See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/283171277

egoSlider: Visual Analysis of Egocentric Network Evolution

Article in IEEE Transactions on Visualization and Computer Graphics · November 2015

DOI: 10.1109/TVCG.2015.2468151

CITATIONS READS
96 872

6 authors, including:

Jian Zhao Guowei Huang

FX Palo Alto Laboratory Huawei Technologies
42 PUBLICATIONS 1,363 CITATIONS 2 PUBLICATIONS 116 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Jian Zhao on 17 October 2016.

The user has requested enhancement of the downloaded file.

Author manuscript, published in IEEE Transactions on Visualization and Computer Graphics, 22(1), pp. 260-269, 2016.
DOI: 10.1109/TVCG.2015.2468151

egoSlider: Visual Analysis of Egocentric Network Evolution

Yanhong Wu, Naveen Pitipornvivat, Jian Zhao, Sixiao Yang, Guowei Huang, and Huamin Qu

Jian Sun

Baining Guo

Kun Zhou

a
Fig. 1. Exploring the ego-network evolutionary histories of several prestigious researchers in the fields of computer graphics, computer
vision, and visualization based on the DBLP collaboration network data. General patterns of all researchers’ ego-networks (such as
clusters and outliers) are revealed in a) the overview. Overall variations of one’s ego-network characteristics (such as alter numbers
and densities) can be learned by viewing the snapshot glyph and transition glyphs in b) the timeline-based visualization. Detailed
structures of alters in the ego-networks (such as different connected components) and their temporal relationship information with the
ego are further visualized in c) the expanded timeline view. For example, we can see that Harry Shum’s ego-network grew significantly
from 2001 to 2005 (from the circle sizes), and meanwhile his collaborators converged from multiple connected components into one
after 2004 (when he became a Managing Director at MSRA). Three of his long-term collaborators can also be identified.

Abstract—Ego-network, which represents relationships between a specific individual, i.e., the ego, and people connected to it, i.e.,
alters, is a critical target to study in social network analysis. Evolutionary patterns of ego-networks along time provide huge insights to
many domains such as sociology, anthropology, and psychology. However, the analysis of dynamic ego-networks remains challenging
due to its complicated time-varying graph structures, for example: alters come and leave, ties grow stronger and fade away, and alter
communities merge and split. Most of the existing dynamic graph visualization techniques mainly focus on topological changes of the
entire network, which is not adequate for egocentric analytical tasks. In this paper, we present egoSlider, a visual analysis system
for exploring and comparing dynamic ego-networks. egoSlider provides a holistic picture of the data through multiple interactively
coordinated views, revealing ego-network evolutionary patterns at three different layers: a macroscopic level for summarizing the
entire ego-network data, a mesoscopic level for overviewing specific individuals’ ego-network evolutions, and a microscopic level
for displaying detailed temporal information of egos and their alters. We demonstrate the effectiveness of egoSlider with a usage
scenario with the DBLP publication records. Also, a controlled user study indicates that in general egoSlider outperforms a baseline
visualization of dynamic networks for completing egocentric analytical tasks.
Index Terms—Egocentric network, dynamic graph, network visualization, glyph-based design, visual analytics.

