WorksheetsBA 3
Total questions: 127
Worksheet time: 1hrs 4mins
Name
Class
Date
1.
In the running case, what is the prediction goal for the Telco churn task?
a)
Predict churn in the next 30 days
b)
Predict monthly revenue for the next year
c)
Predict customer lifetime value over 10 years
d)
Predict plan upgrades within 24 hours
e)
Min-max scaling of numeric features
2.
Which set of terms is presented as core data quality dimensions?
a)
Accuracy, completeness, consistency
b)
Improve model accuracy without validating data types
c)
Treat timeliness as irrelevant once a model is trained
d)
Bias, variance, regularization
e)
Assume schema consistency implies value validity
3.
In the data quality scorecard example, which KPI threshold is explicitly shown?
a)
Missing rate below 2% per feature
b)
Model accuracy above 99%
c)
Always drop any row with a missing value
d)
Training loss below 0.01
e)
Customer churn rate below 1%
4.
Which sequence matches the stated data quality lifecycle?
a)
Ingest, profile, clean, validate, monitor, feedback
b)
Collect, label, train, test, deploy, monetize
c)
Improve model accuracy without validating data types
d)
Plan, build, run, debug, refactor, release
e)
Split, scale, encode, select, validate, predict
5.
Which activity is listed under quick data profiling?
a)
Checking distributions and min/max
b)
Computing Shapley values for all features
c)
Training a generative model to synthesize data
d)
Min-max scaling of numeric features
e)
Designing a data warehouse star schema
6.
Which approach to deduplication is explicitly mentioned?
a)
Exact keys vs. fuzzy matching using edit distance
b)
Only hashing-based similarity joins
c)
Only manual review by domain experts
d)
Only clustering with k-means
e)
Only removing the newest records
7.
What is the purpose of normalizing units and value domains in data quality checks?
a)
Ensure currencies, percentages, and units are consistent
b)
Increase the number of features automatically
c)
Improve model accuracy without validating data types
d)
Guarantee perfect model accuracy
e)
Replace all missing values with zeros
8.
Which missingness mechanism is defined as 'completely at random'?
a)
MCAR
b)
MAR
c)
MNAR
d)
Temporal missingness
e)
Structural missingness
9.
Which method is listed for diagnosing missingness?
a)
Compare distributions in missing vs. non-missing groups
b)
Always assume missing values are MCAR
c)
Remove all rows with any missing value
d)
Impute all columns with zero without analysis
e)
Only inspect a single random sample
10.
In the example by segment, which variable is said to be missing more in the Prepaid segment?
a)
Income
b)
Tenure
c)
Churn label
d)
Plan identifier
e)
Always drop any row with a missing value
11.
Which strategy is explicitly recommended as part of missingness handling?
a)
Add an is_missing indicator
b)
Always drop the target column
c)
Always use global mean imputation
d)
Always use one-hot encoding for numeric features
e)
Always remove outliers before splitting data
12.
Why are median and mode imputation described as robust?
a)
They are robust to outliers
b)
They require no computation
c)
They always preserve correlations perfectly
d)
They guarantee unbiased estimates under MNAR
e)
They remove the need for train/test splits
13.
Which statement about kNN imputation is explicitly listed?
a)
It can be compute-heavy and scale-sensitive
b)
It requires no feature scaling ever
c)
It works only for categorical variables
d)
It always eliminates leakage by itself
e)
It cannot handle missing targets
14.
What is MICE (Multiple Imputation by Chained Equations) described as doing?
a)
Model each incomplete feature iteratively until convergence
b)
Replace all missing values with the global mean once
c)
Remove any column with more than 1% missingness
d)
Encode categories using hashing into a fixed space
e)
Detect outliers using Isolation Forest
15.
Which approach is listed as a time-aware imputation method for time series?
a)
Forward fill or backward fill
b)
One-hot encoding
c)
Chi-square feature selection
d)
Permutation importance
e)
Ridge regression
16.
How do suggest assessing the impact of an imputation strategy?
a)
Compare model metrics like AUC or RMSE across strategies
b)
Only inspect a single record after imputation
c)
Only check the number of columns in the dataset
d)
Always prefer the strategy with the most complex model
e)
Always impute with the mean to keep things consistent
17.
Which item is listed as a common pitfall in missingness handling?
a)
Imputing before the train/test split
b)
Using median instead of mean
c)
Using a pipeline for preprocessing
d)
Checking distribution drift after imputation
e)
Adding an is_missing flag
18.
What is the stated definition of data leakage?
a)
Information from future or test data creeps into training
b)
Any model that performs well on training data
c)
A dataset with duplicate rows
d)
A model with too many parameters
e)
A feature with high cardinality
19.
