PP: Overviewing evolution patterns of egocentric networks by interactive construction of spatial layouts

Problem:

get an overall picture of how ego-networks evolve is a common challenging task.

Existing techniques: inspect the evolution patterns of ego-networks one after another.

Purpose:

how analysts can gain insights into the overall evolution patterns of ego-networks by interactively creating different spatial layouts.

Introduction:

1. What are ego-network and ego-network analysis?

The analysis of individuals in a network context is referred to as egocentric network analysis or ego-network analysis. An ego-network consists of a focal node, the nodes within its one-step neighbourhood and all the edges among these nodes

2. the content in a spatial layout

each dot represents a dynamic ego-network, clusters of dots indicate similar evolution patterns./ outlying dots exhibit uncommon evolution patterns.

3. interpretability and interactivity.

This technique is developed with interpretability and interactivity in mind

Related work:

1. ego-network visualization

i. Most of them focus on visualizing individual ego networks rather than revealing the overall evolution patterns;

ii. tree-ring layout.

iii. ...

2. dynamic network visualization

i. Two major approaches to analyze network evolution are animation and timeline.

• animation: Animation-based technique uses animated transition of visual elements (e.g., nodes and edges in a node-link diagram) to reveal the time dimension. An obvious drawback is that it is cognitively demanding to keep track of the changes.
• Timeline-based approaches, on the other hand, use small multiples (e.g., [6]), vertical or horizontal timeline (e.g., [24]) and circular layout (e.g., [51]) to represent the time dimension. However, as noted by Wu et al. these techniques mainly focus on tracking changes of the entire network rather than the characteristics of ego-networks.

3. techniques for creating spatial layouts for sensemaking

i.

Methodology:

1. 两个要素: interpretability and interactivity. 可解释性和可交互性
2. data model: 142 dynamic ego-networks for 24 months, and generated time series from these dynamic ego-networks( derived from node attributes---CEO,President,Vice President..., derived from network structure---size, density.)
3. data transformation pipeline. 
   1. time series -----> event sequences: input time series and event type, output extracted point/interval events.
   2. event sequences -----> feature vectors: a feature vector records the number of happened events E = {e1,e2,e3,e4...}.
   3. feature vectors -----> distance matrix: pairwise distance.
   4. distance matrix -----> spatial layout: use MDS to project distance matrix onto a spatial layout. Others: force-directed MDS/ t-SNE, slower and less scalable. Spatial layouts are often generated by dimensionality reduction techniques (e.g., PCA [28], MDS [49] and t-SNE [37])
   5. 评论, 由于是从ego-networks中抽取的特征作为time-series, 而又从time series中抽取events, 这一步当中虽然event记录了时间发生开始和结束,但是在step3转化为了feature vectors, 记录的是事件发生的次数,不包含时间发生顺序. 但是原文也提到,step2 and step3 can be replaced by other methods. ??为什么不直接对time series进行距离计算,这样更能发现两个dynamic ego-network之间的evolution 是否相似.
4. The spatial layout reveals the evolution patterns. Each dot in spatial layouts presents a dynamic ego-network (24 ego-networks of one indivisual). If two ego-networks share similar evolution patterns, they will have similar number of events of the same type, thereby pulling them closer together in the spatial layout.

User interface:

1. conducted a formative evaluation of the initial prototype with two experts. Interviewing two experts for about an hour.

Supplementary knowledge:

1. Enron email network dataset: 142 employees. Each individual has a dynamic ego-network. An ego-network snapshot depicts the email communication of an employee with other employees in a given month. The data set spans 24 months. So there are 142 * 24 ego-networks/ 142 dynamic ego-networks, each having 24 snapshots.