1 I NTRODUCTION
Nowadays, social network analysis has become an important approach sis that quantifies relations of a large group of people, and egocentric
for investigating information flows and people relationships in our analysis that studies dynamics between a specific individual, called
societies [7]. Sociologists, psychologists, and anthropologists have ego, and people connected to the ego, called alters. The latter type of
mainly focused on two aspects of social networks: sociocentric analy- networks, often named personal networks or ego-networks, indicates
how an individual is tied to an outside social world. Understanding
how such networks evolve over time can provide huge insights to mis-
• Y. Wu, N. Pitipornvivat, and H. Qu are with the Hong Kong University of cellaneous domains. For example, researchers can better comprehend
Science and Technology. E-mail: {ywubk, npab, huamin}@ust.hk
different communication behaviors in various online social spaces by
• J. Zhao is with Autodesk Research. E-mail: [email protected]
analyzing ego-networks [30]; medical experts find that one’s health
• S. Yang and G. Huang are with the Huawei Technologies Co. Ltd. E-mail:
{yangsixiao, huangguowei}@huawei.com
condition is strongly associated with many ego-network related factors
(e.g., friend degrees) [44]; analysts in management, business intelli-
Manuscript received 31 Mar. 2015; accepted 1 Aug. 2015; date of gence, and information security can make more informed decisions by
publication xx Aug. 2015; date of current version 25 Oct. 2015. identifying the most influencing people in social networks and how
For information on obtaining reprints of this article, please send they affect others along time [2, 18, 39].
e-mail to: [email protected].
 Social network researchers have developed extensive analytical connection density [2, 4, 24]. Romantic relationships between two
methods to measure and model various aspects of ego-networks, such people can be recognized based on what extent that their mutual
as statistical analysis, predicting ties, and detecting alter communities friends are well-connected [11]. Another approach focuses on the
[7, 11, 37, 42, 43]. However, few are able to capture the evolutionary relation type and strength between the ego and alters as well as their
patterns of ego-networks due to the highly dynamic and complex impacts to the ego. Studies show that the existence of stable ego-alter
nature, which requires human supervision in the process of exploration relationships plays a fundamental role in an individual’s life cycle
and analysis. Along an individual’s life, relational ties to alters inter- [25]. There are also some work targeting at the dynamic process of
twine: some relationships emerge and some fade away, some become relationship building, which attempts to answer questions related to
stronger and some turn distant [17]. Not only do the connections where, when, and how relations come into an ego-network [33, 36].
between the ego and alters vary in time, but also the dynamics among Prell described three major properties that are typically studied in
different alters. There exists a series of questions need to be addressed, ego-network analysis: the number of alters (degree), the strengths
for example, when and how new relations come into being, how the of ties connecting the ego and alters (closeness), and the number of
strengths of relations change over time and affect each other, and how interconnections between alters (transitivity) [45].
the alter community structures evolve during a timespan. Macroscopic level analysis looks at the overall patterns in a sub-
Although many dynamic graph visualizations have been proposed group of egos or the entire network. Many researchers have attempted
(e.g., [8, 9, 53]), they mainly focus on tracking changes of the entire to analyze structural properties of ego-networks. Arnaboldi et al.
graph rather than the characteristics of ego-networks. Some methods confirmed that certain attributes of online ego-networks appear to be
that attempt from the egocentric point of view (e.g., [28, 46, 50]) similar to those found ”offline”, including the number of alters and
merely visualize the ego-alter relationships, but omit the connection the relation strength distribution [7]. Network size was found having
strengths and inter-alter relationships, making it impossible to answer a large impact on compositional properties of the network and ego
some alter-related questions, such as insights about alter communities. characteristics [48]. Lubbers et al. suggested that the persistence
To address the above concerns, we propose an interactive visual- of ties was related to tie strength, network density, and other alter
ization system, called egoSlider, for exploring, comparing, and ana- attributes [41]. Moreover, there is some work that characterizes
lyzing ego-network evolution. egoSlider provides a holistic picture ego-networks into various categories, revealing different social com-
of the dataset through three major views, allowing users to browse munication patterns [14, 30]. Dynamic evolution of ego-networks has
ego-networks at various levels of scale. 1) An overview shows overall also been of great interests by researchers. For example, Arnaboldi
temporal patterns, such as clusters, of all individuals’ ego-networks et al. found that people in Twitter have highly dynamic ego-networks
in the database (Fig. 1-a), where users can further dive into different with a large percentage of weak ties and high turnover [6].
regions of interests. 2) A glyph-based timeline visualization sum- Quantitative measurements of graphs are also related to our work.
marizes one’s ego-network evolution with critical statistical features Most proposed metrics focus on static graphs (such as node degrees,
such as alter numbers, densities, and overall changes of ego-alter rela- betweenness, etc.). A few of them have been extended for measuring
tion strengths (Fig. 1-b). Comparative analysis of different people’s dynamic graphs, although they are not specialized for dynamic ego-
dynamic ego-networks can be easily achieved by viewing multiple networks. For example, time-scale degree centrality considers both
timelines. 3) A detailed view of a person’s ego-network timeline presence and duration of links [52]; and change centrality compares
can be further revealed for browsing dynamic structures of alters two graphs based on change events such as added, removed, and
in the ego-networks, such as tracing the relationship with a specific remained links [29].
alter along time, and discovering alter communities (Fig. 1-c). These The design and development of our egoSlider system stem from
three views are seamlessly coordinated with a rich set of interactions, the observations and analysis of the above literature. Based on these
supporting a smooth navigation of the dataset and multi-scope insight studies, we derive metrics that characterize ego-network evolution,
discovery in ego-network analysis. egoSlider also incorporates several and distill our research questions and design goals in Sec. 3.
analytical abilities to filter the data, extract ego-network features, and
conduct similarity measurement of different ego-networks. 2.2 Dynamic Network Visualization
Our main contributions in this paper include: Many visualization techniques have been proposed to address the dy-
• An interactive visualization system, named egoSlider, that en- namic evolving nature of network data. Major methods include using
ables users to explore, compare, and analyze dynamic ego- animated transitions, showing network variations along timelines, and
network evolutionary trends and patterns; a hybrid of the two. A comprehensive survey can be found in [12].
• Novel glyph designs for summarizing critical characteristics of Animation-based approaches were first used by Eades and Huang to
one’s ego-networks, and a new timeline visualization for tracking show changes between time steps of dynamic networks [27]. Staged
the variations of ego-alter connection strengths and alter-alter transitions have been widely used to reduce users’ visual effort for
relationship structures; identifying and understanding the temporal changes [8, 21, 31]. Some
• A usage scenario with real dataset and a controlled user study change highlighting techniques are also applied in those staged an-
that demonstrate the effectiveness and usefulness of egoSlider. imations [8]. Although animation is an effective way to decrease
the complexity of dynamic network evolution, it may lead to a high
2 R ELATED W ORK cognitive load [5]. It is also difficult for users to track the transitions
2.1 Egocentric Network Analysis when comparing multiple networks at the same time.
Ego-network has been extensively studied in the field of anthropology One type of timeline-based techniques leverages a series of node-
and sociology for a long time. When studying ego-networks, most link diagrams to represent networks at different time steps, such
of the works focus on the 1-level ego-network formed by the ego as small multiples of network snapshots juxtaposed to each other
and its 1-degree alters, i.e., the ego’s directly connected friends (e.g., [51]. Itoh et al. extended this idea into 3D spaces so that users can
[2, 4, 24, 45]); and few study ego-networks containing the ego’s observe global differences between graphs more easily [38]. Some
more distant connections (e.g., [33] discusses 2-level ego-networks other methods put nodes on a vertical axis, which can better utilize
including 1- and 2-degree alters). Major literature in those fields falls the screen real estate and highlight the changes of links over time
into two categories: microscopic level and macroscopic level analysis. [23, 32, 53]. We adopted a similar design to visualize the evolutions
Microscopic level analysis studies how structures and attributes of alter connections in ego-networks. Another approach is based
of ego-networks affect the ego’s behaviors. One major focus is to on a matrix-based representation of networks, where the temporal
investigate correlations between the topology of ego-networks and information is encoded as an intra-cell glyph-based timeline, such
the ego’s characteristics. For example, the structural hole theory as simple bar charts or Gestaltlines [20, 22]. However, since the
indicates that an individual may gain strategic advantages over others minimum matrix cell size is restricted by the intra-cell visualization,
when his or her alters are highly separated and have a relatively low such methods are not scaled for large networks. Besides, some works
visualize the adjacency list over time which is suitable for analyzing Data Storage Data Analysis Visual Analysis Analytic Tasks

