Dynamic attribute graph visual analysis method based on embedded community comparison

Abstract

The invention discloses a dynamic attribute graph visual analysis method based on embedded community comparison. According to the method, an embedded community comparison strategy is utilized, and a plurality of exploration tasks related to three aspects (namely the structure, the attribute and the time) in a dynamic attribute graph can be supported in a consistent mode. According to the main technical scheme, the visual framework for exploring the dynamic attribute graph comprises a core step, namely community comparison, for realizing the framework based on an embedding method, a dynamic attribute graph exploring workflow comprising three steps, a visual system for realizing the three-step workflow, and a visual system for exploring the dynamic attribute graph. The method comprises three components: a graph overview component, an embedded projection component and a mode analysis component. According to the method, analysts can be helped to quickly and accurately identify the modes in the dynamic attribute graph, and the overall exploration efficiency is improved. The method has the advantages of high efficiency and good expandability, and is suitable for common dynamic graph and dynamic attribute graph data.

Description

Technical Field

[0001] The present invention mainly relates to the fields of visualization and visual analysis, and in particular to an interactive visual analysis method that supports efficient exploration of dynamic attribute graphs. Background Art

[0002] Dynamic graph exploration is a classic research topic in the field of visualization. Existing work usually combines visualization with machine learning algorithms to analyze the temporal evolution of graphs at different levels, including node, community, and network levels. The complex data structure and analysis tasks have led researchers to generally focus only on the structural level. However, dynamic graph data in the real world usually has multiple attributes and the values ​​of nodes on these attributes change over time along with the structure, which is called a dynamic attribute graph. For example, social networks, transportation networks, communication networks, and global refugee migration networks. The multifaceted features contained in dynamic attribute graphs will form richer tasks, which makes it practical and meaningful to propose a visualization method to analyze them.

[0003] However, it is much more difficult to explore dynamic attribute graphs than dynamic graphs. First, without considering the node attributes, this paper can regard a dynamic attribute graph as a dynamic graph. Exploring dynamic graphs is itself a complex and ongoing research topic. Second, multidimensional attributes and structure influence and evolve together, resulting in richer tasks. For example, analysts can explore (1) how the structure and attributes evolve separately and how these two aspects influence each other at different times (evolutionary analysis) and (2) whether there are nodes with close connections but significantly different attribute values ​​or nodes with highly similar attribute values ​​but sparse connections (anomaly detection). For example, in a dynamic attribute graph formed by refugee migration, analysts can explore how the migration of refugees and the attributes of countries evolve separately and whether there is a correlation between their evolution. In addition, they can also explore whether there are countries with significantly different attribute values ​​(such as GDP, population, etc.) among countries with close refugee exchanges. The rich exploration tasks bring challenges to the design of visualization methods. It is impossible for visualization methods to design independent analysis methods and interaction methods for each task, which will greatly increase the learning burden and operation difficulty of analysts, thereby affecting the usability of the method.

[0004] Therefore, it is challenging to design a visualization method to achieve the full range of exploration tasks involved in dynamic property graphs in a consistent manner. Summary of the invention

[0005] In view of the shortcomings of existing methods and the above challenges, in order to support analysts to complete rich exploration tasks on dynamic attribute graphs with consistent analysis methods and interactive operations, the technical solution of the present invention is a dynamic attribute graph visual analysis method based on embedded community comparison. The details are as follows:

[0006] 1) First, a visualization framework for exploring dynamic property graphs is proposed, whose core idea is to achieve all tasks through a consistent interactive method. The framework consists of two steps. First, according to the structure and attribute values ​​of the graph, the analyst divides the static property graph at each time point into two groups of communities (community partitioning). Second, the analyst can select two groups of communities for comparison, and the changes in the affiliation of nodes in the two groups of communities can reveal rich patterns (community comparison). It is worth noting that there are many ways to select two groups of communities for comparison, and each way corresponds to an exploration task. Therefore, rich tasks can be derived based on this framework and these tasks can be consistently achieved through these two steps.

[0007] 2) Next, this paper proposes an embedding-based method to implement the core step of the framework - community comparison. The core idea is to generate an embedding for each node to represent the change in its community affiliation. To this end, this paper takes the following steps: (1) The community number of the node in the first group of communities is used as the encoding of the first group of communities (a label); (2) The embedding of the node in the second group of communities is learned as the encoding of the second group of communities (a vector); (3) The two community encodings of the node are fused into a single embedding, reflecting its affiliation in the two groups of communities. By projecting the embeddings of all nodes, analysts can explore patterns in the projection. The color and position of the node in the projection can help analysts infer its affiliation in the two groups of communities, thereby improving the interpretability of the patterns in the projection.

[0008] 3) Finally, this paper integrates the above methods into a visualization system to implement the framework. The system follows a three-step workflow and integrates three components (graph overview, embedding projection, and pattern analysis) to address design challenges. First, the graph overview component provides an overview of the dynamic property graph, showing its overall evolution trend, and provides intuitive visualization cues for dividing and selecting communities. Second, the embedding projection component displays the projection formed by the fused node embeddings, allowing analysts to explore various patterns in a consistent manner. Finally, the pattern analysis component can help analysts understand the reasons for the formation of identified patterns. Overall, these integrated visualization methods and quantitative indicators provide analysts with intuitive references, simplify their operations, and increase their confidence in the results.

