空间约束的在线包组推荐优化与公平性研究

Many applications, such as providing personalized and timely information to a group of customers in shopping districts, recommending a package of tourism resources to a self-help tour team in time, etc. belong to a kind of typical needs on smart services. These needs can be summarized as a new problem on package-to-group recommendation with spatial constraints, which requires that a package of items that satisfy the spatial constraints should be recommended to a group of users who are in a limited space in a timely and accurate manner. Different from the traditional strategy for recommendations, which is to construct recommendation models off-line and then to provide a global optimal solution, our proposed recommendation problem is to generate the recommended candidate set instantly and to work out the local optimal solution online, which presents new challenges on both time efficiency of and the satisfactory of recommendations. First, aiming at the effectiveness of online package-to-group recommendation with spatial constraints, we will study a novelty strategy and model by a deep representation and aggregation of multiple recommendation factors, which should provide a solution for accurate recommendation services in a limited space and time. Second, aiming at the time efficiency of package-to-group recommendation with spatial constraints, we will optimize the recommendation computation by a unified consideration on both distributed storage and distributed memory computing to realize an online package-to-group recommendation for large-scale applications. Then, considering the satisfactory of package-to-group recommendation, we will propose the semantics for recommendation fairness to cover the preferences of all users as possible and design methods maximizing the recommendation fairness to improve recommendation quality. We will also design extensive experiments to verify our achievements on massive real data and then establish prototype system.

诸如向商圈内客户群个性化地推送信息、向自助游团队实时推荐旅游资源等是一类典型的智慧服务需求。上述需求可归纳为向一组限定空间内的用户实时精准推荐一个满足空间临近性的项目包，即空间约束的包组推荐。不同于传统的离线生成推荐模型、寻求全局最优解的推荐策略，空间约束的包组推荐需要临时生成推荐候选集、在线计算局部最优解，在推荐模型、推荐时效性、推荐满意度等方面提出了新挑战。首先，针对在线包组推荐有效性，研究满足空间约束、深度聚合多推荐特征的新包组推荐策略和推荐模型，解决限定时空内的信息精准获取问题；其次，针对包组推荐时效性，研究满足包组推荐数据管理和计算需求的分布式存储与分布式内存计算协同优化方法，支持大规模实时包组推荐；最后，针对包组推荐满意度，研究覆盖异构用户组偏好差异的推荐公平性语义及其最大化方法，提升推荐服务质量。同时，基于获取的大量真实数据，设计密集实验开展方法验证，并实现原型系统。

推荐是呈现智慧的一种重要机制，在线包组推荐是提供智慧服务的一种重要模式。如何有效地建立在线包组推荐模型、优化大规模推荐计算时效性、解决推荐公平性评价等挑战性问题，实现高效准确公平的多约束在线包组推荐，既是一项重要的创新研究，也是促进新型智慧服务的关键。空间约束的在线包组推荐具有广泛应用需求，是智慧城市、智慧商圈、智慧旅游等各生态应用的核心功能之一，有效地解决空间约束的在线包组推荐及其公平性问题，对提升各类智慧应用的服务质量具有重要应用价值，对完善智慧城市基础设施建设具有重要意义。. 本项目针对用户组的实时、精准服务需求，提出了满足诸如空间约束等的在线包组推荐问题，针对在线包组推荐模型、大规模推荐计算优化、推荐评价等方面开展了系统研究。主要创新成果包括：提出了嵌入隐式用户影响力的组推荐方法、提出了基于特征级注意力神经网络模型的会话推荐方法、设计了兼顾多样化和公平性的推荐策略、设计了一种分布式矩阵计算调度执行优化策略、提出了兼顾用户影响力和特征交互的组推荐特征融合方法等。空间约束的在线包组推荐是推荐技术的新形态，对推荐应用提出了新挑战，具有重要的创新研究价值，在提供智慧旅游、智慧商圈、智慧城市服务等方面具有显著意义和应用价值。