certain tasks in dynamic graphs such as exploring the overall link Raw Data
Ego-network
Macro-level
Data Overview
change patterns [34, 47]. Filtering

There exist a few visualization systems using a hybrid approach, (a)

Feature Timeline
Meso-level
i.e., combining the animation-based and timeline-based techniques Extraction Summary View

together to achieve better performance for certain tasks. DiffAni Similarity Alter Timeline
Ego Micro-level
supports flexible interactions that allow users to divide the whole graph Network Measuring View
sequence into several aggregation views including diff tiles, animation (a) (b) (c) (d)
tiles, and small multiple tiles [49]. Beck et al. used a rapid serial
visual presentation approach to animate a timeline of graphs at a high Fig. 2. The overview of the egoSlider visualization pipeline. a) The
frequency in order to address the scalability problem [13]. ego-network structures are extracted from the raw data and stored into
However, all the above dynamic graph visualization techniques aim MongoDB. b) The data analysis module incorporates several analytical
to show variations of the entire graph. Egocentric analysis focuses on methods to process the dynamic ego-network sequences. c) Users
specific sub-networks, where particular connections, such as ego-alter can interactively navigate through three major views to perform visual
analysis of the data. d) Each view in egoSlider is aimed to address a
and alter-alter relationships, are more concerned rather than the overall
different level of ego-centric analytical tasks.
topology. Also, it is difficult to use the above techniques to achieve
visual comparisons of different ego-networks, since they all focus on The mesoscopic level questions focus on the overall comparison
one dynamic graph. among a set of individuals’ ego-networks [11, 25, 33]:
There have been several attempts to help users understand network Q3 What are the general similarities between multiple people’s ego-
data from the egocentric perspective. Many approaches adopt a radial networks along time? Do their ego-network sizes increase or
layout where alters are positioned around the ego and the temporal decrease simultaneously? Do they all tend to meet more new
relation information is encoded by the radius length [19, 28, 46]. On alters in a specific time period? Are there any different trends in
the other hand, Shi et al. proposed a 1.5D visual design to reveal the terms of alter numbers?
dynamic pattern of an ego-network [50]. However, many essential Q4 What are the differences between multiple people’s ego-networks
aspects of ego-networks are missing in those visualizations, such as at a specific time step? Do they hold the same alter density? Do
relation strength changes between an ego and its alters, and inter-alter they have a similar number of 1-degree alters? Do the alters share
connection variations over time. Compared with these approaches, common attributes? How about 2-degree alters?
egoSlider covers a much wider range of important properties for The microscopic level questions mainly study the detailed behav-
analyzing ego-network evolution. iors of a particular person’s ego-networks [2, 4, 24, 36]:
Q5 How does the number of an ego’s 1- or 2-degree alters change
3 A NALYTICAL Q UESTIONS over time? Is it increasing or decreasing? Is there a periodical
Experts with different backgrounds may have various interests on pattern? Are there any alter number spikes? Are the 1- and 2-
certain ego-network features. For example, anthropologists could pay degree alter volumes correlated?
more attention to kinship relations which is related to a set of alters Q6 How do the tie strengths between the ego and its alters evolve
who hold strong ties with the ego; health experts could focus on an along time? Do the alters who share stronger ties with the ego
ego’s pro-social behaviors that can be defined as a quantitative ego also preserve longer relations? Does the tie strength become
attribute. We aim to design egoSlider as a general-purpose visual weaker and weaker before an alter leaves the connection? Do
analysis tool that benefits users from different domains with deeper un- the majority of alters follow the same tie strength evolutionary
derstandings of ego-network evolution. Although Ahn et al. presented pattern?
a comprehensive task taxonomy for network evolution analysis [3], Q7 How are the alters of an ego connected over time? How many
it is not specialized for ego-network analysis which demands specific connected components can be divided at different time steps?
attributes such as tie strength to be explicitly described in the tasks. What is the alter flowing trend among the connected components
We choose the approach of first deriving analytical tasks from the in a specific time span? Do the alters tend to diverge into
literature and then validating them with experts. We selected 38 different small subgroups or to merge into a highly connected
dominant ego-network related publications across different domains community?
(Sec. 2.1), then classified them into two categories: macroscopic level Q8 How do new relations come into being? Is a new alter also a
and microscopic level. Next we derived a series of key analytical 2-degree neighbor of the ego previously? If so, who are the
questions that users often need to answer when studying dynamic bridges between the new alter and the ego? Is a new alter the
ego-networks. We then conducted in-depth interviews with two ex- ego’s previous alter long time ago? If so, where did this alter
perts in graph mining. They helped us refine the questions, and go? Staying relatively close to the ego as a 2-degree neighbor, or