[0009] In summary, this paper has accomplished the following work: (1) a visualization framework that can consistently support rich exploration tasks, (2) an embedding-based community comparison method, and (3) a visualization system that supports efficient exploration of dynamic property graphs.

[0010] Compared with the prior art, the advantages of the present invention are:

[0011] 1) The present invention proposes an embedding-based community comparison method, which supports analysts to compare and analyze graphs at different times in a projection manner.

[0012] Identify patterns and analyze graph evolution based on notable visualizations in the projection.

[0013] 2) The method proposed in the present invention has good scalability and takes the community as the analysis object, thereby avoiding paying too much attention to low-level objects (such as nodes and edges).

[0014] 3) The visualization system designed by the present invention can support analysts to efficiently explore dynamic attribute graphs, quickly and accurately identify patterns and understand the reasons why patterns are formed. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The specific embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings:

[0016] Figure 1 It is a common graph pattern in dynamic property graphs;

[0017] Figure 2 It is a representation network that achieves efficient community comparison;

[0018] Figure 3 It is a three-step exploratory workflow;

[0019] Figure 4 It is a visual system interface diagram; DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present invention will be fully described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0021] The implementation of the invention is specifically divided into three steps, namely, dynamic property graph task and pattern summary, embedded community comparison, and the design of a visual analysis system for efficient exploration of dynamic property graphs.

[0022] Step 1: First, systematically summarize the tasks and patterns in the dynamic property graph.

[0023] Exploration task: The visualization framework is divided into two steps, community division and community comparison. Analysts can select two groups of communities for comparison, and the differences between them can reveal potential graph patterns. Therefore, the present invention identifies four ways to select communities for comparison, each of which corresponds to an exploration task. There are four ways in total: (1) Select two groups of communities divided according to edge sets of different static property graphs for comparison to explore structural evolution; (2) Select two groups of communities divided according to attribute matrices of different static property graphs for comparison to explore attribute evolution; (3) Select two groups of communities divided according to the attribute matrix and edge set of the same static property graph for comparison to explore structural anomalies; (4) Select two groups of communities divided according to the edge set and attribute matrix of the same static property graph for comparison to explore attribute anomalies.

[0024] Common graph patterns: This paper summarizes six graph patterns based on the characteristics of dynamic attribute graphs, namely stable, growing, decreasing, merging, splitting, and abnormal patterns. Through these graph patterns, analysts can have a deeper understanding of the evolution laws and specific patterns of dynamic attribute graphs. Common graph patterns are as follows: Figure 1 As shown. Stability means that the number of nodes in a community remains unchanged. Growth means that the number of nodes in a community increases. Decrease means that the number of nodes in a community decreases. Merge means that multiple small communities merge into a large community. Split means that a large community splits into multiple small communities. Anomaly means that there are very few nodes in a community that do not belong to another community.

[0025] Step 2: In traditional community comparison, analysts need to observe the two groups of communities separately and repeatedly switch the analysis focus to identify differences. This analysis method makes it difficult to identify patterns, especially when there are many nodes or the differences between communities are not significant. To solve this problem, this paper proposes an embedding-based community comparison method, whose core goal is to generate a set of embeddings that can represent the nodes in the two groups of communities. Analysts can efficiently explore patterns in the projections of these embeddings.

[0026] Community encoding: First, encode each node in the two contrasting communities. This step will generate two community codes for each node, reflecting its community affiliation in the two communities. First, use the community detection algorithm to generate a set of labels for all nodes in a certain set of communities based on the edge set (or attribute matrix), and use the node labels as the encoding of the first set of communities. Second, generate a set of embedding vectors for all nodes in the other set of communities based on the edge set (or attribute matrix), and use the embedding vectors as the encoding of the second set of communities.

[0027] Code fusion: Design a representation network to fuse the two codes of each node into a single embedding. The present invention uses a dual subnet structure to implement the network. Figure 2The network structure is shown, where f and g are two subnetworks for processing the first encoding and the second encoding respectively. Specifically, the two subnetworks take the fused embedding of the node as input, f is used to predict its label (the first encoding), and g is used to predict its vector (the second encoding).

[0028] Analysis process: Project the fused embedding of all nodes and color each node according to its first code (label). The projection provides two key visual clues to help analysts infer the nodes' affiliation in the two groups of communities. First, the color of the node already reflects its affiliation in the first group of communities. Second, the nodes that form clusters in the projection should belong to the same subcommunity in the two groups of communities. Therefore, analysts can (1) focus on the visual structures in the projection, such as clusters or outliers, (2) understand the nodes' affiliation in the first group of communities based on the node colors, and (3) infer the nodes' affiliation in the second group of communities based on the first two pieces of information.

[0029] Step 3: The goal of the visualization system is to implement all tasks in a consistent manner and provide necessary guidance to analysts. When implementing the system, the present invention overcomes the following three challenges.

