现有的侧信息集成序列推荐模型中存在对用户表示学习及优化不足的问题,针对此问题提出基于多序列交互与对比学习的侧信息集成序列推荐模型(side-information integrated sequential recommendation model based on multi-sequence inter...现有的侧信息集成序列推荐模型中存在对用户表示学习及优化不足的问题,针对此问题提出基于多序列交互与对比学习的侧信息集成序列推荐模型(side-information integrated sequential recommendation model based on multi-sequence interaction and contrastive learning,MICL)。首先,引入多序列交互注意力机制,对项目序列和侧信息序列构建序列内和序列间的深度关联,从项目和侧信息两个角度捕获用户偏好,生成两个视角的用户表示。其次,采用用户表示优化模块,结合动态难负样本采样策略构建正负样本对,利用自监督信号优化用户表示。最后,通过多任务动态权重调整策略在推荐任务与属性预测任务之间实现动态平衡优化目标,提升模型的鲁棒性和泛化能力。在Beauty、Sports、Toys和Yelp四个公共数据集上进行实验,与效果较好的基线模型相比,MICL的召回率(recall)和归一化折损率(NDCG)平均提升了1.63%和2.35%,验证了MICL对学习和优化用户表示方面的有效性。展开更多
It is important to improve the development efficiency of decoupling a coupling task package according to the information relevancy relation between development tasks in the collaborative development process of complic...It is important to improve the development efficiency of decoupling a coupling task package according to the information relevancy relation between development tasks in the collaborative development process of complicated electronic products.In order to define the task coupling model in the development process,the weighted directed graph based on the information relevancy is established,and the correspondence between weighted directed graph model and numerical design structure matrix model of coupling tasks is introduced.The task coupling model is quantized,thereby the interactivity matrix of task package is built.A multi-goal task decoupling method based on improved genetic algorithm is proposed to decouple the task coupling model,which transforms the decoupling of task package into a multi-goal optimization issue.Then the improved genetic algorithm is used to solve the interactivity matrix of coupling tasks.Finally,the effectiveness of this decomposition method is proved by using the example of task package decoupling of collaborative development of a radar’s phased array antenna.展开更多
工业物联网设备会将无法进行本地计算的任务发送至边缘服务器进行处理,但不同设备密度下的覆盖会导致不同边缘服务器的计算任务负载不均衡,进而产生计算时延过大的问题.为了解决这个问题,提出了一种基于改进的深度确定性策略梯度(modifi...工业物联网设备会将无法进行本地计算的任务发送至边缘服务器进行处理,但不同设备密度下的覆盖会导致不同边缘服务器的计算任务负载不均衡,进而产生计算时延过大的问题.为了解决这个问题,提出了一种基于改进的深度确定性策略梯度(modified deep deterministic policy gradient,MDDPG)的任务迁移算法,该算法具有基于深度确定性策略梯度的优先经验重放和随机权重平均机制,以寻求最佳的迁移策略,减少任务的计算时延.实验结果表明,MDDPG算法相较于传统的算法有更好的性能.展开更多
文摘现有的侧信息集成序列推荐模型中存在对用户表示学习及优化不足的问题,针对此问题提出基于多序列交互与对比学习的侧信息集成序列推荐模型(side-information integrated sequential recommendation model based on multi-sequence interaction and contrastive learning,MICL)。首先,引入多序列交互注意力机制,对项目序列和侧信息序列构建序列内和序列间的深度关联,从项目和侧信息两个角度捕获用户偏好,生成两个视角的用户表示。其次,采用用户表示优化模块,结合动态难负样本采样策略构建正负样本对,利用自监督信号优化用户表示。最后,通过多任务动态权重调整策略在推荐任务与属性预测任务之间实现动态平衡优化目标,提升模型的鲁棒性和泛化能力。在Beauty、Sports、Toys和Yelp四个公共数据集上进行实验,与效果较好的基线模型相比,MICL的召回率(recall)和归一化折损率(NDCG)平均提升了1.63%和2.35%,验证了MICL对学习和优化用户表示方面的有效性。
基金supported by the National Defense Basic Research Program of China (No. A1120131044)
文摘It is important to improve the development efficiency of decoupling a coupling task package according to the information relevancy relation between development tasks in the collaborative development process of complicated electronic products.In order to define the task coupling model in the development process,the weighted directed graph based on the information relevancy is established,and the correspondence between weighted directed graph model and numerical design structure matrix model of coupling tasks is introduced.The task coupling model is quantized,thereby the interactivity matrix of task package is built.A multi-goal task decoupling method based on improved genetic algorithm is proposed to decouple the task coupling model,which transforms the decoupling of task package into a multi-goal optimization issue.Then the improved genetic algorithm is used to solve the interactivity matrix of coupling tasks.Finally,the effectiveness of this decomposition method is proved by using the example of task package decoupling of collaborative development of a radar’s phased array antenna.
文摘工业物联网设备会将无法进行本地计算的任务发送至边缘服务器进行处理,但不同设备密度下的覆盖会导致不同边缘服务器的计算任务负载不均衡,进而产生计算时延过大的问题.为了解决这个问题,提出了一种基于改进的深度确定性策略梯度(modified deep deterministic policy gradient,MDDPG)的任务迁移算法,该算法具有基于深度确定性策略梯度的优先经验重放和随机权重平均机制,以寻求最佳的迁移策略,减少任务的计算时延.实验结果表明,MDDPG算法相较于传统的算法有更好的性能.
