智能感知与计算研究中心知识库

网络空间的对象建模是指通过表达学习的方式，将网络空间中对象的复杂信息编码为计算机可以处理的向量形式。高质量的对象建模是完成网络空间中各种相关任务的基础，如计算广告、电子商务等。网络空间中的对象信息通常包含了多种模态、多种结构的属性特征，且对象之间关系复杂。图方法在建模复杂结构方面有着强大的灵活性，这使其在网络空间的对象建模存在着巨大优势。因此，基于图方法的网络空间对象建模在理论和应用上都具有重要的意义和价值。本论文围绕如何有效地采用图方法挖掘对象的属性和关系信息来展开，重点关注三个话题：深度图模型研究、对象属性建模、对象关系建模。
论文的主要工作和创新点归纳如下:

1. 深度动态图模型研究

网络空间的对象结构特性复杂且多变。大多数有关图方法的研究关注在静态图的结构特性挖掘上，从而做到对网络空间对象结构复杂性的建模。但是网络空间的对象特性多变，且随着和用户、商家等实体地频繁交互，对象的特征表达也在不断变化。少有工作关注图结构的动态特性上。本文充分利用了图卷积神经网络中表达传播的思想，并将其运用到动态场景下，设计出每个时间点只进行局部少量节点更新且不用进行微调优化的动态图卷积神经网络，在保证节点表达性能的前提下极大地提升了节点表达更新效率。

在图表达动态学习更新的基础上，进一步学习动态结构变化的规律可以为对象动态特性建模获得更多有效信息。传统的动态图链接预测模型总是依赖于动态表达更新模型来建模图结构变化规律。然而由于表达学习本身存在的结构损失且不同时间的图之间存在对齐偏差，这类方法有着不可避免的动态累计误差。 本文直接建模新增边的映射关系，通过记忆机制来存储历史发生链接的映射，作为动态图链接预测的参考。该方法避免了高频率的节点表达更新，进一步提升了动态图的建模效率。

2. 基于图方法的对象属性建模研究

图结构几乎存在于网络空间对象属性特征的方方面面。如何更好地从复杂的属性特征中建模对象的整体表达是建模对象属性的关键。传统的对象属性建模的方法通常将这些特征单纯地看作增强对象特征表达的辅助信息，简单地将这些属性融合到一起，缺乏对属性的结构建模。本文设计了一种基于二分图匹配的对象解耦表达模型，为每个对象自适应地分配其可能所属的属性特征。该方法直接采用多元属性组合的形式表达对象，既丰富了对象属性特征维度，也提升了对象的表达效率。该表达策略可以直接应用到下游各种基于对象表达学习的任务中去。在此框架下，下游任务可以关注到对象属性的不同方面，提升了诸如推荐系统等任务的效果。

另一方面，对象在不同的场景不同的时间段会表现出不同的状态。针对其时序属性表达建模有着十分重要的意义。本文从生命周期划分的角度，将对象的时序属性随着时间变化的过程，抽象为不同生命周期状态转移的有向图结构。生命周期状态分配过程则可以转化为状态转移有向图中的最短路径搜索问题。通过该策略，可以自适应地判别对象所处的生命周期状态，进而更精确地建模时序场景下的对象表达。

3. 基于图方法的对象关系建模研究

不同对象之间的交互关系多种多样，例如商品之间的共购关系、服装搭配关系、软件使用的逻辑先后关系等。传统的模型有的忽略了对于这些复杂图结构性的建模，有的将对象之间的关系与对象的属性特征完全割裂开来看，没有充分建模特征的交互过程。本文采用图神经网络建模多对象之间的整体交互关系。图神经网络可以使对象信息在考虑了关系结构的基础上进行交互，从而获得了更加整体的表达特征。本文将该方法运用到服装搭配推荐任务中去进行应用验证，利用搭配套装之间的关联性建图，充分建模多件服装对象的交互关系，获得了优秀的效果。

不同对象之间在被用户交互的先后时序上也存在着一定的逻辑关系。根据这样的先后关系进行构图，可以反映全局下的对象之间的时序转移关系。充分把握这样的时序转移关系，对推荐系统、计算广告等业务都有着重要的价值。参考动态图链接预测模型部分的模型思路，本文通过记录并整理历史的时序转移关系，来辅助未来的时序预测。为了增强对于局部结构关系的利用，本文提取历史对象转移关系的图主题特征，从而获得了更加优秀的预测效果。

