Data Visualization for Data Scientists: From Question to Plot

A practical, Python-first guide to common Data Science plots with simulated data, reproducible code and interactive examples.

📊 A good plot is not decoration. It is a question made visible.

Choose the question first, the plot second, and the library third.

Every worked example follows one contract:

QUESTION → DATA SIMULATOR → PLOT CODE → VISIBLE CHART → INTERPRETATION

The chart is always visible. View Data Simulator and View Plot Code expand and collapse independently. This lets a beginner focus on the picture first and inspect reproducible Python when ready.

TipWhat you will learn

Distribution, comparison, relationship, time-series, experimentation, multivariate, model-diagnostic, geographic and interactive visualization using Matplotlib, Seaborn, Plotly and Bokeh, plus guidance on modern alternatives.

1. Visualization Starts with the Question

🔍 Distribution
Histogram · KDE · ECDF · Box · Violin

⚖️ Comparison
Bar · Box · Error bars · Small multiples

🔗 Relationship
Scatter · Regression · Hexbin

⏱️ Time
Line · Rolling trend · Uncertainty band

🤖 Model
Confusion · ROC · PR · Calibration · Residuals

🌍 Location / Flow
Maps · Funnels

2. The Python Visualization Ecosystem

Matplotlib is the low-level foundation and is commonly static-first. Seaborn builds statistical graphics on Matplotlib. Plotly and Bokeh make browser interaction—hover, zoom, selection and widgets—a central strength. Matplotlib can also use interactive backends, so “Matplotlib is never interactive” is not technically correct.

2.1. Matplotlib — Precise Figure Control

Company question: How is weekly revenue changing?

import numpy as np
import pandas as pd
rng=np.random.default_rng(1)
mat_df=pd.DataFrame({"week":np.arange(1,21)})
mat_df["revenue"]=85000+1800*mat_df["week"]+rng.normal(0,4500,20)

import matplotlib.pyplot as plt
fig,ax=plt.subplots(figsize=(8,4))
ax.plot(mat_df["week"],mat_df["revenue"],marker="o")
ax.set(title="Weekly Revenue",xlabel="Week",ylabel="Revenue ($)")
fig.tight_layout(); plt.show()

Best for: custom static analytical figures, reports, annotations and fine control.

2.2. Seaborn — Statistical Visualization

Company question: How does revenue differ across segments?

import numpy as np
import pandas as pd
rng=np.random.default_rng(2)
seg=rng.choice(["Consumer","SMB","Enterprise"],900,p=[.55,.3,.15])
base=pd.Series(seg).map({"Consumer":70,"SMB":125,"Enterprise":220}).to_numpy()
sns_df=pd.DataFrame({"segment":seg,"revenue":np.maximum(5,base+rng.normal(0,32,900))})

import matplotlib.pyplot as plt
import seaborn as sns
fig,ax=plt.subplots(figsize=(7.5,4))
sns.boxplot(data=sns_df,x="segment",y="revenue",ax=ax)
ax.set(title="Monthly Revenue by Segment",ylabel="Revenue ($)")
fig.tight_layout(); plt.show()

Best for: EDA, grouped distributions, statistical comparisons and regression views.

2.3. Plotly — Interactive Exploration

Company question: Which individual customers sit behind the overall pattern?

import numpy as np
import pandas as pd
rng=np.random.default_rng(3); n=650
px_df=pd.DataFrame({"customer_id":np.arange(10001,10001+n),"segment":rng.choice(["Consumer","SMB","Enterprise"],n,p=[.55,.3,.15]),"sessions":rng.poisson(9,n)+1,"tickets":rng.poisson(1.5,n)})
px_df["revenue"]=25+5.5*px_df["sessions"]+px_df["segment"].map({"Consumer":0,"SMB":40,"Enterprise":110})+rng.normal(0,25,n)

import plotly.express as px
fig=px.scatter(px_df,x="sessions",y="revenue",color="segment",size="tickets",hover_data=["customer_id"],title="Interactive Customer Explorer")
fig.show()

Hover, zoom, pan, select and toggle legend groups directly in the rendered chart.