[0030] 1) Intuitive prompts. The system should provide necessary and intuitive visual prompts when analysts divide and select communities. Slight changes in the parameters of the community detection algorithm may lead to significant fluctuations in the community division results. In addition, it is not feasible to propose a parameter optimization method that is applicable to any data set. Therefore, the system should help analysts set reasonable algorithm parameters to obtain stable communities. Dynamic property graphs usually contain a large number of community groups, and analysts can select them for pairwise comparisons, while many community groups do not involve patterns, which will lead to a lengthy analysis process. Therefore, the system should provide a high-level overview of the dynamic property graph to help analysts select community groups involving patterns and improve exploration efficiency.

[0031] 2) Design consistency. The system should display a large number of objects, namely static property graphs, communities, and nodes. The visualization style and visualization language of the interface should be consistent. Specifically, this paper should ensure that the colors, fonts, labels, highlights, and other marks in all components have a consistent style. In addition, similar visualization elements should have similar meanings no matter which component they appear in. The system should support all exploration tasks and graph modes. It is not feasible to design independent interaction methods for each task and graph mode, which will greatly increase the memory burden and operation difficulty of analysts. Therefore, the system should have a consistent interaction method to support all tasks and modes.

[0032] 3) Comprehensiveness of analysis. The system should support comprehensive analysis of the projections formed by the fused embeddings. The system should help analysts identify patterns associated with specific visualization structures. Specifically, it should answer the following research questions: (1) Which visualization structures may involve patterns? (2) How do the community affiliations of the nodes they contain change? (3) What patterns do these visualization structures imply? The system should help analysts reveal how patterns are generated. Specifically, it should answer the following research questions: (1) Which nodes are involved in the visualization structure? (2) Which nodes contribute more to the formation of the pattern? (3) What structural and attribute characteristics do specific nodes have?

[0033] The system follows a 3-step exploratory workflow, such as Figure 3 As shown. First, analysts observe the evolution overview of the graph and select two groups of communities as the starting point for analysis. The representation network will generate projections based on the community encodings of the nodes. Then, they can quickly locate the visual structures related to the pattern from the projections and determine the pattern based on its visualization effect. Finally, they can analyze the reasons for the formation of the identified pattern through inter-group or intra-group patterns. Three components, namely graph overview, embedded projection, and pattern analysis, are integrated in the visualization system to implement these three steps respectively. The visualization system interface is shown as follows Figure 4 shown.

[0034] Figure 1.1 Overview of components. Figure 4 As shown in (a), the graph overview component divides the dynamic attribute graph into several community groups based on the structure and attribute values ​​in each static attribute graph, and supports analysts to select two groups of communities for comparison. It is worth noting that this component provides visual prompts of community quality and overall evolution for all divided communities to help analysts divide and select communities.

[0035] Embeds the projection component. Figure 4 As shown in (b), two encodings are generated for each node, reflecting its community affiliation in the two selected groups of communities. The two encodings of each node are then fused into an embedding, and the embeddings of all nodes are projected. Analysts can observe the projection and select a visualization structure to analyze the underlying patterns. It is particularly important to note that the component should follow a consistent visualization style and visualization language to implement the interface and provide a consistent interaction mechanism to support all tasks and modes.

[0036] Pattern analysis components. Figure 4As shown in (c), the pattern analyzer component is used to mine the patterns involved in the visualization structure selected by the analyst from the projection. It is worth noting that this component supports a comprehensive analysis of the selected visualization structure. First, based on the changes in the nodes' belonging to the two groups of communities inferred by the analysts, this component can help them understand and identify the patterns in the visualization structure. Second, this component can help analysts further analyze the reasons for the formation of the identified patterns.

Claims

1. A dynamic property graph visual analysis method based on embedded community comparison, whose goal is to support analysts to explore dynamic property graphs in a consistent analysis method and interactive operation. The specific contents are as follows: 1) We propose a visualization framework for exploring dynamic property graphs, which consists of two core steps (community partitioning and community comparison) and implements all tasks in a consistent manner. 2) Propose an embedding-based method to implement the core step of the framework - community comparison. In addition, the system summarizes common graph patterns and their visualization effects in dynamic property graphs, allowing analysts to identify various graph patterns through projection.

2. According to the method of dynamic attribute graph visual analysis based on embedded community comparison described in claim 1, a three-step exploration workflow is proposed. First, the analyst observes the evolution overview of the graph and selects two groups of communities as the starting point of the analysis. The representation network will generate projections based on the community encoding of the nodes. Then, they can quickly locate the visual structure related to the pattern from the projection and determine the pattern based on its visualization effect. Finally, they can analyze the reasons for the formation of the determined pattern through inter-group or intra-group patterns.

3. According to the method of dynamic attribute graph visual analysis based on embedded community comparison described in claim 1, a visualization system is designed to implement the above workflow. The system includes three components, namely, graph overview, embedded projection and pattern analyzer, to respectively implement the three steps in the exploration workflow. The graph overview component is used to select two groups of communities as the starting point for analysis. The embedded projection component supports analysts to explore the patterns contained in the projection. The pattern analyzer component helps analysts understand the reasons for the formation of the identified patterns.

Images