WorksheetsBA 4
Total questions: 80
Worksheet time: 40mins
Name
Class
Date
1.
In the running case, what is the goal of analyzing the Telco Churn dataset?
a)
Understand drivers and patterns before modeling
b)
Encrypt customer records for compliance audits
c)
Optimize network routing using shortest paths
d)
Build a rule-based expert system for billing
e)
Benchmark GPU performance on large matrices
2.
Which statement best matches the lecture's definition of EDA?
a)
Systematic exploration of data to form or test hypotheses
b)
A method for compressing data without losing information
c)
A technique for guaranteeing causal conclusions from correlations
d)
An algorithm for training a predictive model end-to-end
e)
A process for converting unstructured text into code
3.
Which sequence is included in the EDA mindset described in the lecture?
a)
Profile, explore, hypothesize, validate
b)
Collect, encrypt, archive, delete
c)
Train, prune, distill, deploy
d)
Compile, link, package, release
e)
Sample, sort, hash, sign
4.
Which pair of data types is highlighted under data types and scales?
a)
Numeric and categorical
b)
Binary and hexadecimal
c)
Textual and executable
d)
Audio and video
e)
Encrypted and decrypted
5.
According to the lecture, a main goal of univariate EDA is to understand:
a)
Shape, center, and spread of a variable
b)
The optimal causal graph of the system
c)
The final model hyperparameters
d)
The exact pixel values of a chart
e)
The network bandwidth between servers
6.
Which set of items is included in the lecture's summary statistics list?
a)
Mean, median, mode, variance, standard deviation, quantiles, IQR
b)
Precision, recall, F1-score, AUC, log-loss
c)
HTTP status codes, request headers, TLS versions
d)
CPU cores, GPU clocks, disk throughput, ping time
e)
Hash functions, salts, keys, signatures, certificates
7.
Which bin-width selection rules are mentioned for histograms?
a)
Freedman-Diaconis (FD) rule and sqrt rule
b)
Bayes rule and Markov rule
c)
Dijkstra rule and Bellman-Ford rule
d)
CAP rule and BASE rule
e)
FIFO rule and LIFO rule
8.
What does the lecture state that boxplots are useful for showing?
a)
Spread and outliers
b)
Exact causal effects
c)
Guaranteed normality of data
d)
The optimal number of clusters
e)
The true underlying data-generating process
9.
In the lecture, a histogram is used primarily to visualize:
a)
The distribution of a numeric variable
b)
The sequence of events in a workflow
c)
The dependencies among software modules
d)
The topology of a network graph
e)
The encryption strength of a cipher
10.
A bar chart is presented as an appropriate univariate visualization for:
a)
Counts by category (e.g., plan type)
b)
Continuous-time signals at high frequency
c)
Model residual autocorrelation
d)
Neural network layer activations
e)
Transaction latency in milliseconds per request
11.
Which is listed as a goal of bivariate EDA?
a)
Quantify pairwise relationships
b)
Guarantee causal inference without experiments
c)
Eliminate all missing data automatically
d)
Replace modeling with visualization
e)
Determine the only correct business decision
12.
For numeric-numeric relationships, the lecture recommends using:
a)
Scatter plots and correlations
b)
Only pie charts
c)
Only stacked bar charts
d)
Only text tables without any plots
e)
Only 3D surface plots
13.
For numeric-categorical comparisons, the lecture suggests using:
a)
Grouped boxplots and strip or dot plots
b)
Only correlation heatmaps
c)
Only line charts with dual y-axes
d)
Only 3D pie charts
e)
Only PCA biplots
14.
For categorical-categorical analysis, the lecture highlights:
a)
Contingency tables and stacked or grouped bars
b)
Kernel density estimates for continuous variables
c)
Backpropagation through time
d)
Monte Carlo tree search
e)
Kalman filtering
15.
In a scatter plot for two numeric variables, the lecture advises looking for:
a)
Nonlinearity and heteroscedasticity
b)
The exact RGB values of points
c)
The font size used in axis labels
d)
The position of the legend on the page
e)
The slide number printed in the corner
16.
To compare distributions of a numeric variable across categories, the lecture shows a:
a)
Boxplot by group
b)
Single-category pie chart
c)
3D cone chart
d)
Word cloud
e)
Sankey diagram
17.