2.4. Bokeh — Interactive Browser Plots

Company question: Can an analyst inspect and zoom into customers?

import numpy as np
import pandas as pd
rng=np.random.default_rng(4); n=500
bk_df=pd.DataFrame({"customer_id":np.arange(1,n+1),"sessions":rng.poisson(10,n)+1})
bk_df["revenue"]=(35+6*bk_df["sessions"]+rng.normal(0,30,n)).clip(5)

from bokeh.plotting import figure,show
from bokeh.models import ColumnDataSource,HoverTool
source=ColumnDataSource(bk_df)
p=figure(width=800,height=380,title="Bokeh: Revenue vs Sessions",tools="pan,wheel_zoom,box_zoom,reset")
p.scatter(x="sessions",y="revenue",source=source,size=7,alpha=.45)
p.add_tools(HoverTool(tooltips=[("Customer","@customer_id"),("Revenue","@revenue{0.0}")]))
show(p)

Bokeh is especially useful for linked selections, widgets and browser applications.

3. Histogram — Distribution Shape

Company question: What does monthly customer revenue look like?

import numpy as np
import pandas as pd
rng=np.random.default_rng(10)
hist_df=pd.DataFrame({"revenue":rng.lognormal(4.35,.62,2500)})

import matplotlib.pyplot as plt
fig,ax=plt.subplots(figsize=(8,4)); ax.hist(hist_df["revenue"],bins=38,edgecolor="white")
ax.set(title="Distribution of Monthly Revenue",xlabel="Revenue ($)",ylabel="Customers"); fig.tight_layout(); plt.show()

Bin width matters. Too few bins hide structure; too many emphasize noise.

4. KDE — Smooth Distribution Comparison

Company question: Do segments have differently shaped revenue distributions?

import numpy as np
import pandas as pd
rng=np.random.default_rng(11)
kde_df=pd.concat([pd.DataFrame({"segment":s,"revenue":rng.lognormal(mu,sd,700)}) for s,mu,sd in [("Consumer",4,.55),("SMB",4.55,.48),("Enterprise",5.25,.40)]],ignore_index=True)

import matplotlib.pyplot as plt
import seaborn as sns
fig,ax=plt.subplots(figsize=(8,4)); sns.kdeplot(data=kde_df,x="revenue",hue="segment",common_norm=False,ax=ax)
ax.set(title="Revenue Density by Segment",xlabel="Revenue ($)"); fig.tight_layout(); plt.show()

KDE is estimated smoothness. Do not interpret every bump as a real population mode.

5. ECDF — Cumulative Distribution

Company question: What fraction of each segment falls below a revenue threshold?

import numpy as np
import pandas as pd
rng=np.random.default_rng(12)
ecdf_df=pd.concat([pd.DataFrame({"segment":s,"revenue":np.maximum(5,rng.normal(c,c*.28,600))}) for s,c in [("Consumer",65),("SMB",125),("Enterprise",225)]],ignore_index=True)

import matplotlib.pyplot as plt
import seaborn as sns
fig,ax=plt.subplots(figsize=(8,4)); sns.ecdfplot(data=ecdf_df,x="revenue",hue="segment",ax=ax)
ax.set(title="ECDF of Revenue by Segment",xlabel="Revenue ($)",ylabel="Fraction ≤ x"); fig.tight_layout(); plt.show()

The ECDF has no histogram-bin choice and makes percentile questions direct.

6. Box Plot — Compact Group Comparison

Company question: Which segment has the widest revenue distribution?

import numpy as np
import pandas as pd
rng=np.random.default_rng(13)
box_df=pd.DataFrame({"segment":np.repeat(["Consumer","SMB","Enterprise"],500)})
c=box_df["segment"].map({"Consumer":70,"SMB":130,"Enterprise":230}).to_numpy(); box_df["revenue"]=np.maximum(5,c+rng.normal(0,c*.25))

import matplotlib.pyplot as plt
import seaborn as sns
fig,ax=plt.subplots(figsize=(7.5,4)); sns.boxplot(data=box_df,x="segment",y="revenue",ax=ax)
ax.set(title="Revenue by Segment",ylabel="Revenue ($)"); fig.tight_layout(); plt.show()

Box plots summarize median, quartiles, spread and potential outliers.

