Review of the Research Progress of Sequential Recommendation = 서열 추천 연구 진전 회고

다국어 초록 (Multilingual Abstract)

In the era of information explosion and the digital economy, recommender systems, as a key technology for solving the problem of information overload, are playing an increasingly important role in e- commerce, social media, online content platforms, and other fields. Traditional recommender systems mainly focus on static user-item interactions, ignoring the inherent sequential and temporal dynamics of user behavior. Sequential Recommendation Systems (SRS), as an emerging paradigm of recommender systems, capture the dynamic evolution of user preferences by modeling the temporal order of user- item interactions, thus achieving more accurate and personalized recommendation services. Human-system interactions exhibit a certain sequence, which needs to be taken into account when providing recommendations; this is the essence of a sequence recommendation system. Sequence recommendation systems analyze user behavior sequences and capture dynamic changes in user preferences to provide accurate personalized recommendation services for various fields such as e-commerce, social media, and online video. This paper comprehensively reviews the current research progress of sequence recommendation systems and explores their importance and application potential in the field of personalized recommendation. It defines the research problems of sequence recommendation, clarifying the core objectives and challenges of recommendation sequences. The paper categorizes and summarizes the main techniques of sequence recommendation, including: traditional methods based on Markov chains, which rely on state transition probabilities when modeling user behavior sequences; deep learning-driven methods, which utilize neural network models to capture long-term dependencies and complex patterns; hybrid model methods, which combine multiple algorithms to enhance the accuracy and robustness of recommendation systems; and emerging methods based on large language models. In the section on traditional methods, this paper analyzes in detail the theoretical basis of Markov chain models, including first-order Markov chains, higher-order Markov chains, factorization personalized Markov chains (FPMC) and their improved models such as HRM, TransRec and TransRec++, and explores in depth the limitations of their probability transition assumptions. In the section on deep learning methods, this paper systematically studies four main techniques: recurrent neural networks (RNNs) and their variants (GRU4Rec, IDLSTM, LSTeM, TLSRec, etc.), convolutional neural networks (Caser, NextItNet, CosRec, etc.), attention mechanisms and Transformers (SASRec, BERT4Rec, TiSASRec, etc.), and graph neural networks (SR-GNN, SURGE, GSASRec, etc.). The section on hybrid models innovatively proposes four fusion paradigms:"GNN+CNN+Attention","GNN+RNN+Attention", "CNN+Attention", and "RNN+Attention". It also provides a detailed analysis of the architectural design and collaborative mechanisms of cutting-edge models such as TiDA-GCN, DGSR, AdaMCT, and DSIM. This paper specifically points out that hybrid models, through multi- level feature extraction and cross-module information complementarity, demonstrate significant advantages in capturing users' long-term preferences and short-term interests. In the section on sequence recommendation methods based on Large Language Models (LLM), a systematic classification framework was constructed, covering cross-domain recommendation (ZESREC, UniSRec, P5, VQ-Rec), domain-specific recommendation (Prompt4NR, TALLRec, PALR, GPT4Rec), multimodal recommendation (VIP5, TransRec), and other cutting-edge directions (DRDT, ReLLa). The implementation mechanisms of emerging technologies such as zero-shot recommendation, contextual learning, and thought chain hints were also explored in depth. Finally, this paper also looks ahead to future research directions in sequence recommendation systems, emphasizing the importance of research areas such as context awareness, multimodal fusion, causal inference, vertical domain-specific large language models, and mitigation of the illusion problem. To provide a comprehensive overview and enable researchers to understand the latest developments in sequence recommendation, this paper investigates this topic, categorizes and summarizes various techniques and related work. The main contributions of this paper include the following: (1) An original sequence recommendation classification framework is proposed, introducing a method based on a large language model, which can contribute insights to existing sequence recommendation classifications. (2) A detailed analysis of relevant sequence recommendation works is conducted based on the provided classification. (3) Valuable future research directions are proposed to address the challenges faced by sequence recommendation.

목차 (Table of Contents)

• I. Introduction 1
• 1.1 Background and Significance 1
• 1.1.1 Background 1
• 1.1.2 Significance 2
• 1.2 Current Status of Research at Home and Abroad 4
• 1.2.1 Traditional Sequence Recommendation Methods 4
• 1.2.2 Application of Deep Learning in Sequence Recommendation 5
• 1.3 Research Content and Main Contributions 7
• 1.3.1 Research Questions 7
• 1.3.2 Main Research Content 8
• 1.3.3 Main Contributions 8
• 1.4 Thesis Framework 9
• II.Basic concepts and methods of sequence recommendation 11
• 2.1 Development History of Recommender Systems 11
• 2.2 Sequential Recommendation: Concept and Problem Statement 12
• 2.2.1 Sequential Recommendation Concept 12
• 2.2.2 Problem Statement 13
• III. Classification of Sequence Recommendation Methods 16
• 3.1 Classification 16
• 3.2 Traditional Sequence Recommendation Methods 16
• 3.3 Deep learning-based sequence recommendation methods 20
• 3.3.1 Recurrent Neural Network (RNN) 20
• 3.3.2 Convolutional Neural Networks (CNNs) 25
• 3.3.3 Attention Mechanism and Transformer Model 31
• 3.3.4 Graph Neural Networks (GNNs) 38
• 3.4 Hybrid Model Sequence Recommendation Methods 42
• 3.4.1 GNN+CNN+Attention 42
• 3.4.2 GNN+RNN+Attention 44
• 3.4.3 CNN +Attention 47
• 3.4.4 RNN +Attention 49
• 3.5 Sequence Recommendation Methods Based on Large Language Models 53
• 3.5.1 A Cross-Domain Sequence Recommendation Model Based on LLM 53
• 3.5.2 Domain-Specific Sequence Recommendation Model Based on LLM 56
• 3.5.3 LLM-Based Multimodal Sequence Recommendation Model 60
• 3.5.4 Other Methods 62
• IV. Future Outlook 66
• 4.1 Context-aware sequence recommendation system 66
• 4.2 Multimodal Sequence Recommendation Systems 67
• 4.3Causal Inference and Intervention in Sequential Recommendation 68
• 4.4 Sequence Recommendation Based on Domain-Specific Large Language Models 69
• 4.5 Sequence Recommendation to Mitigate the Illusion Problem 71
• V. Conclusion 74
• References 75
• Acknowledgments 92