predictors and then retrain the current model.

Answer

Switch to an anomaly detection algorithm such as Isolation Forest or One-Class SVM.

Answer

Collect more transaction data, especially fraudulent examples, to naturally balance the classes.

Question 7

A fraud-detection team at a bank trains a gradient-boosted tree model on two years of transaction data. Only 0.4% of transactions are fraudulent. The model achieves 99.7% accuracy but flags almost no fraud. Which approach best addresses the underlying problem with how the model is being trained and evaluated?

Accepted Answer

Optimize the model using precision-recall AUC and apply class weighting or resampling to the fraudulent class.. Accuracy is a poor metric when one class is extremely rare, because a trivial majority-class predictor scores deceptively high. The real fix is twofold: train with class weighting or resampling so the learner sees the minority class, and evaluate with precision-recall AUC, which reflects performance on the fraud class rather than being swamped by true negatives.

Answer

Remove outliers using a z-score filter and standardize all numeric features before retraining the model.

Answer

Increase the number of boosting rounds and lower the learning rate until training accuracy reaches 100%.

Answer

Switch the evaluation metric to root mean squared error and report it alongside accuracy for each boosting round.

Question 8

A junior data scientist is training a supervised classification model to predict whether a loan applicant will default. The dataset has 40,000 labeled historical records with a clear binary outcome column. The team needs a model that outputs a probability between 0 and 1 for the default class. Which algorithm is the most appropriate choice for this task?

Accepted Answer

Logistic regression. Logistic regression is designed for binary classification and directly estimates the probability that an observation belongs to the positive class through the sigmoid link function. With abundant labeled data and a clear binary target, it satisfies the supervised learning requirement and delivers interpretable, well-calibrated outputs suitable for credit risk decisions.

Answer

Principal component analysis

Answer

Apriori association rule mining

Answer

K-means clustering

Question 9

A junior data scientist is asked to explain the difference between supervised and unsupervised learning to a product manager. She wants to give a single concrete example that clearly illustrates unsupervised learning. Which example should she choose?

Accepted Answer

Grouping retail customers into segments based on purchasing behavior using k-means clustering.. Unsupervised learning finds structure in data that has no labels. Clustering customers by purchasing behavior is a textbook example because k-means discovers groups from feature similarity alone. The other choices involve a known target (default status, sale price) or a reward signal (game playing), which place them in supervised or reinforcement learning rather than unsupervised learning.

Answer

Predicting whether a loan applicant will default using historical labeled outcomes.

Answer

Training an agent to play a board game by rewarding winning moves over many episodes.

Answer

Estimating a house's sale price from its square footage and location features.

Question 10

A machine learning engineer trains a decision tree to predict customer churn. The tree achieves 99 percent accuracy on the training set but only 68 percent on a held-out test set. The engineer wants to reduce this gap. Which single action is most likely to improve test performance?

Accepted Answer

Limit the tree depth and require a minimum number of samples per leaf. The large train-test gap signals high variance, meaning the unconstrained decision tree memorized training noise. Pre-pruning through maximum depth and minimum samples per leaf limits how finely the tree can split, reducing variance and improving held-out accuracy without discarding the model or leaking test data.

Answer

Train the tree on the test set as well so it learns the held-out distribution

Answer

Increase the maximum depth of the tree so it can fit more training examples

Answer

Add more features derived from the training set to give the tree more signal

Question 11

An ML engineer is tuning a random forest classifier for a medical diagnosis task and observes that training accuracy is 99% while validation accuracy is 78%. She wants to reduce the gap without discarding the ensemble approach. Which change is most likely to reduce the generalization gap?

Accepted Answer

Reduce the maximum tree depth and increase the minimum number of samples required at each leaf.. The large train-validation gap indicates that the individual trees are too complex and memorizing the training set. Constraining tree depth and requiring more samples per leaf regularizes each estimator. Because a random forest averages many decorrelated trees, this regularization typically improves validation accuracy while keeping the ensemble's variance-reduction benefits intact.

Answer

Add more trees to the ensemble until the out-of-bag error stops decreasing.

Answer

Increase the maximum tree depth so each tree can capture more interactions in the training data.

Answer

Disable bootstrap sampling so every tree is trained on the full dataset instead of a random subset.

Question 12

A machine learning engineer is training a deep neural network for image classification. The training loss decreases steadily, but the validation loss starts to increase after 20 epochs. The engineer wants to implement a technique that dynamically adjusts the learning rate during training to improve convergence and generalization. Which method should the engineer use?

Accepted Answer

Learning Rate Scheduler with exponential decay. A learning rate scheduler with exponential decay dynamically reduces the learning rate as training progresses, which helps the model converge more smoothly and can improve generalization. This directly addresses the need to adjust the learning rate during training to combat the rising validation loss, unlike fixed learning rates or other regularization techniques that do not modify the learning rate.

Answer

Stochastic Gradient Descent (SGD) with a fixed learning rate

Answer

Batch Normalization

Answer

Dropout regularization

Question 13

A junior ML engineer is asked to evaluate a binary classifier that predicts whether a bank transaction is fraudulent. The model's precision is 0.92 and recall is 0.41. The team wants to improve recall without retraining the model. Which action should the engineer take?

Accepted Answer

Lower the classification threshold for the fraud class.. Recall measures the fraction of actual positives the model captures. When a model has high precision but low recall, it is being too conservative about predicting the positive class. Lowering the decision threshold shifts the operating point along the precision-recall curve toward higher recall, which is exactly what the scenario asks for without retraining the model.

Answer

Apply L2 regularization and retrain the classifier.

Answer

Switch the evaluation metric from F1 score to accuracy.

Answer

Increase the number of training epochs and re-evaluate.

Question 14

A healthcare analytics team trains a model to flag patients at risk of readmission. The dataset contains 9,500 non-readmitted patients and 500 readmitted patients. The model predicts the majority class for every patient and reports 95 percent accuracy, yet it identifies no at-risk patients. Which evaluation approach best reveals the model's failure?

Accepted Answer

Compute the confusion matrix and examine recall and precision for the readmission class. With a 95-to-5 class split, a majority-class predictor achieves 95 percent accuracy while providing zero clinical value. The confusion matrix and minority-class recall and precision expose the absence of true positives, which is the evidence needed to justify resampling, class weighting, or threshold tuning.

Answer

Measure training time and inference latency across the full dataset

Answer

Report overall accuracy on a stratified holdout set

Answer

Calculate the mean squared error between predicted probabilities and labels

Question 15

A team is deploying a sentiment classifier and notices that the model outputs probabilities such as 0.83 for the positive class, but the actual positive rate among examples scored near 0.83 is only about 0.55. Stakeholders need the scores to reflect true likelihoods. Which action should the team take?

Accepted Answer

Apply a calibration method such as Platt scaling or isotonic regression to the model's output scores.. The scenario describes a reliability problem: scores near 0.83 correspond to an actual positive rate near 0.55, so the probabilities are overconfident. Calibration techniques such as Platt scaling and isotonic regression learn a mapping from raw scores to empirical frequencies using a held-out set, producing probabilities that better reflect true likelihoods without changing the model's ranking ability.

Answer

Replace the classifier with a k-nearest neighbors model, which produces inherently calibrated probabilities.

Answer

Raise the decision threshold from 0.5 to 0.83 so only highly confident examples are labeled positive.

Answer

Retrain the classifier with a lower learning rate and more epochs to improve probability estimates.