Network data often contain important attributes from various dimensions such as social affiliations and areas of expertise in a social network. If such attributes exhibit a tree structure, visualizing a compound graph consisting of tree and network structures becomes complicated. How to visually reveal patterns of a network over a tree has not been fully studied. In this paper, we propose a compound graph model, TreeNet, to support visualization and analysis of a network at multiple levels of aggregation over a tree. We also present a visualization design, TreeNetViz, to offer the multiscale and cross-scale exploration and interaction of a TreeNet graph. TreeNetViz uses a Radial, Space-Filling (RSF) visualization to represent the tree structure, a circle layout with novel optimization to show aggregated networks derived from TreeNet, and an edge bundling technique to reduce visual complexity. Our circular layout algorithm reduces both total edge-crossings and edge length and also considers hierarchical structure constraints and edge weight in a TreeNet graph. These experiments illustrate that the algorithm can reduce visual cluttering in TreeNet graphs. Our case study also shows that TreeNetViz has the potential to support the analysis of a compound graph by revealing multiscale and cross-scale network patterns.
Good morning, everyone. My name is Liang Gou, from Penn State. I am presenting our work today: TreeNetViz: Revealing Patterns of Networks over Tree Structures.
Here is an overview of today’s presentation. I will first introduce the problem we have. That is, how to visualize a TreeNet Graph. I will talk about TreeNet graph later. Then I will present our work on TreeNetViz to reveal the network patterns over a tree, include the design of treenetviz and an algorithm to reduce visual complexity in TreeNetViz. Later, I will talk about a case study on a co-authoring social network. Finally, we will see some future directions about this work.
I’ll start with the data we are addressing: A compound graph with a network and a tree is seen in many domain. Here is an example: We have a co-authoring network. The links are the collaborations among authors about papers. And also, the authors has their affiliation information, like univ., country. This affiliation information is actually a tree structure like this. Here we define a TreeNet graph as a compound graph consisting of two sub-graphs of a tree and a network along with a node-mapping schema . Node that, network nodes are linked to leaf nodes of a tree. They are labeled with the same number in the this figure.
Our problem here is how to visualize and explore the patterns of a network over a tree in a treenet graph. Here we are interested in the mutual influence between two sub-graphs of tree and nodes: For example, we have some individual network nodes, which are isolated and the degrees are same. If we put a tree information on them, then network structure is changed. They are fully connected and the degrees are changed. We also want to show the relationship among entities at different levels. This includes multiscale relationships, such as links between individual and individual, university and university, or country and country. It also includes cross-scale relationships from different levels, such as connections between individual and university, individual and country, or university and country.
We need aggregate the network over trees to get these relationships An aggregated network can be generated by applying a cut on a tree to specify nodes of interest, then aggregating edges based on the nodes in the cut.
Here is an example. This is a tree we are working on. We can apply this cut to it. It specifies these nodes in the cut what we are interested in. So we got this kind of sub-tree, with nodes D and C collapsed but nodes ABEF expanded. This is what the original network looks like. Then we aggregate the network based on this sub-tree, we have this view. Basically, we aggregate nodes under C and D, and also, we can aggregate the edges among them based on this sub-tree. This is the basic process. Here we need to design a visualization to support this.
So our goal in TreeNetViz is to design interactive visualization so that: It can provide visual r epresentations of both tree and network structures with an integrated view; It should support interactive aggregation of network over tree structures; It should allow exploration of the patterns of the aggregated networks at different tree levels;
Let’s take a look at the actual design. For tree, we use Radial Space Filling (RSF) technique to show the tree. Traditional node-link tree is very intuitive, but here we use a RSF to not only show the parent-child relationship, but also leverage the node space. More important, we can place an aggregated network over a RSF tree.