The item modeling of cyberspace refers to using representation learning strategy to encode the complex information of items into vector form that can be processed by computer. High quality item modeling is a foundation of the accomplishment of  various related tasks in cyberspace, such as computational advertising, e-commerce, etc. The item information in cyberspace usually contains multi-modal, multi-structured attributes. There are complex structural relations between items. Graph method has full flexibility in modeling complex structure relations, which is a great advantage in item modeling in cyberspace. Therefore, the item modeling based on graph methods has high significance and value in theory and application. This thesis discusses how to use graph methods to mining the attribute and relationship information of items. It focuses on three topics: research on deep graph models, item attribute modeling, item relationship modeling. The main contributions of the thesis are summarized as follows:

1. Research on deep dynamic graph model

The structure of the item graph in cyberspace is complex and changeable. Most of the research on graph method focuses on the mining of the structural characteristics of static graphs. However, the item characteristics of cyberspace are changeable, and the representation of items is changing with the frequent interaction with users and merchants. There are few works focusing on dynamic characteristics of graph structure. In this thesis, the idea of representation and propagation in graph convolution neural networks is fully utilized and extended to dynamic graphs and designs. A dynamic graph convolution neural network with only a small number of local nodes update and no fine-tuning training process is designed. The efficiency of node representation is greatly improved on the premise of ensuring the performance of node representation.

Based on the study of dynamic graph representation, learning the dynamic patterns of structure change further can obtain more effective information for modeling dynamic characteristics of items. Traditional dynamic graph link prediction models always rely on dynamic representation methods to model dynamics of graphs. However, because of the structural loss of graph representation learning itself and the alignment deviation between different time slices, this kind of method has inevitable dynamic cumulative error. This thesis directly models the mapping relationship of new edges, and stores historical emerging links by memory mechanism, which is a reference for dynamic graph link prediction. The method avoids high frequency node representation updating and further improves the efficiency of dynamic graphs.

2. Graph based item attribute modeling

The graph structure exists in almost all aspects of the attributes of the cyberspace items. How to extract the representation of items from complex attributes is the key to modeling the item attribute in cyberspace. Traditional modeling methods of item attributes usually regard these attribute features as auxiliary information to enhance the representation of items. The attributes are roughly integrated together, lacking structure modeling. In this paper, a disentangled item representation model based on bipartite graph matching is designed to adaptively assign the attribute features to each item that they may belong to. The method directly uses the combination of attributes to represent items. It not only enriches the feature dimension of items, but also improves the efficiency of item representation. The representation strategy can be directly applied to various downstream tasks based on item representation learning. In this framework, downstream tasks could focus on different aspects of item attributes, and improve the effectiveness of tasks such as the recommendation system.

Meanwhile, items show different states at different times. It is very important to model the sequential attribute representation for items. From the perspective of life cycle detection, the dynamic properties of items change with time are abstracted into directed graph structures of  different life cycle state transition. The life cycle state allocation process can be transformed into the shortest path search problem in the state transition directed graph. Through this strategy, the life cycle state of the item can be discriminated adaptively, and  the item representation in the sequential scene can be modeled more precisely.

3. Graph based item relationship modeling

There are various kinds of interaction relations among different items, such as the co-purchase relationship between commodities, collocation relations of clothing and the logical sequence relationship between the software applications, etc. Some traditional models ignore the modeling of these complex structures, and some completely separate the relations between items and the attribute features of items, which fail to fully model the interaction between item characteristics. In this paper, the graph neural network is used to model the overall interaction between a set of items. Through graph neural networks, item information can interact on the basis of the relationship structure, thus obtaining more holistic representation features. This thesis applies this method to the task of clothing collocation recommendation, and makes use of the association among matching sets to build a graph, and fully models the interaction of clothing products to obtain excellent recommendation results.

There are also some logical relations between different items in the sequence of user interaction. According to this sequence, the composition reflects the dynamic transfer relationship between items under the global environment. It is of great value to fully grasp such transition relationships for the recommendation system. Referring to the idea of dynamic graph link prediction model, this paper records and arranges the sequential mapping relationship of history to assist the future link prediction. In order to enhance the use of local structure relations, this paper uses the motif structure to extract local sequential features and thus obtains better prediction results.