7. Violin + Strip — Shape and Observations

Company question: Do plans differ in the whole engagement distribution?

import numpy as np
import pandas as pd
rng=np.random.default_rng(14)
violin_df=pd.DataFrame({"plan":np.repeat(["Standard","Premium"],220),"engagement":np.r_[rng.beta(2.2,4,220),rng.beta(4,2.4,220)]})

import matplotlib.pyplot as plt
import seaborn as sns
fig,ax=plt.subplots(figsize=(7,4)); sns.violinplot(data=violin_df,x="plan",y="engagement",inner=None,ax=ax); sns.stripplot(data=violin_df,x="plan",y="engagement",alpha=.28,size=2.5,ax=ax)
ax.set_title("Engagement Distribution by Plan"); fig.tight_layout(); plt.show()

Use violin plots when distribution shape itself matters.

8. Bar Plot — Categorical Summary

Company question: Which segment has the highest churn rate?

import numpy as np
import pandas as pd
rng=np.random.default_rng(15); n=2400
seg=rng.choice(["Consumer","SMB","Enterprise"],n,p=[.55,.3,.15]); prob=pd.Series(seg).map({"Consumer":.19,"SMB":.13,"Enterprise":.08}).to_numpy(); raw=pd.DataFrame({"segment":seg,"churn":rng.binomial(1,prob)}); bar_df=raw.groupby("segment",as_index=False).agg(churn_rate=("churn","mean"))

import matplotlib.pyplot as plt
fig,ax=plt.subplots(figsize=(7,4)); ax.bar(bar_df["segment"],bar_df["churn_rate"]); ax.set(title="Churn Rate by Segment",ylabel="Churn rate",ylim=(0,bar_df["churn_rate"].max()*1.25)); fig.tight_layout(); plt.show()

Bar lengths need an honest baseline. Truncating the y-axis can exaggerate differences.

9. Count Plot — Group Size

Company question: How many customers use each device?

import numpy as np
import pandas as pd
rng=np.random.default_rng(16)
count_df=pd.DataFrame({"device":rng.choice(["Web","iOS","Android"],1800,p=[.46,.28,.26]),"segment":rng.choice(["Consumer","SMB","Enterprise"],1800,p=[.56,.3,.14])})

import matplotlib.pyplot as plt
import seaborn as sns
fig,ax=plt.subplots(figsize=(8,4)); sns.countplot(data=count_df,x="device",hue="segment",ax=ax); ax.set(title="Customer Count by Device and Segment",ylabel="Customers"); fig.tight_layout(); plt.show()

Inspect sample size before interpreting subgroup metrics.

10. Scatter Plot — Numerical Relationship

Company question: Do more sessions correspond to more revenue?

import numpy as np
import pandas as pd
rng=np.random.default_rng(17); n=1300
scatter_df=pd.DataFrame({"sessions":rng.poisson(10,n)+1}); scatter_df["revenue"]=(30+6.2*scatter_df["sessions"]+rng.normal(0,32,n)).clip(5)

import matplotlib.pyplot as plt
fig,ax=plt.subplots(figsize=(8,4.5)); ax.scatter(scatter_df["sessions"],scatter_df["revenue"],alpha=.25,s=18); ax.set(title="Sessions vs Monthly Revenue",xlabel="Sessions",ylabel="Revenue ($)"); fig.tight_layout(); plt.show()

Look for direction, curvature, clusters, changing variance and outliers.

11. Regression Plot — Average Trend

Company question: What average relationship is visible?

import numpy as np
import pandas as pd
rng=np.random.default_rng(18); n=900
reg_df=pd.DataFrame({"sessions":rng.uniform(1,25,n)}); reg_df["revenue"]=25+5.8*reg_df["sessions"]+rng.normal(0,28,n)

import matplotlib.pyplot as plt
import seaborn as sns
fig,ax=plt.subplots(figsize=(8,4.5)); sns.regplot(data=reg_df,x="sessions",y="revenue",scatter_kws={"alpha":.18,"s":18},ax=ax); ax.set_title("Average Revenue Trend with Sessions"); fig.tight_layout(); plt.show()

Warning

A fitted trend summarizes association, not causation.

