5.13: Detailed Figure Captions - Chapter 5

Exploratory Data Analysis (EDA) -- Figure Specifications

Overview

• Chapter: 5 -- Exploratory Data Analysis (EDA)
• Total Figures Proposed: 8
• Figure Types Used: Flowchart, Conceptual Illustration, Data Chart, Comparison Table, Infographic, Architecture Diagram

---

Figure 5.1.1

Title: EDA in the Data Analysis Pipeline 

Detailed Caption: Five gray pipeline tiles run left to right — Raw Data Collection → Data Cleaning & Preprocessing → Exploratory Data Analysis (EDA) → Modeling & Confirmatory → Reporting & Communication — connected by navy arrows. The middle EDA tile is filled in teal and enlarged with a magnifying-glass glyph, drawing the eye, and three warm-gold sub-tiles drop beneath it labeled Summarize, Visualize, and Discover Patterns. A dashed navy "Iterate" arc loops back from the Modeling tile to the EDA tile, signaling that downstream findings often surface new questions worth exploring. By centering EDA visually within an otherwise standard left-to-right pipeline, the figure makes the chapter's positioning argument concrete: EDA is neither cleaning nor modeling, but the bridge between them where analysts develop the questions, hypotheses, and intuitions that downstream confirmatory work then tests rigorously. 

Figure 5.1.2

Title: Exploratory vs. Confirmatory vs. Explanatory Analysis 

Detailed Caption: A three-column matrix sits beneath color-coded header pills — teal "Exploratory," warm-gold "Confirmatory," and navy "Explanatory" — with five row labels down the left margin: Goal, Methods, Audience, Output, and Mindset. Each cell carries a one-line phrase that captures the mode's character: Exploratory pairs "Discover patterns, generate hypotheses" with "Plots, summary stats" for "The analyst (self)" producing "Rough plots, notes" in a "Detective" mindset; Confirmatory pairs "Test specific hypotheses" with "Statistical tests, p-values" for "Peer reviewers" producing "Test results, CIs" in a "Judge" mindset; and Explanatory pairs "Communicate findings to an audience" with "Polished charts, narratives" for "Stakeholders, decision-makers" producing "Reports, presentations" in a "Storyteller" mindset. Alternating row backgrounds aid horizontal scanning. By placing five dimensions of contrast across three modes on a single canvas, the table prevents the common confusion of conflating these activities — the same chart can serve all three roles, but the methods, audiences, and even the analyst's mindset shift completely depending on which mode is in play. 

Figure 5.2.1

Title: Anatomy of a Box Plot 

Detailed Caption: A single vertical box plot dominates the figure, drawn with a navy outline and a pale-teal interior, its key components annotated by leader lines extending right to plain-text labels. A warm-gold horizontal stripe inside the box is labeled "Median (Q2)"; the box's top edge is "Q3 (75th percentile)" and its bottom edge is "Q1 (25th percentile)." A right-side curly bracket spans the full box height labeled "IQR" (Interquartile Range). Whiskers extend upward and downward to dashed reference labels "Q3 + 1.5 × IQR" and "Q1 − 1.5 × IQR" respectively, and a long left-side bracket labeled "Non-outlier range" embraces the entire whisker span. By exposing every quartile, the IQR span, and the whisker formula on a single labeled artifact, the figure converts the box plot from a compact glyph into a working anatomy reference — students can return to it whenever they need to remind themselves what each line, edge, and tick actually represents in the underlying distribution. 

---

Figure 5.3.1

Title: Bivariate Analysis Decision Guide 

Detailed Caption: A navy decision diamond at the top reads "What types are the two variables?" with three colored branches fanning downward to category-specific recommendations. The teal "Both Numeric" branch leads to a Scatter Plot tile and a Correlation (r) tile; the warm-gold "Numeric + Categorical" branch leads to Box Plots, Violin Plot, and Grouped Bar tiles; and the coral "Both Categorical" branch leads to Cross-Tab and Stacked Bar tiles. Each leaf tile is rendered in a pastel of its parent's color so the lineage is visible at a glance. By converting a memorization task into a single decision tree keyed only on variable type, the figure gives students a working reference they can apply mid-analysis: identify the two variable types, follow the matching branch, and the appropriate visualization options sit one click away — eliminating the common stumble of plotting categorical data in a way that implies false continuity, or numeric data in a way that hides distributional shape.

