2nd/10_Wiki/Topics/AI_and_ML/Term-Frequency-Inverse-Document-Frequency.md

---
id: wiki-2026-0508-term-frequency-inverse-document-
title: Term Frequency-Inverse Document Frequency
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [TF-IDF, tfidf, classic IR baseline]
duplicate_of: none
source_trust_level: A
confidence_score: 0.95
verification_status: applied
tags: [ir, nlp, retrieval, baseline, sklearn]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: scikit-learn
---

# Term Frequency-Inverse Document Frequency

## 매 한 줄
> **"매 term frequency × inverse document frequency — 매 word's importance 의 corpus context 매 weight"**. Karen Spärck Jones (1972) 의 IDF formalization. 2026 매 dense retrieval (BGE, E5) 매 default 매도 매 baseline + hybrid (BM25 + dense) 의 second stage 매 still ubiquitous.

## 매 핵심

### 매 Formula
- **TF**: term의 doc 매 count (raw / log-normalized / frequency).
- **IDF**: `log(N / df_t)` — 매 N corpus size, df_t = doc 매 t의 contains 의 count.
- **TF-IDF**: TF(t,d) × IDF(t).
- **L2 norm**: 매 cosine 의 prepare.

### 매 Variants
- Raw TF / log(1+TF) / sublinear.
- IDF smoothing: `log((1+N)/(1+df)) + 1`.
- BM25: 매 TF saturation + length normalization 의 add.

### 매 응용
1. Search baseline (sklearn + scikit-learn).
2. Hybrid retrieval — 매 BM25 + dense embedding의 reciprocal-rank fuse.
3. Feature extraction 매 classical ML (logistic regression, SVM).
4. Keyword extraction (top-k tfidf).

## 💻 패턴

### sklearn TF-IDF
```python
from sklearn.feature_extraction.text import TfidfVectorizer

corpus = [
    "the cat sat on the mat",
    "the dog ate the bone",
    "cats and dogs are pets",
]
vec = TfidfVectorizer(stop_words="english", sublinear_tf=True, ngram_range=(1, 2))
X = vec.fit_transform(corpus)  # sparse (n_docs, n_features)
print(vec.get_feature_names_out())
```

### Cosine search
```python
from sklearn.metrics.pairwise import cosine_similarity

q = vec.transform(["pet animals"])
sims = cosine_similarity(q, X).flatten()
ranking = sims.argsort()[::-1]
```

### Manual IDF (educational)
```python
import math
from collections import Counter

def compute_idf(corpus_tokens):
    N = len(corpus_tokens)
    df = Counter()
    for tokens in corpus_tokens:
        for t in set(tokens):
            df[t] += 1
    return {t: math.log((N + 1) / (df_t + 1)) + 1 for t, df_t in df.items()}
```

### BM25 (preferred over plain TF-IDF for IR)
```python
from rank_bm25 import BM25Okapi

tokenized = [doc.lower().split() for doc in corpus]
bm25 = BM25Okapi(tokenized, k1=1.5, b=0.75)
scores = bm25.get_scores("pet animals".split())
```

### Hybrid search (2026 standard)
```python
import numpy as np

def rrf(rankings, k=60):
    scores = {}
    for ranking in rankings:
        for rank, doc_id in enumerate(ranking):
            scores[doc_id] = scores.get(doc_id, 0) + 1 / (k + rank)
    return sorted(scores, key=scores.get, reverse=True)

bm25_top = bm25.get_top_n("query".split(), corpus, n=100)
dense_top = dense_index.search("query", k=100)
final = rrf([bm25_top, dense_top])[:10]
```

### Top keyword extraction
```python
def top_keywords(doc_idx, vec, X, k=10):
    row = X[doc_idx].toarray().flatten()
    feats = vec.get_feature_names_out()
    top = np.argsort(-row)[:k]
    return [(feats[i], row[i]) for i in top]
```

### Persistence
```python
import joblib
joblib.dump((vec, X), "tfidf_index.joblib")
vec, X = joblib.load("tfidf_index.joblib")
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| 매 small corpus + interpretability | TF-IDF (sklearn) |
| 매 medium corpus + better recall | BM25 |
| 매 semantic / paraphrase | Dense (BGE-M3, E5) |
| 매 production search | Hybrid (BM25 + dense + RRF) |
| 매 keyword extraction / explanation | Plain TF-IDF top-k |

**기본값**: 매 BM25 baseline → 매 hybrid + reranker (cross-encoder) for 2026 production.

## 🔗 Graph
- 부모: [[Information Retrieval]]
- 변형: [[BM25]]
- 응용: [[Search Engine]] · [[RAG]]
- Adjacent: [[Dense Retrieval]]

## 🤖 LLM 활용
**언제**: 매 small corpus 매 lookup, 매 RAG 의 sparse channel, 매 explainability ("matched on 'mat', 'cat'").
**언제 X**: 매 paraphrase / multilingual 매 weak — 매 dense 의 prefer.

## ❌ 안티패턴
- **TF-IDF 만으로 production search**: 매 paraphrase miss.
- **No stopword / lowercasing**: 매 noisy features.
- **Same vectorizer not pickled**: 매 train/serve mismatch.
- **No length normalization**: 매 long docs 의 unfair advantage (use BM25 또는 normalize).

## 🧪 검증 / 중복
- Verified (Spärck Jones 1972; Manning IR Book Ch.6; sklearn TfidfVectorizer 2026).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — TF-IDF formula + sklearn + BM25 + hybrid RRF |