Which leakage type is defined as preprocessing fitted on the full data?
a)
Train-test contamination
b)
Target leakage
c)
Temporal leakage
d)
Label noise
e)
Concept drift
20.
In the churn example, why is the feature 'refund_in_30d' considered problematic?
a)
It occurs after churn and leaks target information
b)
It is a low-cardinality categorical variable
c)
It is perfectly independent of churn
d)
It is always missing for all customers
e)
It uses units that cannot be normalized
21.
Which method is listed for avoiding temporal leakage?
a)
Use time-based splits such as TimeSeriesSplit or rolling origin
b)
Randomly shuffle all data before splitting
c)
Impute using the full dataset
d)
Apply one-hot encoding to numeric variables
e)
Remove rare categories after training
22.
Which principle is stated under anti-leakage pipeline practices?
a)
Fit preprocessing only on training folds
b)
Fit preprocessing on the entire dataset for stability
c)
Compute target encoding using the full dataset without CV
d)
Apply outlier removal after model evaluation
e)
Tune hyperparameters without cross-validation
23.
Which data splitting strategy is recommended when there are multiple rows per customer?
a)
Group-aware split
b)
Purely random shuffle split
c)
Split by feature importance
d)
Split by model prediction score
e)
Split by sorting categories alphabetically
24.
Which encoding method is included in the categorical encoding overview list?
a)
Hashing
b)
Min-max scaling of numeric features
c)
Batch normalization
d)
Median imputation of missing values
e)
IQR-based capping for outliers
25.
What is a stated downside of one-hot encoding with high-cardinality features?
a)
Dimensionality explosion
b)
It forces an incorrect order
c)
It requires binning numeric variables
d)
It cannot handle unknown categories
e)
It always causes label leakage
26.
When is ordinal encoding appropriate according to?
a)
When categories have a true ordering
b)
When categories are unordered identifiers
c)
When features are continuous numeric values
d)
When the dataset is time series
e)
When the model is a decision tree
27.
How do recommend preventing leakage with target/mean encoding?
a)
Compute it inside cross-validation loops
b)
Compute it once using the full dataset
c)
Compute it only on the test set
d)
Replace categories with random integers
e)
Always one-hot encode before target encoding
28.
Weight of Evidence (WoE) is described as widely used in which context?
a)
Credit risk
b)
Customer churn prediction
c)
Demand forecasting
d)
Time series decomposition
e)
Graph clustering
29.
What is a stated drawback of the hashing trick for categorical encoding?
a)
Hash collisions and less interpretability
b)
It cannot handle high cardinality
c)
It requires true ordering of categories
d)
It produces an unbounded number of features
e)
It always leaks the target
30.
Which tactic is listed for handling high-cardinality categorical features?
a)
Frequency or count encoding
b)
Always use ordinal encoding
c)
Always create one binary feature per category
d)
Drop the entire dataset
e)
Replace categories with the label value
31.
Which text feature approach is listed as a quick method?
a)
Bag-of-words or TF-IDF
b)
Min-max scaling of numeric features
c)
Fourier transform of images
d)
Kalman smoothing for state space
e)
One-hot encoding of pixels
32.
What is the basic idea of entity embeddings for categorical variables?
a)
Learn dense representations with neural networks
b)
Always map each category to a unique one-hot vector
c)
Replace categories with their alphabetic rank
d)
Remove categorical variables entirely
e)
Use only frequency counts and never a model
33.
Which algorithm family is listed as scale-sensitive, motivating feature scaling?
a)
kNN
b)
Random forests
c)
One-hot encoding of categorical variables
d)
Rule-based systems
e)
Decision stumps only
34.
What does StandardScaler do to a numeric feature according to?
a)
Subtract the mean and divide by the standard deviation
b)
One-hot encoding of categorical variables
c)
Replace values with the median and IQR
d)
Map values to a uniform distribution
e)
Convert values into categories based on frequency
35.
Which transform is explicitly noted as requiring inputs greater than zero?
a)
Box-Cox
b)
Yeo-Johnson
c)
Standard scaling
d)
Frequency encoding
e)
One-hot encoding
36.
What is the purpose of a quantile or rank-Gauss transform as described?
a)
Map data to uniform or normal distributions
b)
Detect duplicates using edit distance
c)
Guarantee interpretability of hashed features
d)
Prevent temporal leakage automatically
e)
Convert text into n-grams
37.
Which method is listed for outlier detection?
a)
IQR-based thresholds
b)
One-hot encoding
c)
Chi-square filtering
d)
Rolling origin evaluation
e)
Target mean encoding
38.
