AI0-001 AI Models and Data Engineering Practice Test 3 — 15 Questions

Question 1

A data engineer is preprocessing text data for sentiment analysis. Which technique preserves word order while converting text to numeric features?

Accepted Answer

N-grams with frequency counts. (N-grams with frequency counts) is correct because n-grams capture sequences of words, preserving local order. TF-IDF and Bag-of-words ignore word order entirely, treating text as an unordered set of tokens. Word2Vec produces dense embeddings that encode semantic similarity based on context, but it does not explicitly preserve exact word order; it learns from surrounding words but focuses on meaning rather than sequence.

Answer

TF-IDF

Answer

Word2Vec

Answer

Bag-of-words

Question 2

A team is training a deep neural network on a large image dataset. They observe that the training loss decreases smoothly but validation loss oscillates. Which regularization technique should be applied?

Accepted Answer

Dropout. Dropout is the correct regularization technique because it randomly drops neurons during training, which prevents co-adaptation of features and reduces overfitting. This addresses the validation loss oscillation (a sign of overfitting) while allowing the training loss to decrease smoothly, as dropout only applies during training and not during validation.

Answer

Data augmentation

Answer

L1 regularization

Answer

Batch normalization

Question 3

Which TWO strategies are effective for handling missing values in a dataset when the missingness is not random (MNAR)?

Accepted Answer

Multiple imputation using chained equations. Option A (Multiple imputation using chained equations) is correct because MICE models each variable with missingness conditional on the others and produces multiple plausible imputations that propagate uncertainty, which is appropriate when missingness depends on unobserved values (MNAR) and can incorporate auxiliary variables or sensitivity parameters to reflect the non-random mechanism. Option B (Treat missing as a separate category) is correct because for categorical features, encoding 'missing' as its own level preserves the informative nature of the missingness itself, which is exactly what MNAR implies—the fact that a value is absent carries signal that would be lost by imputing a plausible value. Option C (Listwise deletion) is not appropriate because it discards rows with missing data, biasing estimates when missingness is non-random and reducing sample size. Option D (KNN imputation) is not appropriate because it imputes from observed neighbors under a missing-at-random assumption, which is violated in MNAR and can distort the missingness signal. Option E (Mean imputation) is not appropriate because it shrinks variance, ignores the MNAR mechanism, and biases relationships among variables.

Answer

Listwise deletion

Answer

KNN imputation

Answer

Mean imputation

Question 4

A data scientist is building a regression model to predict house prices. The dataset contains features such as square footage, number of bedrooms, and year built. Initial model performance is poor, and the scientist suspects that feature engineering could help. Which approach is most likely to improve model accuracy?

Accepted Answer

Create interaction terms such as bedrooms times square footage. Creating interaction terms like bedrooms × square footage captures non-linear relationships and synergies between features that a linear model alone cannot represent. In real estate, the effect of square footage on price often depends on the number of bedrooms (e.g., a large house with few bedrooms may be less valuable), so interaction terms allow the model to learn these conditional patterns, directly improving predictive accuracy.

Answer

Use only linear features because polynomial terms overfit

Answer

Remove all features except square footage to reduce noise

Answer

Add random noise to the target variable to increase variance

Question 5

A credit risk model is being developed to predict loan defaults. The dataset has 95% non-default and 5% default instances. The data scientist trains a logistic regression model and obtains 95% accuracy, but the recall for defaults is only 10%. Which action is most appropriate to improve the model's ability to identify defaults?

Accepted Answer

Use SMOTE to oversample the default class. SMOTE (Synthetic Minority Oversampling Technique) is the most appropriate action because it generates synthetic samples for the minority class (defaults) rather than simply duplicating existing ones. This directly addresses the severe class imbalance (95% non-default vs. 5% default) that causes the logistic regression model to achieve high accuracy by predicting nearly all instances as non-default, while failing to identify actual defaults (recall of only 10%). By creating realistic synthetic default instances, SMOTE balances the training data and forces the model to learn decision boundaries that better capture the minority class.

Answer

Apply principal component analysis (PCA) to reduce dimensionality

Answer

Collect more data from loan applicants to increase dataset size

Answer

Undersample the non-default class to match the number of defaults

Question 6

A dataset used for training a classification model contains 10% missing values in a feature that is known to be important. The data scientist decides to impute the missing values. Which imputation method is most robust if the data is not missing completely at random?

Accepted Answer

Use multiple imputation to model missing values. Multiple imputation is the most robust method when data is not missing completely at random (NMAR) because it uses a statistical model to account for the relationships between the missing feature and other observed features, generating multiple plausible values and combining them to produce unbiased estimates and valid standard errors. This approach preserves the variability and structure of the data, unlike simpler methods that can introduce bias when missingness depends on unobserved data.

