---
id: wiki-2026-0508-recommendation-systems
title: Recommendation Systems
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [RecSys, Recommender Systems, Recommendation Engine]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [recsys, machine-learning, ranking, retrieval]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: pytorch
---

# Recommendation Systems

## 매 한 줄
> **"매 user × item 의 relevance prediction at scale"**. 매 collaborative filtering (Netflix Prize 2009) → matrix factorization → deep two-tower / sequential / generative recsys. 2026 현재 매 industrial stack 의 multi-stage (retrieval → ranking → re-ranking), 매 LLM-augmented (semantic ID, generative recsys, LLM reranker).

## 매 핵심

### 매 paradigms
- **Content-based**: item features × user profile (cold-start friendly).
- **Collaborative filtering (CF)**: user-item interactions only.
  - **Memory-based**: user-user / item-item KNN.
  - **Model-based**: matrix factorization (MF, ALS, SVD++).
- **Hybrid**: CF + content (LightFM, DCN).
- **Deep**:
  - **Two-tower**: user-tower / item-tower → dot product (retrieval).
  - **Sequential**: SASRec, BERT4Rec, GRU4Rec — 매 user history sequence modeling.
  - **DIN/DIEN**: attention over user behaviors (Alibaba).
  - **Graph**: PinSage, LightGCN.
- **Generative recsys (2024-2026)**: TIGER, semantic ID, LLM-as-recommender.

### 매 pipeline (industrial)
1. **Candidate generation (retrieval)**: 100M items → 1000 (two-tower ANN, FAISS / ScaNN).
2. **Ranking**: 1000 → 100 (heavy DCN / DIN, full features).
3. **Re-ranking**: 100 → 10 (diversity, business rules, MMR, RL).
4. **Serving**: <100ms p99.

### 매 metrics
- **Offline**: Recall@K, NDCG@K, MAP, AUC, MRR.
- **Online (A/B)**: CTR, conversion, dwell time, session length, retention.
- **Diversity / fairness**: ILD (intra-list diversity), exposure parity.

### 매 응용
1. E-commerce (Amazon, Coupang, Taobao).
2. Video / music (YouTube, TikTok, Spotify).
3. Social feed (Facebook, Twitter/X, LinkedIn).
4. News (Toutiao, Yahoo News).
5. Ads (Google Ads ranking).

## 💻 패턴

### Matrix factorization (implicit ALS)
```python
import implicit
from scipy.sparse import csr_matrix

# user-item interactions (rows=users, cols=items)
ui = csr_matrix(interactions)  # 1.0 for click, weighted by dwell

model = implicit.als.AlternatingLeastSquares(
    factors=128, regularization=0.01, iterations=20, use_gpu=True,
)
model.fit(ui)
recs = model.recommend(userid=42, user_items=ui[42], N=10)
```

### Two-tower retrieval (PyTorch)
```python
import torch
import torch.nn as nn

class TwoTower(nn.Module):
    def __init__(self, n_users, n_items, dim=128):
        super().__init__()
        self.u_emb = nn.Embedding(n_users, dim)
        self.i_emb = nn.Embedding(n_items, dim)
        self.u_mlp = nn.Sequential(nn.Linear(dim, dim), nn.ReLU(), nn.Linear(dim, dim))
        self.i_mlp = nn.Sequential(nn.Linear(dim, dim), nn.ReLU(), nn.Linear(dim, dim))

    def user_repr(self, u): return nn.functional.normalize(self.u_mlp(self.u_emb(u)), dim=-1)
    def item_repr(self, i): return nn.functional.normalize(self.i_mlp(self.i_emb(i)), dim=-1)

    def forward(self, u, i_pos, i_negs):
        u_v = self.user_repr(u).unsqueeze(1)        # (B, 1, D)
        pos = self.item_repr(i_pos).unsqueeze(1)    # (B, 1, D)
        negs = self.item_repr(i_negs)               # (B, K, D)
        logits = torch.cat([u_v @ pos.transpose(-1,-2), u_v @ negs.transpose(-1,-2)], dim=-1).squeeze(1)
        labels = torch.zeros(u.size(0), dtype=torch.long, device=u.device)
        return nn.functional.cross_entropy(logits / 0.07, labels)  # in-batch + sampled negs
```

### ANN retrieval with FAISS
```python
import faiss, numpy as np

item_vecs = model.item_repr(torch.arange(n_items)).detach().cpu().numpy()
index = faiss.IndexFlatIP(128)        # inner product
index.add(item_vecs)

def retrieve(user_vec, k=1000):
    D, I = index.search(user_vec[None, :], k)
    return I[0]
```