Which technique is listed for handling outliers?
a)
Capping or winsorization
b)
Always delete any record with high values
c)
Replace outliers with the target label
d)
Use one-hot encoding to remove outliers
e)
Ignore scaling and imputation
39.
Which example ratio feature is explicitly provided?
a)
ARPU equals revenue divided by usage
b)
AUC equals accuracy divided by recall
c)
IQR equals mean divided by variance
d)
TF-IDF equals frequency divided by entropy
e)
WoE equals variance divided by mean
40.
Which is listed as a type of date-time feature?
a)
Holiday flags
b)
Hash collisions
c)
Gini impurity
d)
Edit distance
e)
Ridge coefficients
41.
Which is listed as an example geo feature method when coordinates are available?
a)
Haversine distance
b)
ANOVA F-test
c)
TimeSeriesSplit
d)
Bag-of-words
e)
Lasso regularization
42.
Which is an example of an aggregation or group-by feature?
a)
Counts per user or plan
b)
StandardScaler on all columns
c)
One-hot encoding all categories
d)
Box-Cox transform for all values
e)
t-SNE projection for visualization
43.
Which is listed as a time-series feature type?
a)
Lag features such as t-1 or t-7
b)
Chi-square p-values
c)
Geohash grid cells
d)
Word n-grams
e)
Permutation feature importance
44.
Which item is listed under extended text or NLP features?
a)
Sentiment scores and complaint keyword flags
b)
Haversine distances and clustering
c)
Quantile mapping to uniform
d)
WoE monotone binning
e)
Time-based cross-validation
45.
Which set correctly matches the feature selection categories listed?
a)
Filter, wrapper, embedded
b)
Ingest, profile, clean
c)
Encode, scale, deploy
d)
Train, test, serve
e)
Collect, label, infer
46.
For feature selection, what is Chi-square used for according to?
a)
Categorical features versus the target
b)
Numeric features versus time
c)
Text features versus stopwords
d)
One-hot encoding to remove irrelevant features
e)
Scaling high-cardinality categories
47.
In wrapper methods, how does RFE typically choose the number of features to keep?
a)
Choose k via cross-validation
b)
Choose k by maximizing duplicate rate
c)
Choose k by minimizing missingness only
d)
Choose k by sorting features alphabetically
e)
Choose k equal to the number of categories
48.
Which statement about Lasso is explicitly listed?
a)
Lasso uses L1 regularization for sparsity
b)
Lasso uses only decision trees
c)
Lasso requires target encoding
d)
One-hot encoding to remove irrelevant features
e)
Lasso is a time-series splitter
49.
What is the stated role of t-SNE or UMAP?
a)
Visualization; avoid direct forecasting use
b)
Primary method for forecasting time series
c)
Replacement for train-test splitting
d)
Guaranteed leakage prevention tool
e)
Default method for imputing missing values
50.
Which is stated as a golden rule for pipelines?
a)
Never fit preprocessing on the full dataset
b)
Always fit preprocessing on the full dataset for stability
c)
Always compute features after model training
d)
Always remove rare categories without tracking
e)
Never use ColumnTransformer
51.
Which ordering best reflects a leakage-safe preprocessing and modeling pipeline?
a)
Raw, split, impute, encode, scale, feature select, model, evaluate
b)
Raw, model, encode, scale, split, evaluate, impute, select
c)
Split, raw, evaluate, model, encode, impute, scale, select
d)
Encode, scale, select, impute, split, raw, model, evaluate
e)
Model, evaluate, raw, split, encode, scale, select, impute
52.
In the leakage-safe pseudocode, which scikit-learn utility is used to apply different preprocessing to different column groups?
a)
ColumnTransformer
b)
Fit transforms using all available data to reduce variance
c)
Encode categories after training to avoid drift
d)
Evaluate on the training set to save computation
e)
Tune thresholds using test labels for stability
53.
Which metric is explicitly listed for evaluating models after preprocessing?
a)
Brier score
b)
BLEU score
c)
Min-max scaling of numeric features
d)
Mean Average Precision for ranking
e)
Silhouette score
54.
Which item is listed as part of pre-deploy data validation checks?
a)
Null percentage, ranges, cardinality, and drift
b)
Min-max scaling of numeric features
c)
Number of neurons in the model
d)
Standardization using z-scores
e)
User interface theme colors
55.
What are data contracts described as?
a)
Agreements between producers and consumers on fields, definitions, and SLAs
b)
A type of loss function for classification
c)
A method for hashing categorical variables
d)
A plotting technique for missingness
e)
A feature selection algorithm
56.
Which practice is listed under reproducibility and experiment tracking?
a)
Track parameters, metrics, artifacts, and data snapshots
b)
Always retrain without fixing seeds
c)
Avoid saving environment files
d)
Use only manual notes and no logs
e)
Delete training data after each run
57.