A correlation heatmap in the lecture is used to summarize:
a)
Pairwise correlations among numeric features
b)
Exact causal effects among all variables
c)
Customer identities and account passwords
d)
Server utilization in real time
e)
A complete list of data cleaning steps
18.
According to the lecture, a key goal of multivariate EDA is to:
a)
Reveal joint structures and interactions beyond pairs
b)
Avoid examining segments and conditional patterns
c)
Replace all charts with a single number
d)
Force linear relationships in all data
e)
Compute cryptographic hashes for features
19.
Segmented views and faceting are described as a way to:
a)
Use small multiples to compare segments
b)
Hide differences between groups
c)
Make charts harder to interpret
d)
Guarantee causality from correlations
e)
Avoid any stratification by key variables
20.
A scatter matrix (concept) is described as:
a)
A grid of pairwise scatter plots
b)
A single line chart of cumulative totals
c)
A table of primary keys and foreign keys
d)
A neural network weight matrix
e)
An encryption key schedule
21.
In the EDA lens, PCA is mentioned primarily for:
a)
Linear structure
b)
Detecting malware in binaries
c)
Estimating causal effects from observational data
d)
Choosing bin widths for histograms
e)
Designing database indexes
22.
Which feature-interaction idea is mentioned as an EDA cue?
a)
Ratios or differences and domain-informed composites
b)
Only raw features with no transformations
c)
Only one-hot encoding for every variable
d)
Only random noise features
e)
Only encrypted feature hashes
23.
Why does the lecture emphasize correlation vs. causation?
a)
Decisions require causal reasoning; spurious correlations are common
b)
Correlation always proves causation in business data
c)
Causation can be inferred from a single scatter plot
d)
Correlation is useless and should never be computed
e)
Causation is irrelevant for decision-making
24.
In the lecture's intuition, a confounder is a variable that:
a)
Affects both X and Y
b)
Is affected by both X and Y
c)
Has no relationship to X or Y
d)
Always eliminates missing data
e)
Guarantees randomized assignment
25.
Simpson's paradox is described as a situation where:
a)
A pooled trend differs from within-group trends
b)
A histogram always has the same shape as a boxplot
c)
Pearson and Spearman correlations are identical
d)
A bar chart always requires a dual y-axis
e)
A dashboard cannot include thresholds
26.
What is the lecture's key advice related to Simpson's paradox?
a)
Always check stratified views
b)
Always use a 3D pie chart
c)
Always remove all outliers
d)
Always standardize all variables to mean zero
e)
Always avoid comparing groups
27.
According to the lecture, randomized experiments are primarily used to:
a)
Identify causal effects
b)
Compute correlations more quickly
c)
Guarantee no missing values
d)
Replace all EDA steps
e)
Force linearity in the data
28.
A practical A/B testing step mentioned in the lecture is to:
a)
Define metrics (primary and guardrails), sample size, and duration
b)
Choose a chart color palette before defining metrics
c)
Stop the test as soon as results look positive
d)
Remove randomization to speed up deployment
e)
Ignore guardrails to maximize the primary metric
29.
From EDA to decisions, the lecture recommends moving from patterns to:
a)
Testable hypotheses
b)
Guaranteed causal claims
c)
Random guesses
d)
Irreversible business changes without validation
e)
Only aesthetic visual improvements
30.
In visualization principles, which channels are described as the most precise?
a)
Position and length
b)
Area and volume
c)
Color saturation and texture
d)
Font choice and slide theme
e)
3D depth and perspective
31.
Which mapping of analytical question to chart type is explicitly listed?
a)
Distribution: histogram or boxplot
b)
Distribution: pie chart with 3D effects
c)
Relationship: stacked bar chart only
d)
Composition: scatter plot only
e)
Time series: contingency table only
32.
For bar charts, the lecture recommends which baseline practice?
a)
Use a zero baseline
b)
Always truncate the axis for emphasis
c)
Use a dual y-axis by default
d)
Always plot bars on a log scale
e)
Avoid labeling the axis entirely
33.
Binning and smoothing are described as involving a tradeoff between:
a)
Noise and detail
b)
Speed and encryption strength
c)
Memory and CPU cores
d)
File size and network latency
e)
Accuracy and fairness guarantees
34.
