AI Models and Data Engineering practice questions

SMOTE (Synthetic Minority Oversampling Technique)

SMOTE generates synthetic minority-class samples by interpolating between existing class B neighbours, rebalancing the 9,500-to-500 skew so the classifier no longer favours class A. This directly addresses the stem's imbalance constraint, unlike random undersampling, which would discard valuable majority records.

Use TLS for transfer, AES-256 for storage, and AWS CloudTrail for auditing

TLS encrypts data in transit between services, AES-256 encrypts stored customer records at rest, and AWS CloudTrail logs every API access for audit trails. Together these satisfy all three stem constraints: encryption in transit, encryption at rest, and auditable access to sensitive training data.

The model is overfitting after epoch 2.

Training loss keeps falling while validation loss rises after epoch 2, meaning the network memorises training data instead of generalising. That divergence between the two curves is the defining signature of overfitting, so training should stop or regularisation added.

The policy does not include s3:ListBucket action, which is needed to list objects.

Missing `s3:ListBucket` causes access denied when users list bucket contents, since that action operates on the bucket resource while `s3:GetObject` covers objects. The stem's corporate-network condition is satisfied by the IP restriction; the failure stems from the absent list permission, not the network range.

Implement early stopping

Validation loss rising while training loss falls signals overfitting, so halting training at the point of best validation performance prevents further divergence. Early stopping directly addresses this by monitoring validation loss and restoring the optimal checkpoint, satisfying the need to curb overfitting without altering the model architecture.

Use SMOTE to generate synthetic fraudulent transactions

SMOTE generates synthetic minority-class fraud samples by interpolating between existing fraudulent transactions, rebalancing the 99.9% versus 0.1% class distribution. This gives the supervised model sufficient fraudulent examples to learn distinguishing patterns rather than predicting the majority class.

MobileNet

MobileNet uses depthwise separable convolutions, which cut multiply-accumulate operations and parameter count dramatically compared with standard convolutions. That efficiency satisfies the limited compute, memory and power budget of edge devices while retaining accuracy adequate for real-time inference.

Lowercasing

Lowercasing normalises case variation so tokens such as 'Good' and 'good' map to one form, reducing vocabulary sparsity. This satisfies the stem's multilingual constraint by applying consistently across scripts, though it alone cannot resolve language-specific morphology or tokenisation.

Tokenization

Tokenization splits multilingual text into discrete units (words, subwords or characters), which every downstream vectorisation step requires. For a dataset spanning languages with differing scripts and no whitespace boundaries, subword tokenisation keeps vocabulary manageable and prevents out-of-vocabulary failures, satisfying the cross-language performance constraint.

Removing stop words

Stop-word removal strips high-frequency function words so classification weighs meaningful terms. Because stop-word lists are language-specific, applying per-language lists to the multilingual corpus reduces noise consistently across every language, directly supporting the requirement that the model perform well across all languages.

Use semi-supervised learning with the unlabeled images to improve feature representations

Semi-supervised learning exploits the unlabeled retinal images to learn richer feature representations, directly addressing the 99%-versus-85% train-test gap that signals overfitting. Better generalisation reduces false negatives on unseen clinical images, satisfying the deployment constraint where missed disease is critical.

Replace missing Age with the mean and missing Income with the median.

Mean imputation suits Age's symmetric distribution, preserving its central tendency with only 10% missing. Income is typically right-skewed, so the median resists distortion from extreme earners, addressing the 30% missingness without pulling imputed values toward an inflated mean. This pairing minimises bias across both columns' differing distributions.

Target encoding with regularization.

Target encoding maps each high-cardinality category to a regularised statistic of the target, avoiding the dimensional explosion of one-hot encoding across 200+ neighbourhoods. Regularisation shrinks noisy category estimates, curbing overfitting while retaining a single interpretable numeric column.

Remove duplicate rows.

Removing duplicate rows directly eliminates redundant records, satisfying the requirement to clean a dataset containing duplicates. Deduplication operates on exact row matches, reducing skew and preventing inflated counts during downstream analysis. It is the targeted, minimal intervention for this specific defect, unlike broader transformations such as normalisation or imputation.

Recursive Feature Elimination (RFE) with cross-validation.

RFE with cross-validation suits this high-dimensional, small-sample scenario because it recursively prunes the least important features while cross-validation estimates generalisation on held-out folds, directly countering overfitting when features (10,000) vastly outnumber samples (500). This wrapper method captures feature interactions that filter methods miss.

Use a predictive model to estimate missing values.

Imputation via a predictive model estimates missing values from the relationships between other features, satisfying the requirement to handle gaps without discarding rows. Unlike mean substitution, which ignores correlations, this regression or k-NN approach preserves the dataset's statistical structure, making it appropriate when missingness is systematic rather than random.

Delete rows with missing values if the missing rate is low.

Deleting rows with missing values is appropriate when the missing rate is low, since the loss of information is minimal and bias stays negligible. This satisfies the stem's constraint by removing incomplete records without materially distorting the dataset's distribution, unlike deletion at high missing rates, which would discard substantial data.

Timeliness.

Timeliness ensures features reflect the current state at inference time. Stale data, such as yesterday's prices or lagging sensor readings, silently degrades predictions even when other dimensions pass, so freshness directly underpins the reliability the stem demands.

Consistency.

Consistency ensures identical inputs yield identical outputs across datasets, pipelines and time, preventing contradictory training signals that destabilise learned parameters. For model reliability, this dimension directly satisfies the stem's requirement by eliminating distributional drift and label conflicts between sources, so predictions remain reproducible and trustworthy rather than varying with data provenance.

Accuracy.

Accuracy confirms values correctly represent the real-world events they describe. Systematic measurement errors or mislabelled targets teach the model wrong relationships, so correctness of individual records is fundamental to the reliable predictions the stem asks about.

The event will be rejected because user_id is required.

The stream processor enforces user_id as a required field, so an event lacking it fails schema validation and is discarded rather than processed. This prevents downstream operations that depend on user_id from receiving malformed records.

0.8

Recall equals true positives divided by the sum of true positives and false negatives, so the confusion matrix values in the exhibit yield 0.8. This satisfies the question by correctly applying the recall formula to the displayed counts rather than computing precision or accuracy.

Impute missing values with the median of the column

The median is robust to outliers and skew, so imputing the column's median preserves the central tendency of the distribution better than the mean, minimising the bias introduced by discarding or distorting the 5% missing rows.

Balance the representation of professions across genders

Balancing profession representation across genders removes the statistical correlation between gender and occupation in the corpus, so the model cannot learn spurious associations. This directly addresses the stem's goal of preventing biased gender-profession links during training.

Downsampling to a fixed frequency using mean aggregation

Downsampling to a fixed frequency with mean aggregation converts irregular timestamps into evenly spaced intervals and smooths occasional spikes. Time-series models require uniform spacing, so this satisfies the stem's irregular-arrival constraint while damping outlier noise.