Which item is listed as a common anti-pattern?
a)
Ignoring rare categories and using unbounded one-hot
b)
Using ColumnTransformer for per-type pipelines
c)
Fitting preprocessing only on training folds
d)
Adding an is_missing flag
e)
Using time-based splits for temporal data
58.
In the mini-case (Part 1), what is the stated deliverable?
a)
A data quality metrics table for the top 10 columns
b)
Improve model accuracy without validating data types
c)
A full database migration plan
d)
A set of labeled images for training
e)
A simulation of customer behavior
59.
In the mini-case (Part 2), how is plan_id recommended to be encoded due to high cardinality?
a)
CV-safe target encoding
b)
Ordinal encoding without ordering
c)
StandardScaler
d)
Box-Cox transform
e)
IQR-based capping
60.
In the mini-case (Part 3), which comparison is explicitly suggested?
a)
Logistic regression with scaling vs. XGBoost without scaling
b)
kNN without scaling vs. SVM without scaling
c)
PCA vs. t-SNE for forecasting
d)
WoE vs. hashing for images
e)
ANOVA vs. Chi-square for text
61.
In the in-class hands-on steps, which activity is explicitly required?
a)
Run 5-fold cross-validation and log results
b)
Train on the test set to maximize accuracy
c)
Skip preprocessing to save time
d)
Only compute ROC-AUC without any comparison
e)
Remove all missing values by deleting rows
62.
Which statement is included in the key takeaways summary?
a)
Leakage-safe pipelines are mandatory
b)
Scaling is never needed for any algorithm
c)
Data quality is irrelevant if the model is complex
d)
One-hot encoding always outperforms hashing
e)
Outliers should always be removed automatically
63.
What method is listed for encoding cyclic time features such as hour-of-day?
a)
Use sine and cosine components
b)
Use edit distance on timestamps
c)
Use Chi-square tests on timestamps
d)
Use one-hot encoding on numeric values only
e)
Use kNN imputation on timestamps
64.
In the appendix example, how is group-wise imputation for income implemented in pandas?
a)
Fill missing income with the median within each plan_type group
b)
Replace all income values with the global mean
c)
Drop all rows where income is missing
d)
Encode income using hashing
e)
Scale income to the range [0, 1]
65.
In the appendix, what is the recommended way to use target encoding to avoid leakage?
a)
Use TargetEncoder inside a pipeline and within cross-validation
b)
Compute target means on the full dataset once and reuse them
c)
Fit the encoder on the test set only
d)
Encode categories with random noise only
e)
Always one-hot encode before target encoding and skip CV
66.
In the visual example for missingness, what does the value 1 represent?
a)
A missing value
b)
A duplicated row
c)
An outlier beyond 3 standard deviations
d)
Always drop any row with a missing value
e)
A target class label
67.
Which visualization best summarizes a numeric distribution and highlights potential outliers using quartiles?
a)
Boxplot
b)
Increase the influence of extreme values to improve fit
c)
Replace all values outside the mean with missing values
d)
Sankey diagram
e)
Heatmap of correlations
68.
A data team wants a practical data quality scorecard. Which option best matches the example approach?
a)
Track missing rate per feature and duplicate rate, and define a data SLA for delivery time and schema stability
b)
Track only model accuracy and stop when accuracy exceeds 95%
c)
Silent schema changes without versioning
d)
Track only the number of tables and avoid any thresholds
e)
Track only the number of unique customer IDs and ignore missingness
69.
Which pairing reflects the recommendation for how to validate data quality through the lifecycle?
a)
Use rule-based checks together with descriptive statistics
b)
Use only manual inspection and avoid automated checks
c)
Use only descriptive statistics and avoid any rules
d)
Improve model accuracy without validating data types
e)
Use only unit tests and avoid monitoring in production
70.
Income is missing more often for customers on a particular plan type, while plan type is fully observed. Which missingness mechanism is most consistent with this description?
a)
MCAR, because missingness is completely random
b)
MAR, because missingness depends on an observed variable (plan type)
c)
MNAR, because missingness depends only on the missing income value itself
d)
Temporal leakage, because the plan type occurs after the label
e)
Train-test contamination, because plan type is categorical
71.
You suspect missing income is related to churn. Which analysis is most aligned with for diagnosing this?
a)
Compute missing maps and test the relationship between missingness indicators and the label
b)
Assume income is MCAR and impute with the global mean
c)
Drop the income column immediately to avoid any bias
d)
Fit a model on the full dataset and interpret training accuracy only
e)
Randomly shuffle labels to remove any association
72.
