序列推荐系统的目的是通过对用户连续行为、用户与物品之间的交互，以及用户偏好和商品流行趋势随时间变化的建模，来实现更加精准、个性化和动态的推荐功能。本研究主要针对当前序列推荐中单行为推荐和多行为推荐所存在的问题，提出包括基于对比学习和时间间隔增强的序列推荐模型、基于对比学习和多尺度Transformer的多行为序列推荐模型以及个性化推荐的电商推荐系统平台。

针对当前单行为序列推荐中未考虑到用户序列中可能出现的兴趣漂移等问题，提出基于对比学习和时间间隔增强的序列推荐模型。该模型首先通过调整序列时间间隔，将非均匀序列转换为均匀序列，从而更好地捕捉用户行为的时间动态及解决兴趣漂移问题。其次，对长短期序列进行划分并分别处理。随后，采用对比学习框架，通过拉近长短期兴趣表示，能够更好的捕捉用户的兴趣变化。最后，将长期兴趣与短期兴趣进行融合，生成最后的用户兴趣表示，并计算得到候选物品的得分排序。模型在三个公开数据集下与已有的八种基线方法进行比较，其实验结果表明所提模型性能上均有提升，验证了该模型的有效性。

在电商环境下用户展示出的不同类型行为，针对现有方法未能考虑用户不同尺度下的偏好以及购买意图等问题，提出基于对比学习和多尺度Transformer的多行为序列推荐模型。首先设计了一个基于Transformer的多尺度序列编码器，联合编码来自细粒度和粗粒度级别的序列模式，旨在捕捉用户行为的多粒度兴趣表示。其次，设计一种对比学习方法，不仅考虑了用户的历史购买记录，还综合了用户其他过往行为记录，从而构建出更为深入的用户偏好。最后，通过融合多粒度下的兴趣表示和针对于购买行为的偏好表示，得到最后的用户兴趣表示并计算得到候选物品的得分排序。模型在两个公开数据集下与已有的六种基线方法进行比较，其实验结果表明所提模型性能上均有提升，验证了该模型的有效性。

最后在上述研究的基础上设计了一个个性化推荐的电商推荐系统平台，该平台总体架构分为三大模块，分别为前台用户模块、商品推荐模块、后台管理模块。并可视化展示了推荐系统平台的各功能模块界面，结果表明平台的各功能模块均能正常运行。

Sequential recommendation systems aim to understand and model users' continuous behaviors, interactions between users and items, and the evolution of user preferences and item popularity over time, thereby providing more accurate, personalized, and dynamic recommendations. This study primarily addresses the issues in current sequential recommendation systems, including single-behavior and multi-behavior recommendations, and proposes a sequence recommendation model based on contrastive learning and time interval enhancement, a multi-behavior sequence recommendation model based on contrastive learning and multi-scale Transformers, and a personalized recommendation e-commerce platform.

To tackle the problem of interest drift that may occur in user sequences in current single-behavior sequential recommendation systems, we propose a sequence recommendation model based on contrastive learning and time interval enhancement. This model first converts non-uniform sequences into uniform sequences by adjusting the sequence time intervals, thereby better capturing the temporal dynamics of user behavior and addressing the issue of interest drift. Next, it divides the sequences into long-term and short-term segments and processes them separately. Then, by adopting a contrastive learning framework, it brings the representations of long-term and short-term interests closer together, better capturing the changes in user interests. Finally, it integrates long-term and short-term interests to generate the final user interest representation and calculates the score ranking of candidate items. The model was compared with eight existing baseline methods on three public datasets, and the experimental results show that the proposed model outperforms the baselines, validating its effectiveness.

In the context of e-commerce, users exhibit different types of behaviors. To address the issues of current methods not considering users' preferences at different scales and their purchase intentions, we propose a multi-behavior sequence recommendation model based on contrastive learning and multi-scale Transformers. Firstly, a multi-scale sequence encoder based on Transformers is designed to jointly encode sequence patterns from fine-grained and coarse-grained levels, aiming to capture multi-granularity interest representations of user behavior. Secondly, a contrastive learning method is designed, which considers not only users' historical purchase records but also integrates other past behaviors to construct a more in-depth user preference. Finally, by integrating multi-granularity interest representations and preferences for purchasing behavior, the final user interest representation is obtained, and the score ranking of candidate items is calculated. The model was compared with six existing baseline methods on two public datasets, and the experimental results show that the proposed model outperforms the baselines, validating its effectiveness.

Lastly, based on the aforementioned research, a personalized recommendation e-commerce platform was designed. The overall architecture of the platform is divided into three major modules: the frontend user module, the product recommendation module, and the backend management module. The interfaces of various functional modules of the recommendation system platform were visualized, and the results show that all functional modules of the platform operate normally.