2nd/10_Wiki/Topics/AI_and_ML/Pareto-Principle.md

---
id: wiki-2026-0508-pareto-principle
title: Pareto Principle
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [80/20 Rule, Pareto Distribution, Power Law]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [pareto, 80-20, prioritization, decision-making]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: pandas, numpy
---

# Pareto Principle

## 매 한 줄
> **"매 80% of effects from 20% of causes"**. Vilfredo Pareto (1896) — 매 Italy land ownership 의 observation. 매 modern application: bug triage (top 20% bugs cause 80% crashes), customer revenue (top 20% pay 80%), feature importance (top 20% features carry 80% of model signal). 매 prioritization heuristic 의 default.

## 매 핵심

### 매 origin
- Pareto 1896: 매 80% of Italian land owned by 20% of population.
- Juran 1940s: 매 quality control — "vital few vs trivial many".
- 매 Power Law family — log-log linear distribution.
- 매 80/20 의 mnemonic 일 뿐 — 매 actual ratios vary (90/10, 70/30 등).

### 매 핵심 insight
- Effects are NOT uniformly distributed across causes.
- Sorting by impact 의 long tail 발견.
- ROI: 매 fix top 20% causes → solve 80% of problem with 20% of effort.
- Caveat: 매 remaining 20% of effects 매 important 일 수 있음 (safety, compliance).

### 매 software / ML context
- **Bug triage**: 매 small set of bugs causes most crashes.
- **Performance hotspots**: 매 5% of code = 95% of CPU time.
- **Feature importance**: 매 top features dominate model signal.
- **Customer revenue**: 매 enterprise tail tiny number of users.
- **Test coverage**: 매 80% of bugs in 20% of code paths.

### 매 응용
1. Backlog prioritization (impact × ease).
2. Performance profiling (optimize hot path first).
3. Feature engineering (drop low-importance features).
4. Customer success (focus on high-value accounts).
5. Bug fixing (top crash signatures first).

## 💻 패턴

### Pareto chart for bug triage
```python
import pandas as pd
import matplotlib.pyplot as plt

df = pd.DataFrame({"bug": ["A","B","C","D","E","F"],
                   "occurrences": [500, 300, 100, 50, 30, 20]})
df = df.sort_values("occurrences", ascending=False)
df["cum_pct"] = df["occurrences"].cumsum() / df["occurrences"].sum() * 100

fig, ax1 = plt.subplots()
ax1.bar(df["bug"], df["occurrences"])
ax2 = ax1.twinx()
ax2.plot(df["bug"], df["cum_pct"], "r-o")
ax2.axhline(80, color="gray", linestyle="--")
plt.show()
```

### Find the "vital 20%"
```python
def vital_few(values, threshold=0.8):
    sorted_vals = sorted(values, reverse=True)
    cumsum = 0
    total = sum(sorted_vals)
    for i, v in enumerate(sorted_vals, 1):
        cumsum += v
        if cumsum / total >= threshold:
            return i, sorted_vals[:i]
    return len(values), sorted_vals
```

### Feature importance pruning
```python
import xgboost as xgb
model = xgb.XGBClassifier().fit(X, y)
importance = pd.Series(model.feature_importances_, index=X.columns).sort_values(ascending=False)
cumulative = importance.cumsum() / importance.sum()
top_features = importance[cumulative <= 0.8].index  # vital few
print(f"{len(top_features)} of {len(X.columns)} features carry 80% importance")
```

### Revenue concentration analysis
```python
customers = pd.read_csv("customers.csv").sort_values("revenue", ascending=False)
customers["cum_revenue_pct"] = customers["revenue"].cumsum() / customers["revenue"].sum()
top_20 = customers.head(int(len(customers) * 0.2))
print(f"Top 20% generate {top_20['revenue'].sum() / customers['revenue'].sum():.0%}")
```

### Profiling hot path (Python)
```python
import cProfile, pstats
profiler = cProfile.Profile()
profiler.enable()
run_workload()
profiler.disable()
stats = pstats.Stats(profiler).sort_stats("cumulative")
stats.print_stats(20)  # top 20 functions usually = 80%+ time
```

### LLM cost: top tokens
```python
# 매 prompt token spend tracking
from collections import Counter
spend = Counter()
for log in logs:
    spend[log["prompt_template"]] += log["tokens"] * log["cost_per_token"]
total = sum(spend.values())
running = 0
for template, cost in spend.most_common():
    running += cost
    print(f"{template}: ${cost:.2f}, cumulative {running/total:.0%}")
    if running/total > 0.8: break
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Backlog overload | Pareto-rank by impact, ship top 20% |
| Slow application | Profile, fix hot path 먼저 |
| Too many features | Importance-based pruning |
| Customer support | Tier by revenue, allocate AE coverage |
| Long bug list | Triage by frequency × severity |
| Compliance / safety | Pareto NOT applicable (매 100% 필수) |

**기본값**: 매 sort by impact, take top until cumulative ≥ 80%.

## 🔗 Graph
- 부모: [[Power-Law]]
- 변형: [[80-20-Rule]] · [[Long-Tail]]
- 응용: [[Feature-Importance]]

## 🤖 LLM 활용
**언제**: 매 backlog prioritization, optimization scope, feature selection, customer segmentation.
**언제 X**: 매 safety-critical / compliance — long tail 매 ignore 불가.

## ❌ 안티패턴
- **Treating 80/20 literally**: 매 actual ratio varies — measure, don't assume.
- **Ignoring long tail entirely**: 매 some long-tail items high-leverage (zero-day, churn-risk customer).
- **Cause/effect confusion**: 매 20% of features cause 80% of accuracy ≠ keep only those (interactions matter).
- **Static analysis**: 매 Pareto re-ranks over time — 매 weekly recompute.
- **Pareto in safety domain**: 매 medical, finance, security — 매 100% coverage 필수.

## 🧪 검증 / 중복
- Verified (Pareto 1896 Cours d'économie politique, Juran 1951 Quality Handbook).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — Pareto applications, charts, anti-patterns |