suggested adding an additional mesoscopic level to capture the overall much farther away?
comparison of ego-network evolution of different individuals, because
it is a common scenario in their daily research. 4 S YSTEM OVERVIEW
In this paper, we focus on the visualization of at most 2-level Motivated by the above analytical questions, we designed egoSlider
ego-networks (i.e., containing the ego, and its 1- and 2-degree alters), allowing users to explore and analyze dynamic ego-network data at
which is related to a vast majority of studies in the literature. Unless three different scales: the overall patterns of all people’s ego-networks
specifically stated, when mentioning the ego’s alters, we refer to its with a Data Overview, the similarities and differences of ego-networks
1-degree alters; when mentioning one’s ego-network, we refer to the among individuals of interest with a Summary Timeline View, and the
1-level ego-network. The analytical questions are described below. detailed information of a person’s ego-network history from multiple
The macroscopic level questions aim to obtain a whole picture of perspectives with an Alter Timeline View.
all ego-networks from a large group of people [6, 14, 30, 48]: The whole egoSlider system consists of three major components:
Q1 What are the overall patterns of a large group of people’s ego- data storage and preprocessing module, data analysis module, and
networks at each time step? Are there any clusters of people visual analysis module, as shown in Fig. 2. The data storage and
with similar ego-networks? Are there any outliers? processing module extracts ego-network structures from raw datasets
Q2 What are the evolutionary trends of a large group of people’s such as citation networks, communication networks, and online social
ego-networks? Do they share a common evolutionary path? networks. We used MongoDB as our data storage software since it
Does a number of people’s ego-network evolution always keep can provide a highly flexible and customizable data schema. The data
structurally similar along time, or become different sometimes? analysis module performs data filtering of the extracted ego-networks,
and further characterizes them with essential numerical features for
measuring graph similarities. It thus allows typical computational and microscopic level. In addition, a rich set of user interactions is needed
visual methods, such as multidimensional scaling (MDS) [40], to be in egoSlider to support a holistic analysis of dynamic ego-networks
used to reveal the distributions of the entire dataset and detect common across all levels of tasks.
patterns of ego-network evolutionary trends. Both the above two In this section, we first provide an overview of the egoSlider inter-
modules were developed in Python by leveraging Flask, a Python web face, then introduce the visual encodings of individual views as well
framework, to build the backend. Three major views are integrated as the design alternatives we considered, and finally present the user
in the visual analysis module, i.e., the front-end visualization, to interactions equipped in egoSlider.
support different levels of analytical tasks with smooth interactions.
We implemented the visual analysis module using AngularJS and D3. 6.1 egoSlider Interface
As shown in Fig. 3, the interface of egoSlider consists of four major UI
5 DATA A NALYSIS components: a) a Data Overview panel showing the overall patterns of
In this section, we introduce our analytical approaches used in the data the entire dataset of many people’s ego-network evolutionary histories,
analysis module (Fig. 2-b). We first describe a formal data model for b) a detailed view canvas displaying the Summary Timeline Views
dynamic ego-network datasets, and then a set of numerical metrics for of selected individuals’ ego-networks and their fully expanded Alter
measuring the similarity between two ego-networks. Timeline Views on demand, c) a control panel displaying all the egos
from the dataset in a table with interactive searching, and d) a toolbar
5.1 Data Model on the top where users can select the dataset to visualize and toggle
In egoSlider, we consider dynamic egocentric network data as a se- the overview panel and the control panel.
quence of local friendship networks with time steps. Particularly, an
ego-network at time step t is modeled as an undirected weighted graph 6.2 Data Overview
Gtu = (Vut , Eut ) centered at the ego actor u. The graph nodes, Vut , consist We designed the Data Overview to illustrate the overall ego-network
of the ego u and ntu active alters of u, denoted as Vut = {u ∪ vtiu |1 ≤ i ≤ patterns (Q1 and Q2). By using the ego-network similarity measure-
ntu }. Each edge e j ∈ Eut ⊆ Vut ×Vut , 1 ≤ j ≤ |Eut | has a weight we j ∈ R. ment in Sec. 5.2, for each time step, we used MDS layout [40] to
The difference between Gtu and a general graph is that there are exactly generate the distribution of ego-networks. Different from traditional
MDS-based graph layout where closer nodes share more common
ntu edges Fut = {(u, vtiu )|1 ≤ i ≤ ntu } ⊆ Eut which connect the ego u and
neighbors, closer ego-networks here express more similarity based on
all of its alters. Thus, the dynamic egocentric network of the ego u
the metrics. As shown in Fig. 1-a, each ego-network is represented as
can be represented as Γu = {G1u , G2u , ..., GTu }, a sequence of T graphs
a dot and MDS plots of all time steps are sorted and stacked along a
where Gku , 1 ≤ k ≤ T shares the same ego u. vertical axis in the chronological order. All data points related to the