12. Hexbin — Dense Relationships

Company question: How do two telemetry measures relate across 75,000 events?

import numpy as np
rng=np.random.default_rng(19); n=75000
hex_x=rng.normal(size=n); hex_y=.65*hex_x+rng.normal(0,.75,n)

import matplotlib.pyplot as plt
fig,ax=plt.subplots(figsize=(7,5)); hb=ax.hexbin(hex_x,hex_y,gridsize=45,mincnt=1); fig.colorbar(hb,ax=ax,label="Observations"); ax.set(title="Density Hidden by Overplotting",xlabel="Feature X",ylabel="Feature Y"); fig.tight_layout(); plt.show()

Use hexbin, density, sampling or rasterization when scatter points become an opaque cloud.

13. Line Plot — Change Over Time

Company question: Is revenue rising beneath daily noise?

import numpy as np
import pandas as pd
rng=np.random.default_rng(20); dates=pd.date_range("2026-01-01",periods=220,freq="D"); t=np.arange(220)
line_df=pd.DataFrame({"date":dates,"revenue":120000+180*t+10000*np.sin(2*np.pi*t/7)+rng.normal(0,6500,220)}); line_df["rolling_14d"]=line_df["revenue"].rolling(14).mean()

import matplotlib.pyplot as plt
fig,ax=plt.subplots(figsize=(10,4)); ax.plot(line_df["date"],line_df["revenue"],alpha=.35,label="Daily"); ax.plot(line_df["date"],line_df["rolling_14d"],linewidth=2,label="14-day mean"); ax.set(title="Revenue: Noise vs Trend",ylabel="Revenue ($)"); ax.legend(); fig.tight_layout(); plt.show()

Chronology is information. Keep time on an ordered axis.

14. Uncertainty Band — Show What We Do Not Know

Company question: Is a conversion trend larger than its uncertainty?

import numpy as np
import pandas as pd
rng=np.random.default_rng(21); w=np.arange(1,21); est=.12+.002*w+rng.normal(0,.004,20); se=np.linspace(.012,.006,20)
band_df=pd.DataFrame({"week":w,"estimate":est,"lower":est-1.96*se,"upper":est+1.96*se})

import matplotlib.pyplot as plt
fig,ax=plt.subplots(figsize=(8,4)); ax.plot(band_df["week"],band_df["estimate"],linewidth=2); ax.fill_between(band_df["week"],band_df["lower"],band_df["upper"],alpha=.2); ax.set(title="Conversion Estimate with 95% Band",xlabel="Week",ylabel="Conversion rate"); fig.tight_layout(); plt.show()

Use uncertainty bands for forecasts, experiments and estimated quantities.

15. Error Bars — Experiment Estimates

Company question: How do A/B variants compare with uncertainty?

import numpy as np
import pandas as pd
exp_df=pd.DataFrame({"variant":["Control","Treatment A","Treatment B"],"conversion":[.100,.112,.107],"n":[12000,11850,12120]}); exp_df["se"]=np.sqrt(exp_df["conversion"]*(1-exp_df["conversion"])/exp_df["n"]); exp_df["ci95"]=1.96*exp_df["se"]

import matplotlib.pyplot as plt
fig,ax=plt.subplots(figsize=(7,4)); ax.errorbar(exp_df["variant"],exp_df["conversion"],yerr=exp_df["ci95"],fmt="o",capsize=5); ax.set(title="Conversion ± 95% Approximate CI",ylabel="Conversion rate"); fig.tight_layout(); plt.show()

Effect magnitude and uncertainty belong in the same comparison.

16. Correlation Heatmap — Matrix Pattern

Company question: Which numerical variables move together?

import numpy as np
import pandas as pd
rng=np.random.default_rng(22); n=1200
heat_df=pd.DataFrame({"sessions":rng.poisson(10,n)+1,"tenure":rng.integers(1,72,n),"tickets":rng.poisson(1.5,n)}); heat_df["revenue"]=25+6*heat_df["sessions"]+.6*heat_df["tenure"]+rng.normal(0,30,n); heat_df["satisfaction"]=np.clip(4.5-.28*heat_df["tickets"]+rng.normal(0,.5,n),1,5); corr=heat_df.corr(numeric_only=True)

import matplotlib.pyplot as plt
import seaborn as sns
fig,ax=plt.subplots(figsize=(7,5)); sns.heatmap(corr,annot=True,fmt=".2f",center=0,ax=ax); ax.set_title("Correlation Matrix"); fig.tight_layout(); plt.show()

Correlation is a screening tool, not a causal map.

