exploratory-data-analysis

Exploratory Data Analysis (EDA)

Overview

EDA is the disciplined first pass over a dataset: understand shape, types, distributions, missingness, relationships, and red flags before writing a single model. Skipping it is the #1 cause of silent modeling failures (leakage, broken splits, garbage features).

When to use

• A new dataset just landed.
• You're asked to "look at", "profile", "explore", or "summarize" data.
• A model underperforms and you need to understand the inputs.

Workflow

Follow this order. Do not jump to modeling until every step is answered.

1. Shape & types — rows, columns, dtypes, memory. Are numeric columns actually numeric?
2. Missingness — per-column null counts and patterns (MCAR/MAR/MNAR). Is missingness itself predictive?
3. Target analysis — distribution of the target (class balance / skew). This decides metrics and resampling.
4. Univariate — distributions of each feature (histograms, value counts, describe()).
5. Bivariate — feature vs target relationships; correlation matrix for numeric.
6. Leakage scan — features too perfectly correlated with the target, IDs, timestamps, or post-outcome columns.
7. Cardinality & outliers — high-cardinality categoricals, extreme values.

Reference snippet

import pandas as pd

df = pd.read_csv("data.csv")

# 1. Shape & types
print(df.shape)
print(df.dtypes.value_counts())
print(df.memory_usage(deep=True).sum() / 1e6, "MB")

# 2. Missingness
miss = df.isna().mean().sort_values(ascending=False)
print(miss[miss > 0])

# 3. Target (classification example)
print(df["target"].value_counts(normalize=True))

# 4-5. Numeric summary + correlations with target
num = df.select_dtypes("number")
print(num.describe().T)
print(num.corr()["target"].sort_values(ascending=False))

# 6. Leakage red flag: |corr| ~ 1.0 with target
corr_t = num.corr()["target"].drop("target").abs()
print("LEAKAGE SUSPECTS:", corr_t[corr_t > 0.95].index.tolist())

For a fast automated first look, ydata-profiling (ProfileReport(df)) or df.describe(include="all") is acceptable — but never let a tool replace the manual leakage scan.