5.2 Graph Similarity currently focused egos are highlighted in red. And ego-networks of
the same individual are connected with lines across time steps.
Based on the studies about ego-network and general graph analysis
(e.g., [11, 14, 15, 24]), we derive the following essential metrics to 6.3 Summary Timeline View
characterize an ego-network. We describe those metrics using the data We designed glyphs on a timeline to visually summarize the evolution-
model introduced above: ary process of an individual’s ego-network. This design consists of two
• Number of alters of the ego u: ntu = |Vut − {u}|; parts: a snapshot glyph indicating the network structure and properties
• Density (or clustering coefficient) of the ego u’s ego-network: at one time step, and a transition glyph representing the changes
den(Gtu ) = nt (nt L−1)/2 , L = |Eut | − ntu , where L denotes the num- between two consecutive time steps. Based on the analytical questions
u u
ber of edges between the alters; (Q3 and Q4) in the mesoscopic level, we encoded several key variables
• Average tie strength (or weight) between the ego u and its alters: about 1- and 2-degree alters in those glyphs. Thus, users can easily
avg(wei ), ei ∈ Fut ; capture the main characteristics of a person’s ego-network evolution.
• Number of edges between the alters of the ego u: |Eut − Fut |; It also enables the comparison of multiple people’s ego-networks using
• Number of 2-degree alters of the ego u: |N(Gtu )| = |{w|w ∈ relatively small screen real estate.
Vvt , v ∈ Vut , w ∈
/ Vut }|; Snapshot glyph. As shown in Fig. 4-a, the inner circle color
• Average alter number of the alters’ themselves ego-networks: represents the ego density which measures to what extent the alters
avg(ntv ), v ∈ Vut − {u}; connect to each other, where the darker purple means the higher
• Number of outgoing edges from the ego u’s ego-network: density. The outer parts of the glyph indicate some information about
|E(Gtu )| = |{ev,w |ev,w ∈ Fvt , v, w ∈ Vut , w ∈
/ Vut }|. the ego’s 1-degree alters and 2-degree alters. First, a ring filled with
Leveraging the above metrics, we form a feature signature vector four different colors shows the distribution of four types of 1-degree
for each ego-network and adopted the Canberra Distance proposed in alters: 1) new alters who have no connection with the ego at the
[15] to compute pairwise similarity or distance between ego-networks: previous time step (green), 2) alters whose tie strength with the ego
dCan(P, Q) = ∑ni=1 |Pi − Qi |/(Pi + Qi ), where P and Q represent the increases compared to that at the previous time step (red), 3) alters
feature signatures of two ego-networks. whose tie strength with the ego decreases (blue), and 4) alters who
We choose this signature similarity based approach because it is hold the same tie strength with the ego (light gray). The width of this
less expensive to compute while providing a customizable graph com- circle is mapped to the number of 1-degree alters of the ego. Second,
parison criteria [15]. Also, the Canberra Distance measurement is a gray ring is drawn to indicate the number of 2-degree alters with
sensitive to small changes and normalizes the absolute difference of its width. However, it may be difficult to compare quantities of four
individual comparisons, which benefits users with the detection of types of alters in two glyphs. egoSlider also supports a bar chart based
clusters and outliers of ego-networks at different time steps. design as shown in Fig. 4-b. Similarly, the background square size is
constant and its color represents the ego density. Four bars illustrate
6 V ISUAL D ESIGN the numbers of different 1-degree alters. The top black bar indicates
The main goal of egoSlider visualization design is to provide intuitive the 1-degree alter number and the bottom one shows the 2-degree alter
visual metaphors supporting the analysis of dynamic ego-network number. Since it is easy to compare the total alter numbers in the pie
datasets according to the previously introduced questions at three chart design, and it is also visually attractive with a metaphor of an ego
different scopes (Sec. 3). We aim to offer effective visual summaries surrounded by alters, egoSlider uses this design as the default display.
of important ego-network characteristics to allow temporal pattern dis- But users can switch to the bar chart based design on demand.
covery at different scales, for example, clusters and outliers detection Snapshot glyph design alternatives. We considered several al-
at the macroscopic level, overall ego-network topology comparison at ternative solutions during our glyph design process. Line chart can
the mesoscopic level, and detailed alter information browsing at the also be used in the glyph but it has several drawbacks. First, the data
Fig. 3. The egoSlider interface consists of the following components: a) a Data Overview showing patterns of the entire dynamic ego-network
data, b) a main canvas displaying detailed ego-network evolutionary history of selected individuals, c) a control panel with a search bar and a data
table, and d) a toolbar for selecting datasets. Currently, the main canvas shows the Summary Timeline Views of the top 10 researchers in a DBLP
collaboration network dataset based on their publication numbers, where Kwan-Liu Ma’s timeline is expanded and shown in the Alter Timeline View.
Number of 1-degree alters Ego density Stronger High
presented in snapshot glyph has different natures and value ranges
(e.g., ego density and alter numbers), so using a unified y-axis is Number of constant alters

