---
id: wiki-2026-0508-sparse-data-handling
title: Sparse Data Handling
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Sparse Matrix, Sparse Features, Missing Data, Imputation]
duplicate_of: none
source_trust_level: A
confidence_score: 0.92
verification_status: applied
tags: [sparse, data-engineering, ml, imputation, scipy]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Python
  framework: scipy/scikit-learn/PyTorch
---

# Sparse Data Handling

## 매 한 줄
> **"매 zero/missing 매 majority 일 때 — storage + algorithm 의 sparse-aware 의 switch"**. Sparse data handling 매 high-cardinality categorical (one-hot, TF-IDF, click-stream) + missing-value imputation + sparse model (L1, factorization machine) 의 수렴. 매 2026 매 scipy.sparse + cuSPARSE / PyTorch sparse + Polars LazyFrame 매 production stack.

## 매 핵심

### 매 2 axes 의 "sparse"
1. **Structurally sparse**: 매 most cell = 0 by definition (one-hot, adjacency, term-doc matrix).
2. **Missing**: 매 NaN/None, 매 not-recorded. 매 imputation needed.
- 매 두 매 different problem. Conflate 의 X.

### 매 sparse format (scipy)
- **CSR** (Compressed Sparse Row): row slice fast, matrix-vector fast. 매 ML 의 default.
- **CSC**: column slice fast. 매 column-wise stat.
- **COO**: build-time. 매 fast construct, slow op.
- **DOK / LIL**: incremental build.
- **BSR**: block-sparse. 매 GPU friendly.

### 매 sparse-aware models
- **Linear w/ L1 (Lasso)**: 매 sparse output. SGD/coordinate-descent.
- **Logistic regression**: liblinear / saga 매 sparse input handle.
- **Tree (XGBoost/LightGBM)**: 매 missing 매 native split direction 의 learn.
- **Factorization Machines** (libfm, xLearn): 매 high-card categorical.
- **HashingVectorizer**: 매 fixed-dim feature hashing.

### 매 missing-value strategies
1. **Drop**: row/col 매 sparse-too-much 일 때.
2. **Constant fill**: 0 / mean / median / mode.
3. **Indicator + fill**: 매 missingness 의 carry as feature.
4. **KNN impute**: 매 small data.
5. **Iterative (MICE)**: 매 chained regression. sklearn `IterativeImputer`.
6. **Tree-native**: 매 LightGBM/XGBoost 매 NaN 의 직접 handle.
7. **Deep**: 매 VAE-imputation, GAIN.

### 매 응용
1. NLP (TF-IDF, BoW).
2. Recommender (user-item interaction).
3. Genomics (one-hot variant matrix).
4. Click-stream / session data.
5. Tabular ML w/ missing fields.

## 💻 패턴

### scipy.sparse construction
```python
from scipy.sparse import csr_matrix, coo_matrix
import numpy as np

# COO build then convert
rows = np.array([0, 1, 2, 0])
cols = np.array([0, 2, 1, 3])
data = np.array([1.0, 2.0, 3.0, 4.0])
M = coo_matrix((data, (rows, cols)), shape=(3, 4)).tocsr()

# Slice / op
print(M.shape, M.nnz, M.density if hasattr(M, 'density') else M.nnz / np.prod(M.shape))
print(M[0])             # row slice
print(M @ M.T)          # sparse @ sparse
```

### sklearn sparse pipeline
```python
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.pipeline import Pipeline

pipe = Pipeline([
    ("tfidf", TfidfVectorizer(max_features=200_000, ngram_range=(1,2))),
    ("clf", LogisticRegression(solver="liblinear", penalty="l1", C=1.0)),
])
pipe.fit(texts_train, y_train)
# tfidf returns CSR; LR liblinear handles sparse natively
```

### HashingVectorizer (no vocab, online)
```python
from sklearn.feature_extraction.text import HashingVectorizer
hv = HashingVectorizer(n_features=2**20, alternate_sign=False)
X = hv.transform(stream_of_docs)  # CSR
```

### Iterative imputation (MICE)
```python
from sklearn.experimental import enable_iterative_imputer  # noqa
from sklearn.impute import IterativeImputer
from sklearn.ensemble import HistGradientBoostingRegressor

imp = IterativeImputer(
    estimator=HistGradientBoostingRegressor(),
    max_iter=10,
    random_state=0,
)
X_filled = imp.fit_transform(X_with_nan)
```

