1. Improving Revisitation in Graphs through Static Spatial Features Presented by PourangIrani University of Manitoba SohaibGhaniPurdue University West Lafayette, IN, USA NiklasElmqvistPurdue University West Lafayette, IN, USA Graphics Interface 2011 May 25-27, 2011 ▪ St. John’s Newfoundland, Canada
4. Memorability & Revisitation Memorability The memorability of a visual space is a measure of a user’s ability to remember information about the space Revisitation Revisitation is the task of remembering where objects in the visual space are located and how they can be reached
5. Motivation Graphsprevalent in many information tasks Social network analysis (Facebook, LinkedIn, Myspace) Road networks and migration patterns Network topology design Graphs often visualized as node-link diagrams Node-link diagrams have few spatial features Low memorability Difficult to remember for revisitation Research questions How to improve graph memorability? How to improve graph revisitation performance?
6. Example: Social Network Analysis Interviewed two social scientists who use graphs for Social Network Analysis(SNA) Often experience trouble in orienting themselves in a social network when returning to previously studied network At least 50% of all navigation in SNA in previously visited parts of a graph
7. + People remember locations in visual spaces usingspatial features and landmarks Geographical maps have many spatial features and are easy to remember Evaluate whether staticspatial features to node-link diagrams help in graph revisitation Inspired by geographic maps Idea: Spatial Features in NL Diagrams?
8. Design Space:Static Spatial Graph Features Three different techniques of adding static spatial features to graphs Substrate Encoding (SE) Node Encoding (NE) Virtual Landmarks (LM) But which technique is optimal?
9. Substrate Encoding Idea: Add visual features to substrate (canvas) Partitioning of the space into regions Space-driven: split into regions of equal size Detail-driven: split into regions with equal numbers of items Encoding identity into each region Color Textures Figure 1 Figure 2
10. Node Encoding Idea: Encode spatial position into the nodes (and potentially the edges) of a graph Available graphical variables: Node Size Node Shape Node Color
11. Virtual Landmarks Idea: Add visual landmarks as static reference points that can be used for orientation Landmarks Discrete objects Evenly distributed invisual space
12. User Studies Experimental Platform Node-link graph viewer in Java Overview and detail windows Participants:16 paid participants per study Task:Graph revisitation Phase I: Learning Phase II: Revisitation
13. Phase I: Learning N blinking nodes shown in sequence, Participants visit and learn their positions.
17. Study 1: Substrate Encoding Study Design: Partitioning: Grid and Voronoi Diagram. Identity Encoding: Color and Texture Layout: Uniform and Clustered Hypotheses: Voronoi diagram will be faster and more accurate than grid for spatial partitioning Texture will be more accurate than color for identity encoding
20. Study 2: Node Encoding Study Design: 3 Node Encoding techniques: Size, Color and Size+Color Hypothesis: Size and color combined will be the best node encoding technique in terms of both time and accuracy
23. Study 3: Combinations Best techniques from Study 1 (Grid with Color) and Study 2 (Size+Color) as well as virtual landmarks Study Design: Eight different techniques: SE, NE, LM,SE+NE, SE+LM, NE+LM, SE+NE+LM, and simple graph (SG) Hypotheses: Techniques utilizing substrate encoding will be faster and more accurate than node encoding and landmarks The combination of all three spatial graph feature techniques will be fastest and most accurate
26. Study 3: Results (cont’d) Techniques with substrate encoding significantly faster and not less accurate. SE+NE+LM not significantly faster and more accurate than all other techniques Virtual landmarks promising strategy, performing second only to substrate encoding
27. Summary Substrate encoding (SE) is dominant strategy Space-driven partitioning Solid color encoding Virtual landmarks (LM) help significantly Node encoding (NE) not as good other two Combination of virtual landmarks (LM) and substrate encoding (SE) is optimal
28. Conclusion Explored design space of adding static spatial features to graphs Performed three user studies Study 1: grid with color is optimal substrate encoding Study 2: node size and color is optimal node encoding Study 3:substrate encoding, landmarks, and their combination are optimal techniques
The basic idea of this work is to improve navigation in node-link representations of graphs by adding spatial features to the graph, similar to the geographical features in a map.In the next few slides, I will describe in more detail how we achieve this, and I will also present results from our user studies where we evaluated the efficiency of this idea.