A feature has very high missingness and provides little business value. Which action best matches the strategy overview?
a)
Drop the column to reduce noise and complexity
b)
Force one-hot encoding to make the missingness disappear
c)
Replace missing values with a constant and never record missingness
d)
Scale the feature to [0,1] to address missingness
e)
Use t-SNE to visualize the missingness and keep the feature unchanged
73.
Why might group-wise median imputation by plan_type be preferred over global mean imputation in the telco case?
a)
It can respect segment differences when missingness varies by plan type
b)
It guarantees MCAR assumptions hold
c)
It always increases dimensionality to improve model power
d)
It removes the need for an is_missing indicator
e)
It eliminates the need for a train/test split
74.
A team uses kNN imputation and notices unstable results across features on different scales. What consistent fix addresses the root cause?
a)
Apply feature scaling and do it inside a leakage-safe pipeline
b)
Switch to one-hot encoding for all numeric features
c)
Compute global scaling statistics on the full dataset to stabilize kNN
d)
Drop all numeric variables and keep only categories
e)
Increase k until all neighbors become identical
75.
MICE is described as producing multiple completed datasets. What is the main reason for doing this, according to?
a)
To pool estimates and reflect uncertainty from imputation
b)
To guarantee the data become MCAR
c)
To avoid any need for cross-validation
d)
To increase the number of categories for one-hot encoding
e)
Always drop any row with a missing value
76.
After imputing missing values, what check helps ensure you did not destroy predictive signal?
a)
Check for distribution drift after imputation and compare model metrics across strategies
b)
Only verify that there are no missing values left
c)
Only check that the dataset size did not change
d)
Only re-encode categories after imputation
e)
Only compute correlation heatmaps and skip metrics
77.
Which practice is most likely to cause leakage when handling missingness, based?
a)
Imputing values before performing the train/test split
b)
Using median instead of mean imputation
c)
Adding an is_missing indicator column
d)
Comparing AUC across imputation strategies
e)
Using group-wise medians by plan type
78.
A model shows excellent training and validation scores, but performance drops sharply after deployment. Which defined issue is most consistent with this pattern?
a)
Data leakage from future or test information in training
b)
Dimensionality explosion from one-hot encoding always improves performance
c)
Box-Cox transform always hurts production accuracy
d)
kNN imputation is always compute-heavy and therefore unstable
e)
Feature selection always reduces performance in production
79.
A preprocessing step (imputation and scaling) is fitted once using all available data before cross-validation. Which leakage type does the taxonomy label this as?
a)
Train-test contamination
b)
Target leakage
c)
Temporal leakage
d)
MCAR
e)
WoE leakage
80.
You are forecasting churn over time and must avoid temporal leakage. Which split strategy best aligns with?
a)
Time-based split such as TimeSeriesSplit or rolling origin
b)
Random KFold after shuffling all rows
c)
Split by the largest numeric feature
d)
Split by encoding method (one-hot vs hashing)
e)
Fit transforms using all available data to reduce variance
81.
To evaluate preprocessing choices fairly, what does the anti-leakage guidance recommend?
a)
Run cross-validation on the full pipeline, not on preprocessed data created from all rows
b)
Precompute preprocessing using the full dataset to reduce variance
c)
Fit encoders on the test set to reduce distribution shift
d)
Skip cross-validation and rely on training AUC
e)
Tune the model first and handle missingness afterward
82.
A dataset has multiple records per customer. Why is a group-aware split recommended?
a)
To prevent information from the same customer appearing in both train and test folds
b)
To increase the number of folds without changing the data
c)
To force the label distribution to be uniform
d)
To eliminate the need for feature engineering
e)
To ensure one-hot encoding produces fewer columns
83.
Plan IDs have very high cardinality. If you still want a simple baseline, which consistent mitigation makes one-hot encoding more practical?
a)
Group rare levels into 'Other' or consider hashing instead
b)
Assign an arbitrary numeric order to plan IDs and use ordinal encoding
c)
Compute target means on the full dataset and encode directly
d)
Remove the plan ID feature and replace it with random noise
e)
Increase the number of categories by splitting each ID into characters
84.
Which use of ordinal encoding is discouraged by?
a)
Encoding an unordered identifier by mapping categories to arbitrary integers
b)
Encoding shirt sizes with an inherent order such as S, M, L, XL
c)
Using ordinal encoding only when true ordering exists
d)
Avoiding ordinal encoding for unordered categories
e)
Treating order as meaningful only when the domain defines it
85.
