← notes

Generative Retrieval의 프로덕션 적용 (PLUM) @Youtube

2025-11-10 · recsys, generative-recsys

PLUM: Adapting Pre-trained Language Models for Industrial-scale Generative Recommendations (2025)

Background


Method

PLUM의 3단계 구성: Item Tokenization (SID-v2) → Continued Pre-training (CPT) → Generative Retrieval (SFT)

1. Semantic IDs (SID-v2)

기존 TIGER/RQ-VAE 기반 SID-v1 ([[Semantic IDs, Generative Retrieval (TIGER)]]) 을 개선한 SID-v2 제안

Fused Multi-Modal Content Representation

Hierarchical Refinements in Quantization

Co-occurrence Contrastive Regularization

Lcon=i=12Nbexp(sim(pi,pi+))j=12Nbexp(sim(pi,pj))\mathcal{L}_{con} = -\sum_{i=1}^{2N_b} \frac{\exp(\text{sim}(p_i, p_i^+))}{\sum_{j=1}^{2N_b} \exp(\text{sim}(p_i, p_j))}

최종 SID Training Loss:

L=Lrecon+Lrq+Lcon\mathcal{L} = \mathcal{L}_{recon} + \mathcal{L}_{rq} + \mathcal{L}_{con}

2. Continued Pre-training (CPT)

학습 데이터 (두 가지 소스 50:50 혼합):

학습 규모: 1M training steps, batch size 16, 총 약 260B tokens

In-context Learning

3. Generative Retrieval (SFT)

LSFT=t=1Lr(user,vclick)logP(sidtContextuser,Historyuser,sid<t)\mathcal{L}_{SFT} = -\sum_{t=1}^{L} r(\text{user}, v_{click}) \cdot \log P(\text{sid}_t | \text{Context}_\text{user}, \text{History}_\text{user}, \text{sid}_{<t})

입력 프롬프트: SID 토큰 + numerical feature custom 토큰 + 텍스트 feature

Inference


Experiments

LEM 대비 Generative Retrieval 성능

Sample Efficiency

SID-v2 Ablation

CPT Ablation (2x2)

Scaling Study

MoE-3B 한계


💭