【作者】 黄莹；

【导师】 夏海轮；

【作者基本信息】 北京邮电大学 ， 电子与通信工程（专业学位）， 2021， 硕士

【摘要】 互联网技术的发展使得网络信息资源日益庞大。对于用户而言,海量的数据信息严重干扰其对信息的正确选择,因此信息利用率非常低。对于企业而言,满足用户个性化需求对其扩大用户规模具有不可替代的作用。推荐系统可有效解决信息过载问题,提供个性化服务,因此无论对于用户还是企业,个性化推荐系统的研究都具有重要的影响和意义。本文为获得更好的推荐性能,将强化学习方法应用到推荐算法中进行了研究,同时针对企业需求,设计和实现了基于强化学习的教育资讯个性化推荐系统,具体工作如下:(1)针对推荐算法需适应用户行为特征变化,捕捉用户兴趣演变的问题,提出了一种基于强化学习和生成对抗网络的推荐算法(Model-based Reinforcement Learning with Generative Adversarial Networks and Attention Mechanism for Recommendation,MRLG Rec)。由于无模型的强化学习方法需要与真实环境频繁交互,代价较大,因此本文采用了基于模型的强化学习方法。本文在采用注意力机制充分提取用户状态特征的基础上,基于生成对抗网络构建了一个用户模拟器,用以模拟用户与推荐智能体之间的交互过程,并将这个模拟器作为强化学习的环境模型,进行推荐策略的学习。对比实验表明,所提用户模拟器能够适应用户的行为特征变化,获得较高用户行为预测准确率,基于此用户模拟器的推荐算法也获得了较高的点击率和长期奖励,有效提高了推荐性能。(2)针对企业需求,设计并实现了一个基于强化学习的教育资讯个性化推荐系统。本文首先分析了教育资讯个性化推荐系统的功能和性能需求,进行了总体架构设计和模块设计,将系统分为数据采集模块、数据存储模块、算法模块和系统业务模块;然后将所提算法MRLG Rec应用于此系统,实现了基于强化学习的教育资讯个性化推荐系统;最后进行了算法效果验证和系统测试,验证了该系统在功能和性能方面的有效性。更多还原

【Abstract】 The development of Internet technology has made the network information resources increasing sharply.For users,the massive amount of the resources interferes with their selection of information,so the utilization rate of the information is pretty low.For enterprises,meeting the individual needs of users plays an irreplaceable role in expanding the scale of users.Recommendation system can effectively solve the problem of information overload.Therefore,no matter for users or enterprises,the research on personalized recommendation system has important influence.In order to get better recommendation performancethis,this thesis applies reinforcement learning method to recommendation algorithm and designs a personalized recommendation system based on reinforcement learning according to the needs of enterprises.The specific work of this thesis is as follows:(1)For the problem that recommendation algorithm needs to adapt to the change of user behavior characteristics and capture the evolution of user interests.this thesis proposes a reinforcement learning recommendation algorithm MRLG Rec(Model-based Reinforcement Learning with Generative Adversarial Networks and Attention Mechanism for Recommendation).Model-free reinforcement learning methods require frequent interactions with the real environment,and thus are expensive in model learning,so this thesis uses a model-based reinforcement learning method.In this thesis,the attention mechanism is used to fully extract the user state characteristics,and a user simulator which is used as the reinforcement learning environment model to learn the recommendation strategy is built based on Generative Adversarial Networks to simulate the interaction process between the user and the recommendation agent.The comparative experiment shows that the proposed user simulator can adapt to the change of user behavior characteristics and obtain a high prediction accuracy of user behavior.The recommendation algorithm based on the user simulator also obtains a high click-through rate and long-term reward,which effectively improves the recommendation performance.(2)In order to meet the needs of enterprises,a personalized recommendation system about educational information based on reinforcement learning is designed and implemented.After determining the function and performance requirements of the system,this thesis designs the overall architecture and module design of the system.The system is divided into data collection module,data storage module,algorithm module,systems business module;and after the implementation of the system,the algorithm effect verification and system testing were carried out to verify the effectiveness of the system in terms of function and performance.更多还原