17. Pair Plot — Multivariate First Look

Company question: How do several customer measures relate at once?

import numpy as np
import pandas as pd
rng=np.random.default_rng(23); n=450
pair_df=pd.DataFrame({"segment":rng.choice(["Consumer","SMB","Enterprise"],n,p=[.55,.3,.15]),"sessions":rng.poisson(9,n)+1,"tenure":rng.integers(1,72,n),"satisfaction":np.clip(rng.normal(4,.65,n),1,5)}); pair_df["revenue"]=30+5.5*pair_df["sessions"]+pair_df["segment"].map({"Consumer":0,"SMB":40,"Enterprise":105})+rng.normal(0,25,n)

import matplotlib.pyplot as plt
import seaborn as sns
g=sns.pairplot(pair_df,vars=["sessions","revenue","tenure","satisfaction"],hue="segment",corner=True); g.fig.suptitle("Multivariate Customer Overview",y=1.02); plt.show()

Pair plots are excellent for early EDA but become unwieldy as the number of variables grows.

18. Faceting — Small Multiples

Company question: Does a relationship change by segment?

import numpy as np
import pandas as pd
rng=np.random.default_rng(24); n=1000
facet_df=pd.DataFrame({"segment":rng.choice(["Consumer","SMB","Enterprise"],n,p=[.55,.3,.15]),"device":rng.choice(["Web","iOS","Android"],n),"sessions":rng.poisson(9,n)+1}); facet_df["revenue"]=25+5.8*facet_df["sessions"]+facet_df["segment"].map({"Consumer":0,"SMB":45,"Enterprise":110})+rng.normal(0,30,n)

import matplotlib.pyplot as plt
import seaborn as sns
g=sns.relplot(data=facet_df,x="sessions",y="revenue",col="segment",hue="device",kind="scatter",alpha=.45,height=3.3); g.fig.suptitle("Sessions vs Revenue Across Segments",y=1.05); plt.show()

Small multiples often beat one overloaded plot with too many visual encodings.

19. Interactive Time Series — Zoom into Events

Company question: Can an analyst inspect a long timeline around an incident?

import numpy as np
import pandas as pd
rng=np.random.default_rng(27); dates=pd.date_range("2025-01-01",periods=420,freq="D"); t=np.arange(420)
its_df=pd.DataFrame({"date":dates,"revenue":100000+120*t+8000*np.sin(2*np.pi*t/7)+rng.normal(0,5500,420)}); its_df["rolling_21d"]=its_df["revenue"].rolling(21).mean()

import plotly.express as px
fig=px.line(its_df,x="date",y=["revenue","rolling_21d"],title="Interactive Revenue Timeline"); fig.update_xaxes(rangeslider_visible=True); fig.show()

The range slider and zoom controls are valuable for campaigns, incidents and long monitoring windows.

20. Confusion Matrix — Classification Errors

Company question: Is the churn model missing positives or generating false alarms?

import numpy as np
rng=np.random.default_rng(30); n=2500
cm_y=rng.binomial(1,.18,n); cm_score=np.clip(.12+.55*cm_y+rng.normal(0,.22,n),0,1); cm_pred=(cm_score>=.5).astype(int)

import matplotlib.pyplot as plt
from sklearn.metrics import confusion_matrix,ConfusionMatrixDisplay
cm=confusion_matrix(cm_y,cm_pred); ConfusionMatrixDisplay(cm,display_labels=["No churn","Churn"]).plot(); plt.title("Churn Classifier: Confusion Matrix"); plt.show()

A confusion matrix shows which kind of error occurs, not merely how often the model is correct.

21. ROC Curve — Ranking Across Thresholds

Company question: How well does the classifier separate classes?

import numpy as np
rng=np.random.default_rng(31); n=2800
roc_y=rng.binomial(1,.20,n); roc_score=np.clip(.10+.60*roc_y+rng.normal(0,.24,n),0,1)

import matplotlib.pyplot as plt
from sklearn.metrics import roc_curve,roc_auc_score
fpr,tpr,_=roc_curve(roc_y,roc_score); auc=roc_auc_score(roc_y,roc_score); fig,ax=plt.subplots(figsize=(5.5,5)); ax.plot(fpr,tpr,label=f"AUC={auc:.3f}"); ax.plot([0,1],[0,1],linestyle="--"); ax.set(title="ROC Curve",xlabel="False Positive Rate",ylabel="True Positive Rate"); ax.legend(); plt.show()

ROC summarizes threshold-dependent ranking behavior.