Small multiples and faceting are recommended to compare patterns across:
a)
Groups or time with repeated layouts
b)
Only different font families
c)
Only slide backgrounds
d)
Only chart colors
e)
Only one variable at a time
35.
Avoiding chartjunk includes the practice of:
a)
Removing non-data ink
b)
Adding extra decorative shapes to every chart
c)
Maximizing the number of encodings per chart
d)
Using as many annotations as possible
e)
Replacing titles with artistic icons
36.
Dashboards for decision-making should primarily:
a)
Tie visuals to questions and actions and highlight targets or thresholds
b)
Hide targets to avoid bias
c)
Focus only on decorative design
d)
Exclude any context to keep visuals minimal
e)
Avoid sharing any results with stakeholders
37.
Which item is included in the lecture's EDA workflow and reproducibility guidance?
a)
Versioned data and saved figures or tables
b)
Only ad-hoc screenshots without source code
c)
Untracked manual edits to datasets
d)
Keeping findings private to avoid review
e)
Deleting intermediate outputs immediately
38.
Which item appears in the EDA checklist?
a)
Types, ranges, missingness, and duplicates
b)
Only final model accuracy
c)
Only marketing campaign slogans
d)
Only network packet traces
e)
Only application UI screenshots
39.
Which is listed as a common anti-pattern in EDA and visualization?
a)
Cherry-picking
b)
Stratifying by key segments
c)
Documenting assumptions
d)
Quantifying uncertainty
e)
Sharing EDA artifacts for review
40.
A missingness map is used to help:
a)
Diagnose missingness mechanisms before imputation
b)
Convert categorical variables into numeric codes automatically
c)
Guarantee that no data is missing
d)
Compute Pearson correlation coefficients
e)
Select the best deep learning architecture
41.
An analyst begins EDA by generating many plots without a business question. Which lecture mindset principle is being violated?
a)
Start with questions and context; avoid aimless fishing
b)
Use only multivariate EDA and skip univariate checks
c)
Assume correlations are causal until proven otherwise
d)
Avoid iterating; do EDA in a single pass
e)
Never validate hypotheses once formed
42.
You notice a numeric variable has extreme values and a long tail. Which summary approach is emphasized as more appropriate in such cases?
a)
Use robust summaries rather than relying only on the mean
b)
Drop the variable because tails are always errors
c)
Encode the variable as nominal categories immediately
d)
Replace the distribution with a single correlation value
e)
Assume the data is normal and ignore tails
43.
A team compares two histograms for different segments but uses different bin widths in each. According to the lecture, what is the most direct risk of this choice?
a)
It can make comparisons misleading because bins should be consistent
b)
It forces the data to become normally distributed
c)
It guarantees the presence of outliers in both segments
d)
It eliminates the need for any bivariate analysis
e)
It proves that one segment causes the other to change
44.
A histogram looks very different after you change the bin width. Which lecture topic explains why this can happen?
a)
Histogram bin width selection can change the apparent distribution
b)
Pearson correlation is invariant to binning
c)
Chi-square tests require binning to be identical to boxplots
d)
PCA requires all variables to be binned first
e)
A/B tests are invalid unless you re-bin all metrics daily
45.
You want to compare a numeric variable across plan types while still seeing individual observations. Which combination aligns with the lecture?
a)
Grouped boxplots plus strip or dot plots for raw observations
b)
Stacked bars plus a correlation heatmap
c)
A single pie chart plus a scatter matrix
d)
A 3D pie chart plus a dual y-axis line chart
e)
Only a contingency table without any visualization
46.
A scatter plot of X vs Y shows the spread of Y increasing as X increases. Which issue does the lecture suggest you should check for?
a)
Heteroscedasticity
b)
Perfect causal identification
c)
Guaranteed normal residuals
d)
Duplicate category labels
e)
Primary key violations
47.
You are asked to quantify pairwise relationships and compare distributions across categories before modeling. Which EDA level is being described?
a)
Bivariate EDA
b)
Univariate EDA
c)
Only data ingestion
d)
Only model deployment
e)
Only database indexing
48.