Target/mean encoding can overfit or leak if done incorrectly. Which practice best reduces this risk per?
a)
Compute target encoding within CV folds and use smoothing or noise
b)
Compute target means on the full dataset and reuse them everywhere
c)
One-hot encode the target variable and join it back as features
d)
Replace categories with their row index to prevent leakage
e)
Fit the encoder on the test set to match deployment distribution
86.
In credit-risk style problems, why might WoE be preferred over a naive encoding, according to?
a)
It is interpretable and supports monotonic relationships after binning
b)
It avoids any need for binning or sample size considerations
c)
It guarantees no collisions like hashing
d)
It is designed specifically for text n-grams
e)
It always yields fewer features than frequency encoding
87.
A team chooses the hashing trick for a streaming setting with many categories. What drawback should they anticipate, based?
a)
Hash collisions that reduce interpretability
b)
Min-max scaling of numeric features
c)
A required true ordering of categories
d)
Guaranteed target leakage in every fold
e)
Inability to handle high cardinality
88.
Which combination is listed as a plausible strategy for high-cardinality categorical features?
a)
Collapse rare levels and use frequency/count encoding or CV-safe target encoding
b)
Always use one-hot encoding without any constraints
c)
Always use Box-Cox to reduce the number of categories
d)
Always treat categories as numeric and apply StandardScaler
e)
Always remove categorical variables and keep only text features
89.
Which set of algorithms is explicitly described as scale-sensitive, making scaling important?
a)
kNN, k-means, SVM, PCA, and L1/L2 regression
b)
Decision trees and rule-based systems only
c)
One-hot encoding of categorical variables
d)
Only naive Bayes and topic models
e)
Only hashing and WoE encoders
90.
If a numeric feature has strong outliers, which scaling method is most aligned with for robustness?
a)
RobustScaler using median and IQR
b)
MinMaxScaler
c)
StandardScaler
d)
One-hot encoding
e)
Target encoding
91.
A numeric feature is highly skewed and can take negative values. Which listed transform is applicable?
a)
Yeo-Johnson
b)
Box-Cox
c)
One-hot encoding
d)
WoE
e)
Frequency encoding
92.
Which statement best matches the IQR rule described for outlier detection?
a)
Flag values below Q1 - 1.5*IQR or above Q3 + 1.5*IQR
b)
Flag values outside mean plus or minus 0.1
c)
Flag values that appear in rare categories only
d)
Flag values only if the model accuracy drops
e)
Flag values that are exactly equal to the median
93.
Why do caution against automatically removing outliers without investigation?
a)
Outliers may carry real business signal and should be reviewed with domain context
b)
Outliers always indicate data leakage
c)
Outliers are always duplicates and can be safely deleted
d)
Outliers cannot be handled by robust scaling
e)
Outliers only occur in categorical features
94.
ARPU is defined as revenue divided by usage. What general feature engineering pattern does this represent?
a)
A ratio or rate feature that can normalize for exposure
b)
A categorical encoding method
c)
A data splitting strategy
d)
A filter-based feature selection method
e)
One-hot encoding of categorical variables
95.
For seasonality in cyclic time variables, why might sine-cosine encoding be preferred over using the raw numeric hour?
a)
It represents cyclic structure smoothly so adjacent times across the boundary remain close
b)
It guarantees no missing values remain after encoding
c)
Min-max scaling of numeric features
d)
It replaces the need for scaling and outlier handling
e)
It prevents target leakage without cross-validation
96.
You want regional aggregation from latitude/longitude for modeling. Which listed technique directly supports this?
a)
Grid encoding such as geohash
b)
StandardScaler
c)
RFE
d)
Little's MCAR test
e)
Elastic Net
97.
You engineer group means per plan and worry the feature might leak information when computed on the full dataset. Which listed tactic addresses this?
a)
Use leave-one-out strategies for aggregation features
b)
Use global mean imputation
c)
Use MinMaxScaler on the target
d)
Use t-SNE to embed plans
e)
Use hashing to remove leakage
98.
Which engineered feature is most consistent with the time-series toolkit described?
a)
Rolling mean over past windows such as t-7 to t-1
b)
Using future values such as t+1 in features
c)
Encoding time by the random seed
d)
Replacing all lags with the target label
e)
Dropping the time index after sorting randomly
99.
Which feature selection method is described as suitable for non-linear dependencies?
a)
Mutual Information
b)
Chi-square
c)
ANOVA F-test
d)
StandardScaler
e)
MinMaxScaler
100.
Why do wrapper methods like RFE involve a trade-off?
a)
They can improve accuracy but increase compute time due to repeated model fitting
b)
They require no estimator and are therefore cheap
c)
They are only applicable to text features
d)
They eliminate the need for cross-validation entirely
e)
They are the same as embedded L1 regularization
101.