For the network, we use a circular layout to show node connections. Network nodes are assigned to a circle and links are created to connect them. Then, We can put this circular layout on a RSF tree to integrate both tree and network structure in a single representation. For an aggregated network, this is what we have in the last example, this approach can easily to expand and collapse nodes to show this network. One drawback of using the straight line for edges is that some edges are blocked by node sectors and we see some cluttering. To alleviate this problem, We use Hierarchical Edge Bundling from Holten to bundle the network links. This is our visualization design. Let’s move to interaction part.
Let look at the interaction design. First, we can have multiscale view by controlling the scale level of interest. This is at individual level, the second level and the top level. Users can have a cross-scale view. Users can simply expand or collapse a node. In this pic, we can see that t Ctry C and UnivD are collapsed and others are expanded.
We have some other interactions: like ego-network view shows the local network structure, A critical path view show the shortest path between two nodes.
One issue in the TreeNetViz is how to place the nodes of any aggregated network along with the circle. In this figure with random placement, you can see there is a lot of visual cluttering. We propose a method, called Hierarchy-awareness Weighted Circular Layout. The idea is to order nodes along the circle to reduce visual complexity. It also considers the restriction of a tree structure, which means that Child nodes are sorted under their parent node, it also considers the weight of edge and try to avoid visual cluttering with high weight edges. We use aggregated weight of edges, including the sum of edge number and edge level.
The Key idea here is to optimize node positions to achieve two criteria: We want to fewer edge crossings and Less total edge length. Thus, we used a heuristic approach to try different combinations of node order by switching node pairs. This is an example to reduce edge crossings. We have 3 crossings here and only have 1 after switching. Here we can also reduce total edge length. We have edge length of 3 hops and only have 2 hops after switching.
The two pictures shows the visual improvement by our algorithm HWCL. The left one is a random layout without HWCL and The right now is the optimized layout by HWCL. we can the representation is simple and clean with fewer edge crossings.
We did several experiments to study how different factors impact the perforamcne of the layout. Here is one experiment result which shows the convergence of HWCL algorithm. The x axis is the node number of different graphs and the y axis is the cost which is relative to random layout. We can see that our method can converge very quickly, after round 3 or 4 the cost becomes stable.
We did a case study with a co-authoring social network to exam the usefulness of our design. We want to use our design to Help understand collaboration patterns among diabetes researchers across a Research-I university. The dataset includes 847 authors and around 2,500 links. We identify a social affiliation hierarchy with 10 colleges/schools and 90 departments and centers.
Let’s first take a look at multiscale exploration. Here is a figure at the scale of college. It shows the collaboration patterns among colleges. We can see medical school is a hub connecting with other colleges. There are strong connections among medical school and LAS and school of public health.
When we get the second level of department, we have a more detailed view. It shows general patterns of collaboration among Departments.
We continue to drip down to the scale of individual. We can have an overview of individual collaborations.
On the other hand, some time we don’t need too much details, so we can have cross-scale view and only focus on the organization of interest. Here is an example showing the connection patterns of departments only in LSA with other colleges.
We can even drill down one more level to show the connections across three levels of individual, department and college. This is helpful to see how people connect with each other inside an organization and out side the organization. Of course, we can open more nodes to see in details.
Another useful information we can have with cross-scale view is to find Potential Research Broker between Two Colleges. In this figure, a critical path is highlighted, we can see these two colleges connected by School of Public Health. We can expand this school, and find it’s this department connecting these two schools. Finally, we can find this broker at the last level.
In the future, we are going to extend this work by enabling users to freely modify a tree structure. Besides, we plan to evaluate TreeNetViz with other techniques, include Node-link diagrams, such as ask-graph, or matrix view such as HoneyComb.
Hierarchical Edge Bundling (HEB) [7]
Hierarchical Edge Bundling (HEB) [7]
. The costs of edge crossings and total length with =0, 0.5 and 1: (L). Relative cost of crossings (R) Relative cost of length.
the edge weight has larger impact to reduce the final cost compared with edge level