confusing. Second, it might be difficult to track changes since the

Tie strength change

vertical space is limited. Third, line chart introduces occlusion for

Ego density
the transition glyph introduced later. Pie chart and bar chart do not
have this visual clutter problem. Moreover, a number of alternatives Tie strength change
pie chart designs have been experimented (Fig. 5-a). The first design
Number of 2-degree alters a Alter type
choice uses the radius of a pie chart for the 1-degree alter number
and the opacity of colors for the ego density. However, users can
Number of 1-degree alters Ego density Weaker Low
hardly see the alter type distribution when the 1-degree alter number is Number of constant alters
small. In addition, it is difficult to get an accurate value estimation on New alter:
the opacity channel on many colors. The second design employs the Weaker tie:
inner and outer circle radii to encode the 1-degree and 2-degree alter Same tie:
numbers, and the diverging purple-yellow color of the inner circle to
Stronger tie:
indicate which alter number is greater. Users may feel confused about Tie strength change
this color-coding since the outer circle also employs a categorical color Number of 2-degree alters b Alter type Alter type
scheme. In the third design, the 1-degree and 2-degree alter numbers Fig. 4. The visual encodings of glyphs in the Summary Timeline View.
are encoded by the lengths of two vertical bars respectively. The colors The snapshot glyph represents an individual’s ego-network structure
of those bars represent the ego density. This design is limited when the and properties at each time step. Two design choices are shown: a) pie
vertical bars are short. Besides, the transition glyph may also affect chart based, and b) bar chart based. The transition glyph summarizes
users perceiving the length of the bars. For the bar chart based design, the temporal changes between those time steps.
we also considered to use the size or height of the background square
to indicate 1-degree alter number, but both designs make users difficult users are familiar with. Also, the tilt angle may also impact users
to perceive the ego density when the alter number is small. perceiving the gestaltline thickness. In addition, we attempted to
Transition glyph. As Fig. 4 illustrates, we use the line thickness encode the tie strength changes of all three types of 1-degree alters,
to represent the volume of consistent 1-degree alters between two i.e., alters whose tie strengths are stronger, weaker, and the same
consecutive time steps, i.e., those who remain directly connected to across two time steps, by drawing three parallel lines. But we found
the ego. A diverging blue-red color scheme is used for the transition that the transition glyph became too complicated and overwhelming,
glyph to summarize the tie strength changes of those alters, which is and those lower-level information can be accessed alternatively from
in consistency with the color-coding of the snapshot glyph, i.e., red the Alter Timeline View introduced below.
for a stronger change, blue for a weaker change, and light gray for no 6.4 Alter Timeline View
change. A dashed line is shown when there are no consistent 1-degree
alters between the ego-networks at the two consecutive time steps. The Alter Timeline View aims to support the microscopic level anal-
Transition glyph design alternatives. Fig. 5 shows three design ysis (Q5-Q8) by providing detailed information of alters rather than
alternatives of the transition glyph that we experimented with, where overall statistics. The Summary Timeline View can only partially
different types of overall constant alters’ tie strength changes are listed address Q5, whereas here users can track the temporal variations
from top to bottom. In the first design, the line thickness encodes the of tie strength between the ego and any individual alters as well as
1-degree alter number and the tilting angle encodes the tie strength inter-alter connection structures. We considered node-link based and
change. In this design, the tilt angle is somewhat restricted by the line matrix based design choices before we adopted the current design,
thickness, so users may feel more difficult to discern different levels but they have the scalability problem as indicated in the literature
of tie strength when the line becomes thicker. We also explored the (see Sec. 2 and [12]), whereas the line chart based design is more
possibility to embed the gestaltline [20] in the second design choice. compact. As the direct-connected alters are more important, we use
However, the gestaltline is not a common visualization encoding that different visual aggregation methods to encode 1-degree and 2-degree
 curve is shown starting at the alter bar and connecting to the boundary
of the flow chart at the next time step. In some situations, an alter
may become 2-degree from 1-degree, stay as 2-degree for a while, and
reappear as 1-degree, which is important to be captured in ego-network
analysis. Thus, we encoded this alter behavior with a lurking curve in
interior of the 2-degree alter flow to indicate such future reconnection.
a Alternative 1 Alternative 2 Alternative 3 The position of the lurking curve is determined by the timespan while
remaining as a 2-degree alter, where the longer the time interval, the
closer the lurking curve to the bottom.
Design alternatives. We discussed several candidate designs of
the Alter Timeline View before we made the final decision. One
design choice is to position the lurking curves below the 2-degree alter
flow separately (rather than on top of the flow and overlapped). This
approach reduces the visual clutter when multiple lurking curves share
b Alternative 1 Alternative 2 Alternative 3 a similar timespan, especially when the 2-degree alter number is small
Fig. 5. Design alternatives of a) the default snapshot glyph and b) the (making the flow tiny). However, it is not space efficient and much
transition glyph. For the transition glyphs, three different types of tie screen real estate is actually allocated to less important information.
strength changes are indicated: increasing, constant, and decreasing Another design choice is to encode 2-degree alters who changed to and
(from top to bottom). from the ego’s 1-degree neighbors with similar glyphs to Fig. 4 on top
High
a
Alter type: New alter Same tie of the alter flow chart. However, we found those glyphs significantly
Weaker tie Stronger tie
occlude the lurking curves which represent alters constantly remaining
as 2-degree neighbors. Therefore, we chose to draw curves connected
Tie strength