---

Figure 5.4.1

Title: Multivariate EDA Techniques Overview 

Detailed Caption: Three horizontal cards stack vertically, each topped by a colored header bar that names a strategy family. The teal "Visual Encoding" card lists three techniques: Color/Shape by 3rd Variable, Faceted (Trellis) Plots, and Size Encoding — each using cosmetic dimensions of a 2-D chart to carry extra variables. The warm-gold "Matrix Views" card lists Scatter Plot Matrix (SPLOM) and Correlation Heatmap, both of which exhaust pairwise relationships in a single grid. The navy "Dimensionality Reduction" card lists PCA, t-SNE / UMAP, and Hierarchical Clustering, projecting many variables down to a manageable display. The deliberate ordering (encoding → matrix → reduction) traces the analyst's escalation path: when extra variables can be carried as visual channels, do so; when too many pairs exist to inspect individually, use a matrix; and when even a matrix is too large, reduce dimensionality. The three-tier taxonomy gives students a portable vocabulary for choosing among multivariate techniques rather than reaching for whichever method they saw most recently. 

Figure 5.5.1

Title: Missing Data Mechanisms and Handling Strategies 

Detailed Caption: Two side-by-side panels split the figure into the diagnosis side and the prescription side. The left "Missing Data Mechanisms" panel lists three rows: a teal circle marking MCAR — Missing Completely at Random ("Missingness unrelated to any variable"), a warm-gold triangle marking MAR — Missing at Random ("Missingness related to observed data"), and a coral square marking MNAR — Missing Not at Random ("Missingness related to the missing value itself"). The right "Handling Strategies" panel stacks four teal-bordered tiles: Deletion (listwise/pairwise), Simple Imputation (mean/median), Advanced Imputation (KNN, regression), and Flagging (create indicator variable). Color-coded dashed lines bridge the two panels, each line connecting a mechanism to its appropriate handling strategies — MCAR routes to Deletion and Simple Imputation, MAR routes to Advanced Imputation, and MNAR routes to Advanced Imputation and Flagging. The two-panel layout makes the chapter's central rule visible: the right way to handle missingness depends on why the data is missing, and the same imputation technique that's safe under MCAR can introduce serious bias under MNAR. 

Figure 5.6.1

Title: Outlier Detection and Decision Framework 

Detailed Caption: A navy header tile at the top — "Identify Potential Outliers (IQR Rule, Z-Score, Visual Inspection)" — opens a top-to-bottom flow through three teal decision diamonds. The first diamond, "Is it a data error?", branches Yes (right) to a warm-gold "Correct or Remove" terminal and No (down) to the next diamond. The second, "Relevant to analysis goal?", branches No (right) to a gray "Exclude with Documentation" terminal and Yes (down) to the third diamond. The third, "Distorts summaries?", branches No (right) to a gray "Keep and Analyze" terminal and Yes (down) to a final terminal off the bottom of the figure. By converting the abstract advice "decide what to do with outliers" into a concrete three-question funnel, the chart prevents two common failure modes — blanket removal of outliers (which destroys signal when they're real high-value events) and blanket retention (which lets data-entry errors corrupt summary statistics). The structured questions force the analyst to justify each handling choice rather than applying a uniform rule. 

Figure 5.7.1

Title: The EDA Workflow Cycle 

Detailed Caption: Six rounded-rectangle stages are arranged clockwise around a central "EDA · Iterate as needed" label, color-coded to mark progression: navy "Data Understanding" at the top, slate-blue "Data Cleaning" on the upper right, teal "Univariate Exploration" on the lower right, teal "Bivariate Exploration" at the bottom, teal "Multivariate Exploration" on the lower left, and warm-gold "Synthesis & Next Steps" on the upper left. Navy connector lines join each stage to the next around the loop, and a dashed marker in the center reinforces the iterative nature of the workflow. By drawing EDA as a closed loop rather than a one-pass checklist, the figure delivers the chapter's core argument: every round of synthesis raises new questions that flow back into Data Understanding and re-trigger the cycle, and the practical signal that EDA is "done" is not running out of techniques but reaching a synthesis the analyst feels confident handing off to confirmatory analysis or stakeholder communication.