A contingency table is created for two categorical variables. Which high-level test concept is paired with this in the lecture?
a)
Chi-square test using expected vs observed counts
b)
t-test comparing two sample means
c)
ANOVA requiring numeric-numeric relationships
d)
PCA to reduce the number of categories
e)
A/B testing with randomization and bucketing
49.
A correlation heatmap shows several strong correlations. According to the lecture, what is the most appropriate interpretation stance for decision-making?
a)
Treat correlations as suggestive signals and avoid causal claims without further tests
b)
Treat the largest correlation as proof of causality
c)
Ignore scatter plots because heatmaps are sufficient
d)
Assume spurious correlations cannot occur in business data
e)
Conclude that multivariate analysis is unnecessary
50.
A pooled trend between charges and churn looks different when you split customers by plan type. Which lecture concept directly motivates checking these stratified views?
a)
Simpson's paradox and the need for stratification
b)
Histogram binning rules
c)
Marks and channels precision
d)
Avoiding chartjunk through non-data ink removal
e)
Dimensionality reduction for linear structure
51.
Before building a model, the lecture suggests a scatter matrix can be used primarily to:
a)
Sanity check pairwise relationships and patterns
b)
Prove the best causal graph
c)
Guarantee the best hyperparameters
d)
Replace all missing values automatically
e)
Compute expected counts for chi-square tests
52.
Which statement best matches the lecture's warning about t-SNE and UMAP in EDA?
a)
Use them for visualization only, not as definitive modeling evidence
b)
Use them to estimate causal effects in observational data
c)
Use them to choose histogram bin widths
d)
Use them to define guardrail metrics for A/B tests
e)
Use them to convert categorical variables into ordinal scales
53.
You create a ratio feature during EDA to capture a domain-informed interaction. What does the lecture recommend doing next to keep the analysis decision-oriented?
a)
Record the hypothesis so it can be tested later
b)
Assume the ratio feature is causal and act immediately
c)
Delete the original variables to avoid multicollinearity
d)
Avoid any further validation because EDA is sufficient
e)
Only report the ratio in a dashboard without context
54.
A variable Z affects both X and Y. In the lecture's causal intuition terminology, Z is best described as a:
a)
Confounder
b)
Collider
c)
Guardrail metric
d)
Bin width
e)
Non-data ink
55.
A variable Z is influenced by both X and Y. The lecture notes that conditioning on Z can distort relationships. Z is best described as a:
a)
Collider
b)
Confounder
c)
Primary metric
d)
Scatter matrix
e)
Log scale
56.
Which action is most aligned with the lecture's guidance when you suspect confounding in an observational dataset?
a)
Use stratification and DAG-thinking to reason about variables
b)
Increase chart decorations to make patterns clearer
c)
Treat correlation coefficients as causal effects
d)
Avoid segmentation to keep the analysis simple
e)
Only compute summary statistics and skip visualization
57.
Which approach is highlighted as the most direct way to identify causal effects?
a)
Randomized experiments
b)
Correlation heatmaps
c)
Stacked bar charts
d)
Scatter matrices
e)
Histogram binning rules
58.
In practical A/B testing, which pairing of metric types is explicitly recommended?
a)
Primary metrics and guardrail metrics
b)
Expected counts and observed counts
c)
Position channels and volume channels
d)
Pearson correlations and kurtosis values
e)
IQR values and chi-square p-values
59.
An analyst repeatedly checks interim A/B test results and stops early when the primary metric looks good. Which lecture warning does this violate?
a)
Avoid peeking
b)
Always use dual y-axes
c)
Always prefer 3D pies
d)
Never stratify by segments
e)
Always ignore uncertainty
60.
To compare magnitudes as accurately as possible, the lecture recommends favoring which visual encoding channels?
a)
Position or length
b)
Area or volume
c)
3D perspective
d)
Texture density
e)
Random color gradients as the main encoding
61.
Which chart type is explicitly discouraged in the lecture's chart selection guidance?
a)
3D pies
b)
Histograms
c)
Boxplots
d)
Scatter plots
e)
Stacked bars
62.
Which axis practice is discouraged unless necessary, according to the lecture?
a)
Using dual y-axes
b)
Using clear titles and labels
c)
Using a zero baseline for bars
d)
Using a log scale for heavy tails
e)
Using small multiples to compare segments
63.
