2nd/10_Wiki/Topics/AI_and_ML/Dimensionality-Reduction.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-dimensionality-reduction	Dimensionality Reduction	10_Wiki/Topics	verified	self	PCA t-SNE UMAP autoencoder curse of dimensionality feature extraction	none	A	0.93	applied	dimensionality-reduction pca tsne umap autoencoder visualization manifold-learning		2026-05-10	pending	language framework Python scikit-learn / umap-learn / PyTorch

Dimensionality Reduction

매 한 줄

"매 high-dim 의 essence 의 low-dim". 매 PCA (linear) → 매 t-SNE / UMAP (nonlinear, 시각화) → 매 Autoencoder / VAE (deep). 매 modern: 매 embedding (CLIP, sentence-transformers) 의 implicit dim reduction.

매 핵심 method

Linear

PCA (Principal Component Analysis)

• 매 variance 의 maximum direction.
• 매 orthogonal axis.
• 매 SVD.
• 매 fast + interpretable.

LDA (Linear Discriminant Analysis)

• 매 class separation 의 maximize.
• 매 supervised.

Factor Analysis

• 매 latent factor 의 explain variance.

Nonlinear (manifold)

t-SNE (Maaten 2008)

• 매 local neighborhood 의 preserve.
• 매 visualization 강.
• 매 global structure 의 weak.
• 매 stochastic.

UMAP (McInnes 2018)

• 매 t-SNE 의 successor.
• 매 faster + 매 global structure 도 better.
• 매 default for high-dim viz.

Isomap

• 매 geodesic distance 의 preserve.

LLE (Locally Linear Embedding).

Neural

Autoencoder

• 매 bottleneck 의 dim reduce.

VAE (Variational AE)

• 매 probabilistic.

Self-Supervised Embedding

• 매 CLIP, BERT, sentence-transformers.
• 매 implicit dim reduction.

매 PaCMAP / TriMap (recent)

• 매 UMAP 의 variant.
• 매 better global structure.

매 응용

1. Visualization (2D / 3D): 매 t-SNE, UMAP.
2. Speed (preprocess): 매 PCA.
3. Anomaly detection: 매 autoencoder.
4. Feature extraction: 매 embedding.
5. Compression: 매 quantization + 매 embed.
6. Clustering preprocessing.
7. RAG (vector DB): 매 PCA / quantization.

매 curse of dimensionality

• 매 distance 의 meaningless.
• 매 sparsity in 매 high-dim.
• 매 sample requirement 의 exponential.

💻 패턴

PCA (sklearn)

from sklearn.decomposition import PCA
from sklearn.preprocessing import StandardScaler

X_scaled = StandardScaler().fit_transform(X)

pca = PCA(n_components=0.95)  # 매 95% variance 의 keep
X_reduced = pca.fit_transform(X_scaled)

print(f'Original: {X.shape[1]}, reduced: {pca.n_components_}')
print(f'Explained variance: {pca.explained_variance_ratio_.cumsum()}')

t-SNE

from sklearn.manifold import TSNE

tsne = TSNE(
    n_components=2,
    perplexity=30,
    n_iter=1000,
    random_state=42,
)
X_2d = tsne.fit_transform(X[:5000])  # 매 t-SNE 의 slow → 매 sample

UMAP (modern)

import umap

reducer = umap.UMAP(
    n_components=2,
    n_neighbors=15,
    min_dist=0.1,
    metric='cosine',  # 매 embedding 에 좋음
    random_state=42,
)
X_2d = reducer.fit_transform(X)

Autoencoder (PyTorch)

import torch.nn as nn

class AE(nn.Module):
    def __init__(self, input_dim, latent_dim=32):
        super().__init__()
        self.encoder = nn.Sequential(
            nn.Linear(input_dim, 128), nn.ReLU(),
            nn.Linear(128, 64), nn.ReLU(),
            nn.Linear(64, latent_dim),
        )
        self.decoder = nn.Sequential(
            nn.Linear(latent_dim, 64), nn.ReLU(),
            nn.Linear(64, 128), nn.ReLU(),
            nn.Linear(128, input_dim),
        )
    
    def forward(self, x):
        z = self.encoder(x)
        return self.decoder(z), z

