在我们这个充斥着海量信息的数字时代，如何从繁杂的数据中获得有价值的 信息成了一个巨大的挑战。这促使个性化推荐系统的发展受到了前所未有的重视。 其中序列推荐系统因其不以简单静态方式建模用户偏好，而是着眼于用户偏好变 化的动态过程，而受到了广泛的研究和关注。本论文针对序列推荐系统面临的关 键难题，包括如何高效建模稀疏数据、如何对用户的长期偏好进行建模、以及如 何处理复杂多行为交互的问题，提出了解决方案。本文的主要贡献和创新点包括: 首先，我们开发了 DualRec 模型，旨在深度挖掘并充分利用稀疏数据。DualRec 通过引入双重网络结构以及执行双重预测任务的方式，在模型训练中引入了未来 交互数据作为上下文信息，同时巧妙地避免了在模型训练期间和推理时出现的常 见不一致性问题。通过这种设计，DualRec 为我们提供了一个全面的用户行为视 角，极大地增强了模型在处理用户有限互动数据时，识别其偏好的能力。 为了在不增加计算负担的情况下捕捉用户行为长期依赖，我们设计了循环记 忆与排列增强 Transformer(PTM)。PTM 通过将长用户行为序列切分为多个段落， 并利用一种排列增强上下文建模技术，以提高模型在每个段落内把握序列语义的 能力。此外，模型在不同段落之间嵌入了循环记忆模块，这使得它能够以较低的 计算成本，有效捕捉到用户行为的长期依赖关系。这种设计让 PTM 模型在确保高 效率的同时，也保持了良好的推荐效果，尤其适合那些有着丰富交互行为序列的 活跃用户。 最后，我们考虑现实世界中的多行为交互，我们提出了多行为序列推荐 Trans- former(MB-STR)，专门处理用户行为中的多样性和复杂性。MB-STR 通过精巧的 设计，能够建模用户不同类型行为间的序列依赖性和时序变化模式。更重要的是， MB-STR 能够高效地利用大量的辅助行为数据，显著提升目标行为的推荐结果的 准确性。这一模型不仅在提高推荐准确性方面取得了显著成效，还深化了我们对 用户多样行为互动模式的理解。
In the current digital age, the vast amount of online information presents a significant challenge in information retrieval, leading to an increased emphasis on developing personalized recommender systems. These systems are crucial in filtering and directing relevant information to users. Among them, sequence recommender systems, which do not model user preference statically but focus on the evolving dynamics of user preferences, have received widespread attention and research. This thesis addresses the critical challenges of sequential recommender systems, including data sparsity, long-term preference modeling, and multi-behavior interactions. The main contributions and innovations are as follows:Firstly, we propose DualRec to facilitate deep mining of sparse data. DualRec tackles the challenge of utilizing future interaction data without falling prey to the training-inference gap. DualRec has a dual network structure and performs dual prediction tasks, facilitating the integration of future context information into training. DualRec offers a holistic view of user behavior while avoiding the prevalent training-inference discrepancy. This approach significantly enhances the model’s ability to discern user preferences from limited interactions, showcasing a robust framework for sequential pattern recognition.In addressing the challenge of capturing long-term dependencies in user behavior without excessive computational demand, we present the Permutation-augmented Transformer with Recurrent Memory (PTM). To capture users’ life-long preferences, we separate long user behavior sequences into segments and use a permutation context modeling objective to enhance the model’s ability to capture sequential semantics within each segment. A recurrent memory module is employed between segments to capture long-term dependencies with minimal computational overhead. This model demonstrates a balanced approach to effectiveness and efficiency, particularly for active users who generate extensive interaction sequences.Lastly, we introduce Multi-Behavior Sequential Transformer Recommender (MB-STR) to navigate the complexities of heterogeneous multi-behavior interactions. Through a sophisticated design, MB-STR can capture diverse multi-behavior sequential dependencies and temporal patterns. Besides, MB-STR can utilize massive auxiliary behavior data to enhance the quality of the recommendation. This model not only achieves superior accuracy in recommendations but also provides deeper insights into multi-behavior user interactions, paving the way for future research in recommender systems.