22. Precision–Recall Curve — Rare Positive Classes

Company question: What tradeoff is available when positive cases are scarce?

import numpy as np
rng=np.random.default_rng(32); n=3500
pr_y=rng.binomial(1,.08,n); pr_score=np.clip(.05+.58*pr_y+rng.normal(0,.23,n),0,1)

import matplotlib.pyplot as plt
from sklearn.metrics import precision_recall_curve,average_precision_score
p,r,_=precision_recall_curve(pr_y,pr_score); ap=average_precision_score(pr_y,pr_score); fig,ax=plt.subplots(figsize=(5.5,5)); ax.plot(r,p,label=f"AP={ap:.3f}"); ax.set(title="Precision–Recall Curve",xlabel="Recall",ylabel="Precision"); ax.legend(); plt.show()

Often more operationally revealing than accuracy for rare events such as fraud.

23. Calibration — Can We Trust 0.8?

Company question: Do predicted probabilities correspond to observed frequencies?

import numpy as np
rng=np.random.default_rng(33); n=4500
x=rng.normal(size=n); true_p=1/(1+np.exp(-1.3*x)); cal_y=rng.binomial(1,true_p); cal_score=np.clip(1/(1+np.exp(-.75*x)),.001,.999)

import matplotlib.pyplot as plt
from sklearn.calibration import calibration_curve
obs,pred=calibration_curve(cal_y,cal_score,n_bins=10); fig,ax=plt.subplots(figsize=(5.5,5)); ax.plot(pred,obs,marker="o"); ax.plot([0,1],[0,1],linestyle="--"); ax.set(title="Calibration Plot",xlabel="Mean predicted probability",ylabel="Observed positive rate"); plt.show()

Ranking quality and probability calibration are different properties.

24. Residual Plot — Regression Diagnostics

Company question: Does prediction error change systematically?

import numpy as np
import pandas as pd
rng=np.random.default_rng(34); n=1200
actual=rng.normal(100,25,n); predicted=actual+.10*(actual-100)+rng.normal(0,12,n); resid_df=pd.DataFrame({"predicted":predicted,"residual":actual-predicted})

import matplotlib.pyplot as plt
fig,ax=plt.subplots(figsize=(7,4.5)); ax.scatter(resid_df["predicted"],resid_df["residual"],alpha=.28); ax.axhline(0,linestyle="--"); ax.set(title="Residual Plot",xlabel="Predicted",ylabel="Actual − Predicted"); fig.tight_layout(); plt.show()

Residuals can reveal bias, nonlinearity and changing variance hidden by a single RMSE.

25. Feature Importance — What Does the Model Rely On?

Company question: Which inputs contribute most under this importance definition?

import pandas as pd
imp_df=pd.DataFrame({"feature":["last_login_days","support_tickets","tenure","monthly_spend","sessions"],"importance":[.31,.25,.19,.14,.11]}).sort_values("importance")

import matplotlib.pyplot as plt
fig,ax=plt.subplots(figsize=(7,4)); ax.barh(imp_df["feature"],imp_df["importance"]); ax.set(title="Model Feature Importance",xlabel="Importance"); fig.tight_layout(); plt.show()

Importance is not causality. A predictive feature can be a proxy, consequence, correlated signal or leakage.

26. Funnel — Where Do Users Drop Out?

Company question: At which purchase stage is the largest loss?

import pandas as pd
funnel_df=pd.DataFrame({"stage":["Visitors","Product View","Cart","Checkout","Purchase"],"users":[100000,62000,21000,12000,9200]})

import plotly.express as px
fig=px.funnel(funnel_df,x="users",y="stage",title="Purchase Funnel"); fig.show()

Common in acquisition, onboarding, checkout and application workflows.