Which embedded method is specifically associated with sparsity?
a)
Lasso (L1)
b)
Ridge (L2)
c)
t-SNE
d)
Box-Cox
e)
Isolation Forest
102.
Which practice best matches the 'golden rules' guidance when tuning hyperparameters?
a)
Wrap hyperparameter search around the full preprocessing pipeline
b)
Tune hyperparameters on preprocessed data created from the full dataset
c)
Fit encoders on the test set to improve generalization
d)
Skip preprocessing during tuning to avoid leakage
e)
Use only a single train/test split and no pipeline
103.
Which evaluation metric listed most directly targets probability calibration quality?
a)
Brier score
b)
ROC-AUC
c)
PR-AUC
d)
Min-max scaling of numeric features
e)
Duplicate rate
104.
A producer changes a column's type unexpectedly during normal operations, breaking downstream features. Which scorecard or contract element is most directly violated per?
a)
Stable schema outside maintenance windows
b)
Duplicate rate below 0.5%
c)
Out-of-range below 0.1%
d)
Missing rate below 2% per feature
e)
Holiday flag completeness
105.
Income is missing more in Prepaid customers, and within Prepaid the missingness is higher for customers with very low income. Which interpretation best matches guidance?
a)
Primarily MCAR because the pattern is random
b)
MAR by plan_type, with a possible MNAR component for income
c)
Pure MNAR only because plan_type is observed
d)
Train-test contamination because plan_type is categorical
e)
Target leakage because income predicts churn
106.
You impute income but the act of being missing is strongly related to churn. Which approach best preserves this predictive signal while still filling values?
a)
Impute and add an is_missing indicator feature
b)
Impute with global mean and remove the missingness indicator
c)
Drop all rows with missing income to avoid bias
d)
Scale income to [0,1] and ignore missingness
e)
Use t-SNE on income values only
107.
Missing tenure is more common among churners. If you compute a single imputation value using the entire dataset before splitting, what is the most likely failure mode?
a)
It can leak label-related distribution information into training via train-test contamination
b)
It reduces dimensionality explosion from one-hot encoding
c)
It guarantees unbiased estimates under MNAR
d)
It prevents temporal leakage automatically
e)
It improves interpretability of hashing collisions
108.
You want to use kNN imputation but must avoid leakage. Which implementation is most consistent with?
a)
Fit scaling and kNN imputation inside a Pipeline so they are learned only from training folds
b)
Fit scaling on the full dataset once, then run kNN imputation inside CV
c)
Impute first using the full dataset, then split and train
d)
Impute only the test set so training remains clean
e)
Avoid scaling because kNN is not scale-sensitive
109.
Several features are incomplete and you want to preserve correlation structure under an MAR assumption. Which imputation method is best aligned with?
a)
MICE (multiple imputation by chained equations)
b)
Global mean imputation
c)
Row deletion
d)
One-hot encoding of missing values
e)
Isolation Forest imputation
110.
You predict churn at time T0. A feature records refunds that occur within 30 days after T0. Which mitigation best matches the guidance?
a)
Drop the feature or time-shift it to use only information from before T0
b)
Keep it because it improves validation ROC-AUC
c)
Replace it with one-hot encoding to remove leakage
d)
Fit it on the full dataset to stabilize estimates
e)
Scale it with RobustScaler to prevent leakage
111.
A rolling average feature is computed using values from t-3 through t+1. For forecasting, what is the key problem according to?
a)
It introduces temporal leakage by using post-hoc (future) information
b)
It violates the assumption that one-hot encoding is bounded
c)
It causes hashing collisions
d)
It guarantees MNAR missingness
e)
It prevents drift monitoring
112.
Which procedure most directly prevents target-encoding leakage during model evaluation?
a)
For each CV split, compute category target means on the training fold and apply them only to the corresponding validation fold
b)
Compute category target means once using all data and reuse in every fold
c)
Compute target means on the validation fold to mimic deployment
d)
Replace categories with random integers without any CV
e)
One-hot encode the target and join it to features
113.
You have a high-cardinality categorical feature in a streaming system, and you need fixed memory with acceptable approximation. Which encoding choice best matches?
a)
Hashing trick to map categories into a fixed-dimensional space
b)
Unbounded one-hot encoding for perfect fidelity
c)
Ordinal encoding with arbitrary integers
d)
Target encoding computed on the full stream without CV
e)
Chi-square feature selection
114.
A WoE model is built but several bins have very few observations. Based, what is the most likely concern?
a)
Unstable WoE estimates because binning needs enough samples
b)
Dimensionality explosion because WoE creates one column per category
c)
Hash collisions reduce interpretability
d)
Temporal leakage is unavoidable with WoE
e)
WoE cannot be used in credit risk
115.