to the flow boundaries to convey the same message, which is also

d
consistent with the line-based visual encodings in this view.
6.5 User Interactions
c
Low
To allow users to smoothly perform ego-network analysis from the
b
three levels of scales and gain deeper insights, egoSlider incorporates
a set of intuitive interactions to help users browse data through the
multiple visualization views introduced above.
Navigating through multiple views. Users can glance at the
t t+1 t+2 t+3 t+4 t+5 t+6 t+7
entire dynamic ego-network dataset with the Data Overview to identify
Fig. 6. The visual encodings of the Alter Timeline View: a) the 1-degree clusters and outliers. Alternatively, users can browse those egos
alter collections at one time step; b) the 2-degree alter volume flow, c) through a table on the Control Panel showing the egos’ names and
a new 1-degree alter who was the ego’s 2-degree alter in the previous some attributes. Once individuals of interest are identified, users can
time step; and d) an alter transits to the ego’s 2-degree neighbor and select them to compare their ego-network evolutional history with
returns to the 1-degree neighbor after several time steps.
Summary Timeline Views on the main canvas. Further details-on-
alters at different levels of scales (Fig. 6). In addition, two Alter demand explorations can be achieved by clicking a button to expand
Timeline Views can be displayed in one window together which makes the Summary Timeline View with the Alter Timeline View.
comparison possible at this level. Filtering and searching. Users can interactively select an area
Timeline encodings. As shown in Fig. 6, each 1-degree alter is rep- of interest in the Data Overview, thus filtering out other egos and
resented by a small horizontal bar which uses the same color-coding revealing how the similarities among those ego-networks evolve along
schema as the snapshot glyph. Alter bars are organized vertically time. In the Alter Timeline View, further filtering of alters based on
into groups based on their tie strength with the ego at each time their degrees is supported. Moreover, a search bar is provided for
step. From top to bottom, the tie strength between the ego and its looking up a particular ego’s name on the Control Panel.
alters decreases. We also support an interaction to group alters into Synchronizing and desynchronizing timeline. By default, all the
connected components. In that case, the alters within a group belong to timeline visualizations are synchronized globally along one time axis.
the same connected component rather than share the same tie strength. Users can navigate through time using a global slider on the top. To
Note that the colors and positions of those alters are flexible in this compare two pieces of timelines from different time ranges, users can
visualization design. For example, we can use the color to encode detach a timeline and shift it individually using a local slider.
alters’ influence to the whole network, and group the alters into graph Modifying alters’ color-coding and/or positioning. Users can
connected components vertically to help verify the “structural hole” switch between different glyph designs (Fig. 4) in the Summary Time-
theory [24]. As for 2-degree alters, which are less important, we line View. Users can also dynamically order the 1-degree alters in the
aggregate them as a light blue flow chart shown at the bottom of this Alter Timeline View based on different criteria, such as by tie strength,
view to illustrate their temporal volume changes. connected graph components, and other alter attributes. Similarly, the
Same alters between consecutive time steps are linked by lines. color-coding of alter bars is flexible to be changed to represent alter
Thus, each curve depicts an alter’s tie strength evolutionary history types, alter’s overall influence, and so on.
along time. If an alter remains as the ego’s 1-degree neighbor, we can Highlighting and brushing. Most of the visual elements in
simply draw a curve connecting them. To avoid edge crossing, we egoSlider are associated with informative tooltips, allowing users to
adopted an alter sorting algorithm. At each time step, all the alters are have more information and learn about the visual encodings. Certain
first sorted by the types of their tie strengh changes. Then, the constant highlighting of visuals, such as alter evolutionary curves in the Alter
alters are positioned in the same order as those in the previous time Timeline View, can be fixed when selected. In addition, brushing &
step. Finally, the positions of alters who become 2-degree alters in the linking techniques are applied between different visualization objects.
next time step are lowered. In addition to the tie strength, the degree For example, hovering over an ego-network in the Data Overview
of an alter may change over time. For example, if an alter becomes the reveals all other ego-networks sharing the same ego; hovering over
ego’s 1-degree neighbor from its 2-degree neighbor in the network, a an ego snapshot glyph or transition glyph in the Summary Timeline
curve is drawn starting at the boundary of the 2-degree alter flow and View indicates the corresponding alters in the Alter Timeline View;
ending at the alter bar at the next step (Fig. 6-c); similarly, if an alter and hovering over an alter in the Alter Timeline View highlights its
changes from a 1-degree neighbor of the ego to a 2-degree neighbor, a directly connected alters in the same ego-network.
 a b c

Fig. 7. Partial visualization of the DBLP collaboration network dataset in the Data Overview: a) earlier years (1993–1999), and b) more recent years
(2008–2014). c) Summary Timeline Views of the top 10 researchers based on their publication numbers; compared with Fig. 3-b which shows the
recent years, it shows the early years of those authors.