Answer

Delete all rows with missing values

Answer

Replace missing values with the mean of the feature

Answer

Fill missing values with 0

Question 7

A company wants to forecast monthly sales for the next year using historical sales data over three years. The data shows strong seasonality and a slight upward trend. Which model type is best suited for this task?

Accepted Answer

SARIMA model with seasonal order (1,1,1)[12]. The SARIMA (Seasonal ARIMA) model with seasonal order (1,1,1)[12] is best suited because it explicitly captures both the strong seasonality (period 12 for monthly data) and the slight upward trend through its non-seasonal differencing and seasonal components. This model extends ARIMA by adding seasonal terms, making it ideal for time series with clear seasonal patterns and trends.

Answer

Simple moving average of the last 12 months

Answer

ARIMA model without seasonal terms

Answer

Linear regression with time as the independent variable

Question 8

A data engineer is splitting a dataset into training, validation, and test sets for a machine learning project. The dataset is large and representative of the population. Which split ratio is commonly recommended?

Accepted Answer

80% training, 10% validation, 10% test. (80% training, 10% validation, 10% test) is commonly recommended for large, representative datasets because it provides sufficient data for model training while retaining enough samples in the validation and test sets to reliably evaluate model performance and detect overfitting. This split balances the need for a robust training set with the requirement for statistically meaningful holdout sets, as recommended in standard machine learning practices for AI model development.

Answer

90% training, 5% validation, 5% test

Answer

70% training, 20% validation, 10% test

Answer

50% training, 25% validation, 25% test

Question 9

A natural language processing (NLP) team is building a sentiment analysis model. The raw text data contains punctuation, stop words, and URLs. Which TWO preprocessing steps are most appropriate to improve model performance? (Choose two.)

Accepted Answer

Remove all punctuation and URLs. Option A is correct because punctuation and URLs are noise in sentiment analysis: URLs carry no sentiment signal and punctuation marks (aside from possibly emoticons) add tokens that increase vocabulary size and sparsity without contributing to polarity, so stripping them cleans the feature space. Option C is correct because stop words such as "the", "is", and "and" occur with high frequency but carry little sentiment information, and removing them reduces dimensionality and lets the model focus on content-bearing terms. Option B is not appropriate here because stemming can distort sentiment-bearing words (e.g., conflating distinct forms) and is not one of the two most fitting steps for this scenario. Option D, lowercasing, is a normalization step that is generally useful but is not marked correct for this question. Option E, tokenization, is a required foundational step rather than a noise-reduction preprocessing improvement, so it is not among the two selected answers.

Answer

Apply stemming to reduce words to root forms

Answer

Convert all text to lowercase

Answer

Tokenize the text into individual words

Question 10

A data scientist is evaluating a binary classification model for fraud detection. The dataset is highly imbalanced (99% non-fraud, 1% fraud). Which TWO metrics are most appropriate for assessing model performance? (Choose two.)

Accepted Answer

Precision. Precision (A) is appropriate because in a 99:1 imbalanced fraud dataset it measures the fraction of transactions flagged as fraud that are truly fraudulent, directly quantifying the cost of false positives (blocking legitimate customers), which accuracy would obscure. Recall (B) is equally appropriate because it measures the fraction of actual fraud cases the model catches, addressing the severe cost of false negatives (missed fraud) that dominates in imbalanced settings. Together, precision and recall expose the trade-off that matters for fraud detection, whereas the other metrics hide it. F1 score (C) is a single blended measure and is not one of the two marked answers here, AUC-ROC (D) can look deceptively high under extreme class imbalance because the large non-fraud class inflates the true-negative rate, and accuracy (E) is misleading since a trivial all-non-fraud classifier already achieves 99% accuracy.

Answer

F1 score

Answer

Area under the ROC curve (AUC-ROC)

Answer

Accuracy

Question 11

A computer vision team is building an image classifier for rare wildlife species. The dataset has only 500 images per class, and the model overfits. Which THREE data augmentation techniques are most likely to reduce overfitting? (Choose three.)

Accepted Answer

Horizontal flip. Horizontal flip (A) is correct because mirroring images creates label-preserving variations that effectively enlarge the 500-image-per-class dataset and force the model to learn orientation-invariant features, directly combating overfitting. Color jitter (D) is correct because varying brightness, contrast, and saturation simulates realistic lighting and camera differences, adding diversity without changing the class label, which regularizes the classifier. Random rotation by ±10 degrees (E) is correct because small rotations produce plausible new viewpoints of wildlife while keeping the object recognizable, increasing effective sample size and reducing memorization. Adding Gaussian noise (B) is not among the marked answers because it perturbs pixel values without creating semantically meaningful new samples and can degrade learning rather than reliably reduce overfitting. Random cropping (C) is also unmarked because aggressive cropping can cut off the animal and destroy the discriminative features needed for classification, making it less suitable here than the three selected techniques.