You smooth or bin data to reduce noise, but the plot now hides meaningful variation. Which lecture tradeoff explains this tension?
a)
The tradeoff between noise and detail
b)
The tradeoff between correlation and causation
c)
The tradeoff between primary and guardrail metrics
d)
The tradeoff between confounders and colliders
e)
The tradeoff between expected and observed counts
64.
A decision-making dashboard is being designed for executives. Which element is emphasized as especially important in the lecture?
a)
Highlighting targets or thresholds tied to actions
b)
Maximizing decorative elements to keep attention
c)
Hiding context to avoid influencing decisions
d)
Using as many encodings per chart as possible
e)
Avoiding any documentation to keep the work lightweight
65.
During EDA, an analyst finds a striking correlation after trying many different filters and then reports only that view to justify a decision. Which combination best matches the lecture's critique and remedy?
a)
This is cherry-picking; start from explicit questions and triangulate with multiple views
b)
This is good practice; correlations alone are sufficient for decisions
c)
This is overplotting; fix it by adding more encodings per chart
d)
This is a binning problem; fix it by always using a 3D pie chart
e)
This is a PCA issue; fix it by applying t-SNE for causal inference
66.
A team compares income distributions across two segments using histograms. Segment A uses fine bins and Segment B uses coarse bins, and the x-axis ranges differ. Based on the lecture, what is the most defensible critique?
a)
The comparison can be misleading because binning and scales should be made comparable across segments
b)
The comparison is always valid because histograms are invariant to bin width and scale choices
c)
The issue is only that boxplots cannot show outliers
d)
The issue is that Pearson correlation should be used instead of histograms
e)
The issue is that contingency tables require expected counts
67.
A variable has a heavy tail and extreme outliers. Which approach best aligns with multiple lecture recommendations for summarizing and visualizing it without overreacting to extremes?
a)
Use robust summaries (e.g., median and IQR) and consider a log-type transform or log scale if appropriate
b)
Use only the mean and standard deviation because they ignore tails
c)
Convert the variable to nominal categories and skip distribution checks
d)
Rely only on a correlation heatmap to understand its behavior
e)
Remove all extreme values before any inspection and proceed directly to modeling
68.
A correlation heatmap suggests two variables are strongly related, but a scatter plot shows a curved pattern and changing spread. According to the lecture, what is the best interpretation and next step?
a)
Do not infer causality; inspect nonlinearity and heteroscedasticity and treat the pattern as a hypothesis to validate
b)
Conclude the relationship is causal because the correlation is strong
c)
Ignore the scatter plot because heatmaps provide the final answer
d)
Remove one variable immediately because strong correlation proves redundancy
e)
Switch to a 3D pie chart to confirm the correlation visually
69.
You observe that the overall relationship between tenure and churn differs from the relationship within each plan type. Which paired explanation and action is most consistent with the lecture?
a)
Simpson's paradox may be present; check stratified and faceted views before drawing conclusions
b)
This proves the pooled trend is correct; within-group trends should be ignored
c)
This proves a collider is present; remove the plan type variable
d)
This is a histogram binning artifact; change the FD rule constant
e)
This indicates PCA has failed; use t-SNE to estimate causality
70.
An analyst conditions on a variable that is influenced by both X and Y to 'control for it' and then reports the adjusted relationship. What does the lecture warn could happen, and why?
a)
Conditioning on a collider can distort the relationship because the collider is affected by both X and Y
b)
Conditioning always removes bias because any controlled variable is a confounder
c)
Conditioning guarantees causal identification in observational data
d)
Conditioning is irrelevant because correlations never change under conditioning
e)
Conditioning only affects histograms, not relationships
71.
A bivariate analysis shows little relationship between a feature and churn overall, but faceted views by complaints_30d reveal clear differences. What does this most directly illustrate in the lecture?
a)
Multivariate EDA can reveal conditional patterns and segment interactions beyond pairs
b)
Univariate EDA is sufficient and multivariate EDA is unnecessary
c)
Correlation heatmaps always reveal conditional patterns without segmentation
d)
Chartjunk improves when you add more encodings per chart
e)
A/B testing is the only way to visualize segments
72.