# 매 latent 의 use
model = AE(input_dim=784)
# ... train ...
_, latent = model(X_test)

Visualization combo (UMAP + scatter)

import matplotlib.pyplot as plt

X_2d = umap.UMAP().fit_transform(X)
plt.figure(figsize=(10, 8))
plt.scatter(X_2d[:, 0], X_2d[:, 1], c=labels, cmap='tab10', alpha=0.5, s=10)
plt.colorbar()
plt.title('UMAP projection')
plt.show()

PCA for speed (vector DB preprocessing)

from sklearn.decomposition import PCA
import faiss

# 매 매 768 의 OpenAI embedding → 매 256
embeddings = get_embeddings(documents)
pca = PCA(n_components=256)
reduced = pca.fit_transform(embeddings).astype('float32')

# 매 Faiss
index = faiss.IndexFlatIP(256)
index.add(reduced)

Quantization (vector DB modern)

import faiss

dim = 768
quantizer = faiss.IndexFlatIP(dim)
index = faiss.IndexIVFPQ(quantizer, dim, nlist=100, m=8, nbits=8)
# 매 8 byte 의 768-dim 의 represent — 매 매 100× compression.

index.train(embeddings_np)
index.add(embeddings_np)

Word2Vec / CLIP-style (implicit reduction)

from sentence_transformers import SentenceTransformer

model = SentenceTransformer('all-MiniLM-L6-v2')  # 매 384-dim
embeddings = model.encode(sentences)
# 매 매 sentence (potentially infinite words) → 매 384-dim.

Reconstruction error (anomaly)

def detect_anomaly(model, X, threshold):
    X_recon, _ = model(X)
    error = ((X_recon - X) ** 2).mean(dim=1)
    return error > threshold

Choose dimension (elbow / cumvar)

import numpy as np
import matplotlib.pyplot as plt

pca = PCA().fit(X_scaled)
cumvar = np.cumsum(pca.explained_variance_ratio_)

plt.plot(cumvar)
plt.xlabel('Component')
plt.ylabel('Cumulative variance')
plt.axhline(0.95, color='r', linestyle='--')
plt.show()

n_components = np.argmax(cumvar >= 0.95) + 1

Manifold visualization comparison

def viz_compare(X, labels):
    fig, axes = plt.subplots(1, 3, figsize=(20, 6))
    
    for ax, (name, reducer) in zip(axes, [
        ('PCA', PCA(n_components=2)),
        ('t-SNE', TSNE(n_components=2, random_state=42)),
        ('UMAP', umap.UMAP(n_components=2, random_state=42)),
    ]):
        proj = reducer.fit_transform(X)
        ax.scatter(proj[:, 0], proj[:, 1], c=labels, cmap='tab10', s=5)
        ax.set_title(name)

매 결정 기준

상황	Method
Speed (preprocess)	PCA
Visualization	UMAP
Cluster preserve	UMAP
Variance interpret	PCA
Class-aware	LDA
Text → embedding	Sentence-transformer
Image → embedding	CLIP
Vector DB compress	PCA / PQ quantization
Anomaly	Autoencoder
Generative	VAE

기본값: PCA (preprocess) + UMAP (viz) + embedding (semantic).

🔗 Graph

• 부모: Feature Engineering
• 변형: PCA · t-SNE · UMAP · Autoencoder · VAE
• 응용: CLIP · Sentence-Transformers · Faiss · Anomaly-Detection
• Adjacent: Auto-Encoding · Bag of Words (BoW) · Bias vs Variance Trade-off

🤖 LLM 활용

언제: 매 visualization. 매 vector DB. 매 cluster preprocessing. 매 anomaly detection. 언제 X: 매 already low-dim. 매 lossless 필수.

❌ 안티패턴

• PCA without standardize: 매 wrong principal component.
• t-SNE 의 cluster size 의 interpret: 매 not preserved.
• UMAP 의 distance 의 absolute interpret: 매 local 만.
• Too aggressive reduction: 매 information loss.
• Forget train-test split: 매 leakage in PCA.

🧪 검증 / 중복

• Verified (Jolliffe PCA, van der Maaten t-SNE, McInnes UMAP).
• 신뢰도 A.
• Related: Auto-Encoding · Bag of Words (BoW) · CLIP · Sentence-Transformers · Bias vs Variance Trade-off.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — methods + 매 PCA / t-SNE / UMAP / AE / Faiss / quantization code