Answer

Adding Gaussian noise

Answer

Random cropping

Question 12

A healthcare startup is deploying a machine learning model to predict patient readmission within 30 days using electronic health records (EHR). The data pipeline uses Apache Spark for preprocessing and training on an Amazon EMR cluster. The training dataset is 50 GB and composed of structured numeric and categorical features, along with unstructured clinical notes. The data scientist observes that training takes over 12 hours and frequently fails due to out-of-memory (OOM) errors, especially when processing the clinical notes via TF-IDF vectorization. The cluster has 10 nodes with 64 GB RAM each. The data engineer has already tried increasing spark.sql.shuffle.partitions to 400 and using Kryo serialization, but OOM persists. Which action should the data engineer take next to resolve the OOM errors?

Accepted Answer

Repartition the clinical notes data into 2000 partitions before TF-IDF. Repartitioning the clinical notes data into 2000 partitions before TF-IDF vectorization increases parallelism and reduces the memory pressure per partition. The default partition count (often based on spark.default.parallelism) is too low for 50 GB of data, causing individual partitions to exceed executor memory limits. By increasing partitions, each executor processes smaller chunks, preventing OOM errors during the memory-intensive TF-IDF stage.

Answer

Broadcast the TF-IDF model to all executors to avoid shuffling

Answer

Add 10 more nodes to the cluster to increase total memory

Answer

Use a single executor with 64 GB and increase driver memory to 128 GB

Question 13

A financial services company has a real-time fraud detection system that uses Apache Kafka to stream transaction events, a TensorFlow Serving model for scoring, and a Redis cache for lookup of historical fraud patterns. The system processes 10,000 transactions per second with an SLA of 100ms latency per transaction. Recently, after a model update, the latency for some transactions spiked to over 500ms, causing timeouts. The model uses a deep neural network with 10 million parameters. The engineering team suspects the issue is due to increased model inference time. Which action should be taken to reduce latency without significant loss in accuracy?

Accepted Answer

Quantize the model weights from FP32 to FP16. The latency spike is caused by increased model inference time after a model update. Quantizing model weights from FP32 to FP16 reduces memory bandwidth and computation requirements, directly speeding up inference on compatible hardware (e.g., GPUs with Tensor Cores) with minimal accuracy loss. This addresses the root cause—model inference latency—without changing the system architecture.

Answer

Add more Redis nodes to the cache cluster

Answer

Increase the number of Kafka partitions and consumer threads

Answer

Decrease the inference batch size from 32 to 1

Question 14

A logistics company uses a machine learning model to predict delivery times based on historical data including distance, traffic, weather, and driver performance. The model is deployed as a REST API using Flask and run on a single server. Recently, the model has been returning predictions with high latency (over 2 seconds) during peak hours when the API receives 500 requests per second. The server has 8 CPU cores and 32 GB RAM. The model is a gradient boosting model (XGBoost) with 500 trees. The engineer wants to reduce inference latency to under 500ms without retraining the model. Which action is most effective?

Accepted Answer

Prune the model by reducing the number of trees to 100 and limit tree depth. Inference latency for an XGBoost model scales roughly with the number of trees and their depth, since each prediction traverses every tree. Reducing 500 trees to 100 and limiting depth directly cuts the computation per prediction, which is the most effective way to reduce latency without retraining from scratch (pruning is a form of model simplification). This addresses the root cause: model complexity driving CPU-bound inference time.

Answer

Replace XGBoost with a linear regression model

Answer

Scale horizontally by deploying additional servers behind a load balancer

Answer

Increase server RAM to 128 GB

Question 15

A data scientist is preparing a dataset for a classification model. The dataset contains several categorical variables with high cardinality. Which TWO encoding methods are appropriate for converting these categorical variables into numerical features?

Accepted Answer

One-hot encoding. One-hot encoding (C) is appropriate because it converts each category into a separate binary column, which handles high-cardinality categorical variables without imposing an artificial ordinal relationship, though it can create many sparse features. Label encoding (E) is also appropriate because it assigns each unique category an integer value, efficiently transforming categorical variables into numerical features that tree-based models can consume. Min-max scaling (A) is a numerical normalization technique and does not encode categorical data. K-means clustering (B) is an unsupervised clustering algorithm, not an encoding method. Principal component analysis (D) is a dimensionality-reduction technique for numerical features, not a categorical encoder.

Answer

Min-max scaling

Answer

K-means clustering

Answer

Principal component analysis (PCA)