You want an EDA technique to reveal linear structure in high-dimensional numeric data, but you also want a purely visual tool for exploring clusters without treating it as evidence for modeling. Which pairing best matches the lecture?
a)
PCA for linear structure; t-SNE or UMAP for visualization only
b)
t-SNE for linear structure; PCA for visualization only
c)
Boxplots for linear structure; chi-square for visualization only
d)
Scatter matrices for linear structure; dual y-axes for visualization only
e)
Stacked bars for linear structure; 3D pies for visualization only
73.
A product team proposes an A/B test plan. Which plan best satisfies the lecture's practical guidance as a coherent set of choices?
a)
Define primary and guardrail metrics, set sample size and duration, randomize and bucket users, and avoid peeking during the run
b)
Pick a chart style first, then adjust metrics daily until results look good, and stop as soon as the primary metric increases
c)
Run the test without randomization to speed up, and monitor only the primary metric with no guardrails
d)
Randomize users but do not define duration or sample size, and report interim results every hour to decide when to stop
e)
Use observational correlations instead of experimentation because experiments are unnecessary for causal effects
74.
A stakeholder asks to act immediately on an EDA pattern. Which response best aligns with the lecture's 'from EDA to decisions' guidance?
a)
Translate the pattern into a testable hypothesis, quantify uncertainty, and document assumptions before acting
b)
Declare causality from the pattern and implement the change without validation
c)
Remove all outliers to make the pattern stronger and then act
d)
Avoid documenting assumptions because it slows down decision-making
e)
Replace the analysis with a single dashboard chart and skip validation
75.
You must choose a visual encoding to compare small differences in values across groups. Based on the lecture's marks and channels guidance, which choice is most defensible?
a)
Use position or length rather than area or volume encodings
b)
Use volume encodings because 3D is more realistic
c)
Use area encodings because they are always more precise than length
d)
Use as many encodings as possible to make differences obvious
e)
Use decorative shapes to improve precision
76.
A dashboard uses bars to compare totals but the y-axis starts above zero to make differences look larger. Which lecture principle is being violated and why is it risky?
a)
Zero baseline for bars is recommended; truncation can exaggerate differences
b)
Dual y-axes are recommended; a single y-axis hides information
c)
Area encodings are recommended; bars should be avoided
d)
PCA requires non-zero baselines for stability
e)
Chi-square tests require truncated axes to show expected counts
77.
A report includes a chart with many decorative elements, multiple encodings, and heavy annotation. Which lecture guidance most directly challenges this design?
a)
Remove non-data ink, limit encodings per chart, and use annotations sparingly
b)
Maximize non-data ink to increase engagement
c)
Add more encodings because more channels always improves clarity
d)
Use 3D pies to reduce cognitive load
e)
Avoid titles and labels to keep visuals minimal
78.
A missingness map shows that missing values are concentrated in a specific customer segment. According to the lecture, what is the most appropriate immediate response before imputation?
a)
Diagnose the missingness mechanism, potentially using stratified views, before choosing an imputation approach
b)
Impute immediately with a single method because missingness patterns do not matter
c)
Drop the segment entirely because missingness implies fraud
d)
Treat the missingness map as proof of causality
e)
Replace the missingness map with a correlation heatmap to decide imputation
79.
A team wants their EDA to be auditable and repeatable. Which combination best matches the lecture's reproducibility guidance?
a)
Use notebook scripts, version the data, save figures or tables, and share artifacts for review
b)
Rely on manual steps only, avoid saving outputs, and keep the dataset unversioned
c)
Store only screenshots in email threads and delete code after presenting
d)
Change the dataset in place without tracking and avoid peer review to reduce bias
e)
Use dual y-axes everywhere so results are more convincing
80.
An EDA dashboard highlights a correlation and suggests a policy change. Which critique best synthesizes the lecture's cautions about decision-making from EDA?
a)
Correlation can be spurious; consider confounding, use stratified checks, and validate with experiments or quasi-experiments before causal decisions
b)
If a dashboard highlights a correlation, it is automatically causal and action-ready
c)
Stratification is unnecessary because pooled trends are always correct
d)
Confounders and colliders only matter in machine learning, not business decisions
e)
EDA is complete once you have a single strong correlation
100 %