A pipeline combines a tree-based model with a linear model in a stacked ensemble. According to, why might scaling still be beneficial even though trees are less scale-sensitive?
a)
Scaling can help when mixing models because some components are scale-sensitive
b)
Scaling is required to make one-hot encoding work
c)
Scaling eliminates the need for target encoding
d)
Scaling prevents all forms of leakage automatically
e)
Scaling replaces feature selection
116.
You want a variance-stabilizing transform, but your feature can be zero or negative. Which choice is consistent with?
a)
Yeo-Johnson
b)
Box-Cox
c)
Log transform only
d)
MinMaxScaler
e)
WoE
117.
A linear model performs poorly because a feature violates normality assumptions. Which listed transform is designed to map data closer to a normal distribution?
a)
Quantile or rank-Gauss transform
b)
Deduplication using edit distance
c)
Group-aware splitting
d)
Hashing trick
e)
RFE
118.
A spike in monthly charges may indicate fraud or a premium segment. Which handling policy best matches the guidance on outliers?
a)
Investigate with domain context and consider robust methods (capping, robust loss) rather than unconditional deletion
b)
Delete all high-charge records automatically to stabilize training
c)
Impute outliers using global mean computed on all rows
d)
One-hot encode outliers as separate categories for numeric variables
e)
Apply Box-Cox and assume outliers disappear
119.
Which option contains only feature transformation techniques explicitly listed?
a)
Binning, interactions, polynomials, and ratio/rate features
b)
Min-max scaling of numeric features
c)
Standardization using z-scores
d)
Median imputation with an indicator
e)
IQR-based outlier capping
120.
Why does the cyclic encoding use both sine and cosine rather than a single raw time index?
a)
To avoid a discontinuity at the cycle boundary and preserve circular distance relationships
b)
To ensure missing values become MCAR
c)
To reduce hashing collisions
d)
To eliminate the need for cross-validation
e)
To guarantee monotonicity for WoE
121.
You compute an average churn rate per plan and use it as a feature. What listed practice best reduces leakage risk in such group-by features?
a)
Leave-one-out aggregation so each row's feature excludes its own target contribution
b)
Compute the aggregation on the full dataset to increase stability
c)
Compute the aggregation on the test set only
d)
Replace the aggregation with one-hot encoding
e)
Drop the plan feature entirely because leakage is unavoidable
122.
You have limited compute budget but want strong feature selection. Which statement best captures the trade-off between wrapper and embedded methods?
a)
Wrapper methods can be more accurate but are compute-heavy, while embedded methods like L1 regularization select features during training
b)
Wrapper methods are cheaper because they do not require an estimator
c)
Embedded methods always require repeated refits across many subsets
d)
Filter methods require the most compute because they train many models
e)
t-SNE is the recommended feature selection method for deployment
123.
You want dimensionality reduction in a predictive pipeline (not just visualization). Which approach is more consistent with?
a)
Use PCA for variance-preserving components rather than t-SNE/UMAP for forecasting
b)
Use t-SNE because it is designed for forecasting
c)
Use UMAP and treat its 2D plot as a forecast
d)
Skip splitting and use visualization to evaluate generalization
e)
Use WoE to reduce numeric dimensions
124.
A team runs GridSearchCV but precomputes imputation and scaling once on the full dataset before the search. What is the key issue per the pipeline rules?
a)
Train-test contamination because preprocessing was fitted on full data instead of within the CV pipeline
b)
Dimensionality explosion because scaling increases features
c)
Missingness becomes MNAR due to scaling
d)
Hash collisions in one-hot encoding
e)
WoE requires binning and therefore grid search fails
125.
In production, a categorical feature starts producing many new levels. Which listed monitoring check most directly flags this issue?
a)
Cardinality monitoring with alerts and thresholds
b)
StandardScaler mean tracking only
c)
ANOVA F-test reruns on a fixed holdout set
d)
Box-Cox transform logging
e)
Little's MCAR test on the label
126.
A new field is added to a dataset used by multiple consumers. Which practice is highlighted to manage this safely?
a)
Schema versioning with planned backfill strategies
b)
One-time manual patching without version control
c)
Dropping the field immediately in all consumers
d)
Encoding the field with hashing to avoid schema changes
e)
Randomly renaming columns at each release
127.
Which combination best supports reproducible experiments as described?
a)
Fix random seeds, record environment files, and track parameters, metrics, artifacts, and data snapshots
b)
Rely on memory of the last run and skip logging
c)
Change preprocessing rules between folds to explore variability
d)
Avoid saving any artifacts to reduce storage costs
e)
Use the full dataset for preprocessing to reduce randomness
100 %
