가중 K-shell을 활용한 특징 적응형 그래프 어텐션 기반 추천 시스템 = Data-Efficient Graph Attention Network for Recommendation with Weighted K-Shell Features

국문 초록 (Abstract)

사용자-아이템 상호작용 그래프 기반 추천 시스템은 희소성, 노이즈가 포함된 관계, 그리고 심층 메시지 패싱에서 발생하는 과잉 평활화 문제를 겪는다. 본 논문은 무거운 메타데이터에 의존하지 않으면서 데이터 효율성과 강건성을 향상시키는 2단계 프레임워크를 제안한다.

Method-I에서는 평점 강도, 시간적 최신성, 시청 연도 엔트로피, 시간적 모멘텀을 엣지 가중치로 변환하는 특징 인식 가중 k-shell 기법을 도입하고, 수정된 k-shell 프로세스를 통해 노드 중요도를 도출한다. 이 점수를 활용하여 고코어 노드는 제거하고 저코어 노드는 보강하는 개인화된 엣지 교란 뷰를 구축하며, 대조 학습을 통해 그래프 인코더를 훈련하여 강건하고 메타데이터 의존도가 낮은 사용자/아이템 표현을 획득한다.

Method-II에서는 특징 적응형 다중 헤드 그래프 어텐션 네트워크인 FAGAT를 제시한다. 동일한 헤드를 사용하는 대신, FAGAT는 RatingEncoding, PopEncoding(차수/통계), TimeEncoding이라는 세 가지 특징 기반 헤드를 사용하며, 각 헤드는 어텐션 집계 전에 LayerNorm과 학습 가능한 게이팅 스칼라로 안정화된다. 최종 사용자/아이템 임베딩은 BPR 손실을 통해 랭킹에 최적화된다.

MovieLens-100K, MovieLens-1M, 그리고 극도로 희소한 Amazon Digital Music 데이터셋에서 수행한 실험 결과, 7가지 랭킹 지표(Recall@K, Precision@K, F1@K, MRR@K, MAP@K, HR@K, NDCG@K)와 다양한 K 값에서 일관된 성능 향상을 보였다. 특히 제안 모델은 강력한 GNN 베이스라인(GCN, GIN, GAT, GraphSAGE, AGNN)에 비해 명확한 개선을 보였으며, 희소한 음악 데이터에서는 기본 BPR 대비 특히 큰 상대적 성능 향상을 달성하여 희소성 환경에서의 강건성을 입증하였다. 반면, 최신 SSL 기반 모델에 비해서는 다소 낮은 성능을 보여, 극도로 희소한 데이터셋에서 순수 지도 학습의 한계를 드러냈다.

목차 (Table of Contents)

• I. 서론························································································································· 1
• II. 관련연구················································································································ 5
• 1. Mixture of Experts 기반 그래프 추천 시스템·········································6
• 2. Multi_view Learning 기반 그래프 추천 시스템······································7
• 3. 그래프 신경망 관련 연구 ·············································································8
• III. 제안하는 연구···································································································13
• 1. 전체 시스템 구조도·······················································································13
• 2. 데이터 수집 및 전처리 ···············································································15
• 3. Feature-aware K-shell based representation Learning ······················18
• 3-1. K-shell 알고리즘의 데이터기반 엣지 변형···································18
• 3-2. 특징 가중치기반의 Weighted Kshell··············································23
• 3-3. K-core 기반 그래프 데이터 증강기 설계·····································24
• 3-4. 그래프 증강기 학습 및 결과 시각화··············································29
• 4. Feature-adaptive Graph Attention Network ········································30
• 4-1. GAT에서의 Multi-head 연산 ·························································30
• 4-2. 데이터 효율적인 FAGAT 모델의 헤드 설계·······························33
• 4-3. 안정적인 특징 주입방법 및 헤드 설계 ········································37
• 4-4. 전체 학습 파이프라인 ······································································39
• IV. 성능 평가 ········································································································44
• 1. BaseLine 모델 선정······················································································44
• 2. 학습 환경 및 하이퍼파라미터 설정 ·························································47
• 3. 추천 시스템에서의 성능 평가 방법 및 지표···········································49
• 4. 성능 비교 ·······································································································53
• V. 결론······················································································································ 59
• 부록···························································································································· 61
• 참고 문헌·················································································································· 66
• ABSTRACT············································································································ 71
• 감사의 글·················································································································· 74