Memorability is thus closely linked to revisitation.
Our work is motivated by collaborations with social scientists who use visualization tools for social network analysis (SNA).we performed structured interviews with two of our collaborators both faculty members at our university and SNA experts.We observed that these scientists would often experience some trouble orienting themselves when returning to a previously studied social network. Moreover, ad hoc observations of social scientists performing SNA showed that more than 50% of all navigations in a node-link diagram was between previously visited parts of a graph.
Based on our survey of the literature, we study three different classes of static spatial graph features: substrate encoding, node encoding, and virtual landmarks.
Substrate encoding mimics geographical maps by adding graphical features to the visual representation of the graph. In a map, thesefeatures are typically spatial regions, such as roads, city limits, state lines, etc. The regions themselves are generally identifiable throughunique colors or textures. The features can then be used as reference points.We identify two degrees of freedom for substrate encoding: the partitioning of the space into regions, and the encoding of identity into each region to allow the user to separate them.The advantage of a detail-driven approach is that if nodes are clustered in a small area of the whole graph, then we will allocate more partitions in that area. For uniform partitioning, a majority of the nodes may end up in the same partition.Figure 1 shows detail driven partitioning with color encoding and figure 2 shows space driven partitioning with texture encoding.
This approach has the advantage of not introducing a high degree of visual clutter. However, some of these graphical variables may already be utilized to convey underlying information about the data in many existing graph visualizations.Figure shows example of node encoding where node color is varied on x-axis and node size on y-axis.
The basic idea with virtual landmarks mimics the role of landmarks in the real world—they serve as static reference points that can beused for orientation (e.g., the Eiffel tower in Paris). Landmarks typically give rise to less visual clutter than substrate and node encoding techniques without affecting the visual representation of the graph itself.We used 9 virtual landmarks of different shapes as shown in figure.
we include an overview of the visual space so that the visual space could be larger than the screen, preventing participants from remembering nodes by absolute positions on the screen rather than by spatial features. Furthermore, the overview was scaled down to a factor of about 10, making it difficult for participants to remember nodes using just the overview.
Figure shows first node blinking in red color. User will learn its position and click on it than second node start blinking.
Figure shows 2nd node blinking in red color. In this way N blinking nodes are shown to the user.In learning phase we use N=4 for first two studies and N=5 for third study. We increase N in 3rd study so that we get a separation between the techniques.
After the learning phase participants were asked to revisit the learned node in the same as order as before.
In this way user will revisit N nodes learned in the learning phase.
A regular grid is the simplest partitioning technique for equal-sized regions. We use a 3*3 grid to divide the space into the 9 regions (derived by pilot study).Partitioning space into regions with equal numbers of items requires us to group the graph nodes into 9 disjoint clusters. We then use a Voronoi diagram, summing up the cells for node in each cluster, to find the regions covered by these nodes. This yields an irregular partitioning focused on areas of high detail.we used two separate layouts: one yields uniform node distribution with uniform edge lengths , and the second clusters similar nodes based on the graph topology.
Color and Textures are used to encode each region. A solid color is chosen as it is the straightforward way to differentiate between regions. A texture will yield more internal detail to a region, potentially increasing its memorability. However, texture will increase visual clutter as well.
Space-driven partitioning using a grid yields significantly faster and more accurate performance than detail-driven partitioning using a Voronoi diagram.Encoding regions using a solid color yields significantly faster performance, with no significant difference in errors, than encoding using a texture.There is no significant effect of graph layout on completion time or accuracy.
In the second study, the spatial position of nodes was encoded in their size and color.
We use 3 approaches for node encoding. In first approach size of the nodes are varied such that width is varied on x-axis and height on Y axis. In second approach Color of nodes is varied such that Hue is varied on X-axis and Brightness on Y-axis. In 3rd approach both size and color are varied on Y and X-axis respectively.
The combination of size and color for encoding position is both significantly faster and more accurate than each of these techniques separately.There was no significant difference between size and color alone.
Figure shows significant pairwise differences in completion time. Arrows indicate which technique was faster than another. Results suggests that Substrate encoding and landmarks are best approaches for graph revisitation. Both of these techniques especially in combination performed significantly better than the competing techniques. Node encoding seems not to make much difference either way, which is perhaps why the combination of all three approaches is good, but not significantly better than others