### SASRec (sequential recsys)
```python
class SASRec(nn.Module):
    def __init__(self, n_items, dim=64, max_len=200, n_heads=2, n_layers=2):
        super().__init__()
        self.item_emb = nn.Embedding(n_items + 1, dim, padding_idx=0)
        self.pos_emb = nn.Embedding(max_len, dim)
        layer = nn.TransformerEncoderLayer(dim, n_heads, batch_first=True, activation="gelu")
        self.tr = nn.TransformerEncoder(layer, n_layers)

    def forward(self, seq):  # (B, L)
        L = seq.size(1)
        pos = torch.arange(L, device=seq.device).unsqueeze(0)
        x = self.item_emb(seq) + self.pos_emb(pos)
        mask = torch.triu(torch.ones(L, L), diagonal=1).bool().to(seq.device)
        h = self.tr(x, mask=mask)
        # next-item prediction
        return h @ self.item_emb.weight.T   # (B, L, V)
```

### DIN-style attention over user history
```python
class DINAttention(nn.Module):
    def __init__(self, dim):
        super().__init__()
        self.mlp = nn.Sequential(nn.Linear(4*dim, dim), nn.ReLU(), nn.Linear(dim, 1))

    def forward(self, target, history, mask):
        # target (B, D), history (B, L, D), mask (B, L)
        T = target.unsqueeze(1).expand_as(history)
        feats = torch.cat([T, history, T - history, T * history], dim=-1)
        attn = self.mlp(feats).squeeze(-1)
        attn = attn.masked_fill(~mask, -1e9).softmax(-1)
        return (attn.unsqueeze(-1) * history).sum(1)
```

### LLM reranker (2026)
```python
from anthropic import Anthropic
client = Anthropic()

def llm_rerank(user_history, candidates):
    msg = client.messages.create(
        model="claude-opus-4-7",
        max_tokens=500,
        messages=[{"role": "user", "content": f"""
User watched: {user_history}
Rerank these candidates by relevance, return top-10 IDs only as JSON array:
{candidates}
"""}],
    )
    return parse_json(msg.content[0].text)
```

### Implicit feedback BPR loss
```python
def bpr_loss(u, i_pos, i_neg, model):
    s_pos = (model.u(u) * model.i(i_pos)).sum(-1)
    s_neg = (model.u(u) * model.i(i_neg)).sum(-1)
    return -torch.log(torch.sigmoid(s_pos - s_neg) + 1e-12).mean()
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Cold-start (new user / item) | content-based + popularity |
| Small data (<10k users) | item-item KNN |
| Mid (10k-1M) | ALS / LightFM |
| Large (>1M, sequence behavior) | two-tower retrieval + DIN/SASRec ranking |
| Strict latency budget | two-tower + ANN |
| Need explanation / control | LLM reranker on top-100 |
| Cross-domain (text + image) | multimodal embeddings (CLIP-style) |

**기본값**: 매 industrial 의 two-tower retrieval + DCN/DIN ranking + business-rule rerank, 매 ANN (FAISS / ScaNN), 매 implicit feedback + sampled softmax. 매 LLM-as-reranker 의 emerging 2026 pattern for top-K refinement.

## 🔗 Graph
- 부모: [[Machine-Learning]] · [[Information-Retrieval]]
- 변형: [[Collaborative-Filtering]]
- 응용: [[E-commerce]]
- Adjacent: [[Embeddings]] · [[FAISS]]

## 🤖 LLM 활용
**언제**: cold-start (zero-shot recommendation from item description), reranker on top-100, explanation generation, semantic ID encoding.
**언제 X**: full-funnel retrieval at scale (latency / cost prohibitive). 매 LLM 의 reranker only, 매 retrieval 의 ANN.

## ❌ 안티패턴
- **Random negative sampling only**: easy negatives, model 의 saturate — use hard negatives + in-batch negatives.
- **Train on biased logged data**: position bias / popularity bias not corrected → IPS / counterfactual.
- **Offline metric chasing**: NDCG up but online CTR flat — online A/B 의 truth.
- **Cold-start ignore**: pure CF 의 fail on new items — hybrid fallback.
- **No exploration**: greedy ranking → filter bubble. ε-greedy / Thompson / contextual bandit.
- **Single objective**: CTR-only optimization 의 clickbait. Multi-objective (dwell, retention).

## 🧪 검증 / 중복
- Verified (Koren 2009 Netflix Prize, He LightGCN 2020, Kang SASRec 2018, Zhou DIN 2018, Covington YouTube DNN 2016, Google TIGER 2024).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — full canonical recsys with two-tower/SASRec/DIN/LLM-reranker |