Visual Analytics & Data Storytelling

Question

What should a visualization actually communicate?

A chart can be technically correct and still fail. It can bury a finding, make a weak model look convincing, or make an association appear causal.

Across three analytical contexts, I used visualization for different purposes with one objective: identify the message the evidence supports, then design the visual around making that message easier to understand.

Context

These projects cover manufacturing classification, public-opinion visualization, and cross-country happiness comparisons. They demonstrate a shared communication discipline across different questions; they are not parts of one research study.

Data

The manufacturing analysis contains 26,570 observations: 20,921 successful outcomes and 5,649 failures. Its exploratory work examines class balance, missingness, product groups, and measured predictors.

The public-opinion exhibit compares polling percentages and observation totals across pollster groups. The happiness exhibits compare reported happiness with GDP per capita and corruption-perception measures across countries.

Evidence on this page consists of the original project exports and accompanying analysis narrative. Dashboard images are static exhibits; select an image to inspect the full-size export.

Approach

Message first, visual second

Effective visualization starts before choosing a chart. The intended message determines which comparisons matter, how the method should be explained, and which limitations must stay visible.

1. MessageWhat should someone leave understanding?
2. EvidenceWhat data supports that message?
3. MethodWhich technique fits the question?
4. Visual formWhat makes the evidence easiest to interpret?
5. InterpretationWhat should the viewer notice first?
6. BoundaryWhat might someone incorrectly infer?

Predict

A model can look accurate and still be weak

Initial screening identified loading as the strongest relationship with failure, with Measurement 17 also showing a positive relationship. The analysis began by making class imbalance, missing data, and potential signal visible.

The logistic model reported approximately 0.79 accuracy; the decision tree reported 0.78. Those headline metrics do not establish whether the models reliably distinguish failures from successful outcomes.

For context, predicting the majority class for every observation in the full dataset would be correct about 78.7% of the time (20,921 ÷ 26,570). This is a full-dataset reference, not a reproduced test-set benchmark.

The original findings export reports testing accuracy of 0.7892756 and testing ROC-AUC of 0.3925843. Before operational use, the evaluation needs review—including positive-class coding and probability direction—alongside further validation and model development.

Potential signal is not deployment evidence

The decision-tree exhibit identifies potentially useful variables, but feature importance alone does not establish reliable predictive performance.

Decision: communicate the model’s limitations before asking anyone to trust its predictions.

View original model findings

Explore

A dashboard should make the next question easier to answer

The public-opinion visualization organizes several comparisons around the same question: how do polling percentages and observation totals differ across pollster groups?

The viewer first sees the spread of polling percentages, then the difference in total observations, then the individual pollsters behind those totals. Each view provides context for the next.

The observed spread invites investigation of sample size, timing, population, and methodology. Larger aggregate observation totals alone do not establish better poll quality or prove why the groups differ.

Decision: organize reporting around analytical questions and comparability, with definitions visible alongside the evidence.

Explain

A strong relationship is not proof of causation

The happiness project compares reported happiness with economic and governance-related measures. The visual sequence begins with geography and country rankings, then narrows to the relationships themselves.

From the broad pattern to a focused comparison

The original visual analysis suggested a stronger association between corruption-perception measures and reported happiness than between GDP per capita and happiness in the analyzed data. The side-by-side plots make that interpretation inspectable.

The precise conclusion is: governance perceptions showed a stronger association with reported happiness than GDP in the original visual analysis. This is not a claim that governance perceptions cause happiness or have a larger causal effect.

The export includes null indicators, and the supplied exhibits do not provide a reproducible comparison of correlation coefficients or model fits. The apparent relationship should therefore be read as the project’s visual interpretation, not a newly validated statistical ranking.

Decision: make the evidence clearer while preserving the boundary of what it supports.

Findings

Communicate the clearest message the evidence can legitimately support.

Decision

Treat communication as part of the analytical method.

Visualization is not something added after analysis. The intended message informs data preparation, comparison design, method selection, interpretation, and the limitations that need to remain visible.

Outcome

The work demonstrates three complementary capabilities: translating model evaluation into an understandable judgment, organizing related comparisons so viewers can investigate a question, and explaining statistical relationships without removing uncertainty.

Turning analysis into a message another person can understand, question, and use.

No measured business lift or deployment outcome is claimed for these visualization projects.

Limitations

• The projects use different datasets and methods and should not be interpreted as one study.
• Manufacturing model performance requires further validation and evaluation review before operational use.
• Polling comparisons require attention to populations, dates, definitions, and methods; aggregate sample totals do not establish quality.
• The happiness analysis is observational. Missing-value indicators and the lack of underlying data here limit independent verification.
• Original exports retain their original wording. The surrounding interpretation qualifies causal language and unsupported implications.

Tools / Methods

Tableau · Logistic regression · Decision trees · Model evaluation · Geospatial visualization · Statistical visualization · Data preparation · Analytical communication