### Missing indicator + fill
```python
from sklearn.compose import ColumnTransformer
from sklearn.impute import SimpleImputer, MissingIndicator
from sklearn.pipeline import FeatureUnion

union = FeatureUnion([
    ("fill_median", SimpleImputer(strategy="median")),
    ("ind",         MissingIndicator(features="all")),
])
X = union.fit_transform(X_raw)
```

### LightGBM native NaN handling
```python
import lightgbm as lgb
# No imputation needed
model = lgb.LGBMClassifier(
    n_estimators=500,
    learning_rate=0.05,
    num_leaves=63,
    use_missing=True,
    zero_as_missing=False,
)
model.fit(X_train_with_nan, y_train)
```

### PyTorch sparse tensor
```python
import torch

i = torch.tensor([[0, 1, 2], [2, 0, 1]])
v = torch.tensor([3.0, 4.0, 5.0])
sp = torch.sparse_coo_tensor(i, v, (3, 3)).coalesce()
sp_csr = sp.to_sparse_csr()
out = torch.sparse.mm(sp_csr, dense_matrix)
```

### cuSPARSE-backed sparse @ dense (GPU)
```python
import cupy as cp
from cupyx.scipy.sparse import csr_matrix as cp_csr

X_gpu = cp_csr(cp.array(X.data), ...)  # or cp_csr.from_scipy
y = X_gpu @ w_gpu
```

### Polars lazy missing handling
```python
import polars as pl

df = (
    pl.scan_parquet("events/*.parquet")
      .with_columns([
          pl.col("price").fill_null(strategy="median").alias("price"),
          pl.col("category").fill_null("unknown"),
          pl.col("price").is_null().cast(pl.Int8).alias("price_was_missing"),
      ])
      .collect()
)
```

### Sparsity diagnostics
```python
def sparsity_report(X):
    if hasattr(X, "nnz"):
        density = X.nnz / (X.shape[0] * X.shape[1])
    else:
        import numpy as np
        density = np.count_nonzero(X) / X.size
    print(f"shape={X.shape}, density={density:.4%}, sparse={1-density:.4%}")
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Density < 5%, structurally sparse | scipy CSR + sparse-aware model |
| Missing < 5% | Drop or median fill |
| Missing 5-30% | Indicator + iterative impute |
| Missing > 30% on key col | Drop column or model NaN-native (LightGBM) |
| Tabular w/ NaN | LightGBM/XGBoost (native) |
| Online stream | HashingVectorizer + SGDClassifier |
| GPU @ scale | cuSPARSE / torch.sparse |

**기본값**: scipy CSR + sklearn pipeline; tabular missing 매 LightGBM native.

## 🔗 Graph
- 부모: [[Data-Engineering]] · [[Feature Engineering|Feature-Engineering]]
- 변형: [[Sparse-Matrix]] · [[Imputation]]
- 응용: [[TF-IDF]] · [[One-Hot-Encoding]] · [[Recommender-Systems]]
- Adjacent: [[Lasso]] · [[LightGBM]]

## 🤖 LLM 활용
**언제**: 매 imputation strategy rationale, 매 sparse pipeline scaffolding, 매 missingness mechanism (MCAR/MAR/MNAR) 의 explain.
**언제 X**: 매 numerical impute (use IterativeImputer/MICE), 매 distribution test (statsmodels).

## ❌ 안티패턴
- **Sparse → dense conversion**: 매 OOM 의 instant. `X.toarray()` 매 N×D > 10^9 매 ban.
- **Mean-fill skewed dist**: 매 long-tail 매 distort. 매 median 의 default.
- **Drop NaN row 매 30% lose**: 매 power loss. 매 indicator + impute.
- **Imputation 의 leakage**: 매 train+test 동시 fit 매 leak. 매 train-only fit, transform both.
- **MNAR ignored**: 매 missing-not-at-random 매 imputation 의 bias. 매 indicator 의 critical.
- **Sparse + StandardScaler(with_mean=True)**: 매 densify. 매 `with_mean=False`.

## 🧪 검증 / 중복
- Verified (scipy.sparse docs 1.14; sklearn IterativeImputer; LightGBM missing-value docs; Little & Rubin "Statistical Analysis with Missing Data" 3rd ed).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — full content (CSR/MICE/LightGBM/cuSPARSE patterns + missingness taxonomy) |