27. Geographic Plot — When Location Matters

Company question: Where are orders concentrated around a service area?

import numpy as np
import pandas as pd
rng=np.random.default_rng(35); n=500
geo_df=pd.DataFrame({"latitude":30.27+rng.normal(0,.055,n),"longitude":-97.74+rng.normal(0,.065,n),"orders":rng.poisson(5,n)+1})

import plotly.express as px
fig=px.scatter_map(geo_df,lat="latitude",lon="longitude",size="orders",hover_data=["orders"],zoom=9,height=500,center={"lat": 30.2672, "lon": -97.7431},title="Interactive Geographic Distribution of Orders"); fig.show()

Use a map when spatial relationships matter. If the question is only “which region is largest?”, a sorted bar chart may be clearer.

28. Other Modern Visualization Libraries

• Altair — declarative grammar-of-graphics and concise interactive statistical charts.
• hvPlot / HoloViews — high-level interactive plotting for DataFrame/xarray workflows.
• Datashader — rasterizes/aggregates huge datasets rather than drawing every point.
• PyDeck / deck.gl ecosystem — rich large-scale geospatial visualization.

The state of the art is not “use the newest library.” Match the tool to data volume + analytical question + audience + interaction + deployment environment.

29. Which Library Should I Reach For?

Need	Good starting choice
Custom static figure	Matplotlib
Statistical EDA	Seaborn
Publication/report control	Matplotlib
Interactive exploration	Plotly
Browser widgets / linked interaction	Bokeh
Declarative grammar	Altair
Huge point clouds	Datashader ecosystem
Interactive geospatial work	Plotly / PyDeck

30. Visualization Mistakes That Mislead

• Truncated bar axes can exaggerate small differences.
• Too many categories turn legends into puzzles.
• Dual y-axes can manufacture apparent relationships through scaling.
• Many-slice pie charts make precise comparison difficult.
• Rainbow scales can create false perceptual boundaries.
• Lines between unrelated categories imply continuity that does not exist.
• Means without uncertainty/sample size can hide noise.
• Overplotting calls for transparency, sampling, hexbin, density or rasterization.

31. Exploratory vs Explanatory Visualization

Exploratory: “What is happening?” → many views, interaction, diagnostics.

Explanatory: “Here is what matters.” → fewer views, intentional ordering, annotation, uncertainty, restrained color.

EDA → DISCOVERY → ANALYSIS → DECISION → COMMUNICATION

32. A Data Scientist’s Plot Selector

ONE NUMERIC → Histogram / KDE / ECDF / Box
NUMERIC + CATEGORY → Box / Violin / Bar + CI
TWO NUMERIC → Scatter / Regression / Hexbin
TIME → Line / Rolling trend / Uncertainty band
MATRIX → Heatmap
CLASSIFIER → Confusion / ROC / PR / Calibration
REGRESSION → Actual-vs-predicted / Residuals
LOCATION → Map only when spatial structure matters

33. A Company Visualization Workflow

If Product asks “Why did conversion fall last month?”:

1. Line plot → when did it begin?
2. Facets → which device/segment/region?
3. Funnel → which stage?
4. Distributions → did population mix change?
5. Event annotations → what changed around that date?
6. Statistical analysis → is the change beyond noise?
7. Explanatory chart → communicate the finding.

34. The Visualization Ladder

SEE → COMPARE → RELATE → SEGMENT → DIAGNOSE → EXPLORE → EXPLAIN → SHIP

35. Final Perspective

✨ Choose the question first. Choose the plot second. Choose the library third.

A sophisticated interactive chart cannot rescue the wrong visual encoding. A simple plot can reveal the entire story when it represents the right question.

Summary

• Distribution → histogram, KDE, ECDF, box, violin.
• Comparison → bars, box/violin, error bars, facets.
• Relationship → scatter, regression, hexbin.
• Time → line, rolling trend, uncertainty.
• Classification → confusion, ROC, PR, calibration.
• Regression → residual diagnostics.
• Plotly/Bokeh → browser-native exploration.
• Matplotlib/Seaborn → static-first analytical communication and statistical EDA.
• Every worked chart in this chapter has independently expandable data simulator code and plot code, with the visualization left visible.

Learning Resources

• Matplotlib: https://matplotlib.org/stable/
• Seaborn: https://seaborn.pydata.org/
• Plotly Python: https://plotly.com/python/
• Bokeh: https://docs.bokeh.org/en/latest/