【作者】 张旭；

【导师】 欧中洪；

【作者基本信息】 北京邮电大学 ， 计算机科学与技术， 2023， 硕士

【摘要】 随着互联网时代的高速发展,用户难以在海量信息中快速筛选出感兴趣的内容,面临信息过载问题,推荐系统应运而生。推荐系统根据用户历史行为、个人信息等数据计算出用户可能感兴趣的物品集,为每个人实现个性化推荐。多兴趣个性化推荐系统对于每一个用户,算法能够学习到用户的多种不同的兴趣偏好,从个性化推荐的“千人千面”效果升级为“千人万面”。目前推荐系统在新闻资讯、旅游、电子商务等领域都有了广泛的应用,然而目前基于用户多兴趣的推荐系统依旧存在许多问题:(1)目前主流的多兴趣推荐系统只考虑了用户短期历史记录,忽视了用户长期历史记录中蕴含的丰富信息;(2)当前主流的推荐系统点击率预估算法忽视了用户和物品特征中的高阶特征交互信息或对高阶特征交叉信息挖掘不充分;(3)个性化推荐系统在科技资源领域落地应用较少,实践应用经验不足。针对上述问题,本文研究并实现一个基于用户多兴趣的推荐系统,主要研究工作包含:(1)设计并实现了一种基于用户长短期历史的多兴趣召回算法,通过不同的神经网络模型结构分别建模用户长短期兴趣偏好,并通过门控融合网络融合用户长短期兴趣偏好最终得到用户多个兴趣偏好。在MovieLens数据集上于HR@50指标较先前最优模型提升了 4.49%,Taobao数据集上于HR@100指标较先前最优模型提升了 8.55%,实现了个性化推荐召回。(2)研究并实现了一种基于子空间投影神经网络的点击率预估算法,通过子空间投影机制在不同的子空间中实现不同的高阶交互特征,并通过堆叠子空间投影层实现复杂特征交互,在公开数据集Criteo和Avazu上优于先前最优模型,并通过可视化展现了各个特征的重要度,实现了个性化推荐排序。(3)基于以上研究内容,面向科技资源领域设计并实现了一套个性化推荐系统,实现了页面交互、模型训练部署、账户管理和详情页展示等功能,为用户个性化推荐科技资源,在实际应用中验证了上述两个算法的有效性。更多还原

【Abstract】 With the fast development of the Internet era,making it difficult for users to quickly sift through the massive amount of information and find content of interest.Users face with information overload,recommendation systems were born to solve this problem.Based on user historical behavior and personal information,recommendation systems recommend a set of items that users may be interested in.Recommendation systems realizing personalized recommendations for everyone.A multi-interest personalized recommendation system can learn a user’s multiple different interests,upgrading the "thousand faces,thousand worlds" effect to "thousand faces,ten thousand worlds." Currently,recommendation systems have been widely used in areas such as news,travel,and e-commerce.However,there are still many problems with multi-interest recommendation systems:(1)Current mainstream multi-interest recommendation systems only consider a user’s short-term history,ignoring the rich information contained in a user’s long-term history;(2)Current mainstream recommendation system’s click-through rate prediction algorithms ignores high-order feature interactions in the user and item features or inadequately models high-order feature interactions;(3)Personalized recommendation systems have been underutilized in the field of technology resources,resulting in insufficient practical application experience.In light of the above-mentioned issues,this paper studies and implements a recommendation system based on user multi-interest.The main research work includes:(1)Designing and implementing a multiinterest recall algorithm based on user long and short-term history.Different neural network models are used to model user long and shortterm interests,and the gate fusion network is used to fuse these preferences to obtain the user’s multiple interests.The proposed algorithm achieved a 4.49%improvement in HR@50 on the MovieLens dataset and an 8.55%improvement in HR@100 on the Taobao dataset compared to the previous best model,implement the personalized recommendation recall.(2)Studying and implementing a click-through rate prediction algorithm based on a subspace projection neural network.The model implements different high-order feature interactions in different subspaces through a subspace projection mechanism,and achieves complex feature interactions through stacked subspace projection layers.The proposed algorithm outperforms the previous best model on the Criteo and Avazu datasets,and the importance of each feature is shown through visualization,implement the personalized recommendation sort.(3)Based on the above research,a personalized recommendation system for the technology resource domain is designed and implemented.The system includes features such as page interaction,model training and deployment,account management,and detail page display,recommend technology resources personalized for users,the effectiveness of the aforementioned two algorithms has been validated in practical applications.更多还原