7 E VALUATION
Here we describe a usage scenario to demonstrate the effectiveness
and usefulness of egoSlider using the DBLP collaboration network,
and present a controlled user study to quantitatively compare the user
performance of egoSlider with a baseline dynamic graph visualization.
7.1 Usage Scenario: DBLP Collaboration Network
Academic collaboration network, which is a specific kind of social a b
networks, is a common dataset interested in many applications, be-
cause it is typical and has the challenges of being huge and dense. The Fig. 8. Comparison of the Alter Timeline Views of a) Kwan-Liu Ma and
ego-networks extracted from a large collaboration network indicate re- b) Daniel Cohen-Or after filtering out their alters who collaborate with
searchers’ collaboration circles. The temporal evolution of a person’s them less than 5 times. Alters are ordered by tie strength and colored
collaboration network can also reflect his or her career development. by their publication performance (the darker the higher).
Do top researchers start to collaborate extensively in their early career transition glyphs. Fig. 7-c indicates the early years of those authors’
time? Do they tend to keep any long-term collaborators? Many careers, where Xiaoou Tang’s timeline is detached and shifted from
interesting questions can be asked. Our experts from the graph mining the global time axis since his publication history starts from 2002.
domain are particularly interested in this kind of data. Thus, we asked Most researchers’ collaborator numbers are small except Xiaoou Tang
them to try egoSlider and conducted interviews with them. We derived in their early academic years. We can also see that many dashed
the following usage scenario based on their observations. transition glyphs exist, which may indicate some stage transitions.
From the DBLP dataset [1], we select 52038 papers from 31 confer- For example, Harry Shum received his PhD in 1996, Kwan-Liu Ma
ences and journals in the fields of information visualization, computer in 1993, and Daniel Cohen-Or changed his job between 1994 and
graphics, computer vision, and human-computer interaction. We then 1996. As we further compare Ma and Cohen-Or’s ego-networks along
identify 64892 authors and extract their ego-network evolutionary time, their 1-level ego-network sizes are similar but the 2-degree alter
histories from 1975 to 2014. The tie strength between two authors are numbers are different, e.g., 2005–2010 in Fig. 3-b (Q4). Ma’s 2-degree
defined as the number of their collaborations in a year. We also define alter volume presents a periodic fluctuation pattern while Cohen-Or’s
one’s publication performance based on the publication number. is more stable, which can be better revealed after expanding their
We first toggle the Data Overview to gain a big picture about the timelines into the Alter Timeline Views in Fig. 8 (Q3).
entire dataset. From the MDS plots which summarize all authors’ By using the colors of the alter bars to represent the publication per-
ego-networks, we observe different clustering patterns along time. In formance, we notice that Ma’s 2-degree alter volume increases as the
recent years (2008–2014), there appears one giant cluster, a much number of his highly-performed 1-degree alters (Fig. 3-b where purple
smaller cluster, and several outliers (Fig. 7-b). As we move earlier, the means higher performance). This may suggest that his 2-degree alters
sizes of different clusters tend to be more equalized, showing three or are strongly connected to those 1-degree alters, and other 1-degree
four groups where one is slightly larger (Fig. 7-a). This might indicate alters have much fewer collaborators in general (Q5). Thus, we filter
that nowadays most of the authors collaborate in a similar pattern (Q1). out all the 1-degree alters who collaborate with him less than 5 times,
We then select the top 10 authors from the data table on the Control and find that their publication performance is not very high (Fig. 8-a).
Panel based on their publication number. In recent years (2004–2014), On the contrast, there still exist many highly-performed 1-degree alters
we observe that their ego-networks (highlighted in red) are similar and in Cohen-Or’s ego-networks after the same filtering (Fig. 8-b). This
distributed within a big cluster, but they are more spread out in earlier might be because that Cohen-Or works with many other professors
time (Q2). Another interesting fact we find is that Edwin Hancock’s whereas Ma usually collaborates with his own students (Q3 and Q4).
ego-networks are a slightly different (or far) from others in general. Further, in Fig. 3-b, by comparing the alter types over years, we find
To dive into more details, we leverage the Summary Timeline Views that more than 50% percent of Harry Shum’s collaborators are constant
of those top 10 authors, as shown in Fig. 3-b and Fig. 7-c. We imme- alters in a consecutive sequence of years. His more stable collabora-
diately identify that Hancock keeps a smaller group of collaborators tion relations may either be because he worked at an industry research
compared with others, based on the sizes of snapshot glyphs and lab where the employee turnover is less frequent than the student
 a

Xiaoou Tang

c a

Fig. 9. Alter Timeline View of Harry Shum’s ego-networks (2001– Xiaogang Wang

2010). Alters are ordered by tie strength and colored with publication
performance (the darker the higher). Three long-term coauthors are
highlighted: a) Baining Guo, b) Jian Sun, and c) Kun Zhou.

enrollment at universities, or because he worked at a leading position b

where he may have more stable colleagues (Q3). Moreover, Shum’s
egocentric network sizes expand significantly after 2004, which might Fig. 10. Alter Timeline Views of a) Jian Sun and b) Xiaoou Tang, where
be related to that the Managing Director position gave him a wider alters are ordered by tie strength and colored with alter types. The
highlighted alters are Xiaoou Tang and Xiaogang Wang respectively.
impact. After 4 years, his ego-network size has a big shrink, which
could be because he was promoted as a Corporate Vice President Comparison system. We chose the timeline-based approach with
and started to focus on Bing after 2007. He took the Executive Vice small multiples (Sec. 2.2) as our baseline for visualizing dynamic ego-
President position at Microsoft in 2013; and the visualization indicates networks, where the network at each time step was presented using
that he has no further publications after 2012 (Q5). the most widely-used node-link layout. The ego was shown in the
To further explore Harry Shum’s collaborators, we open his Alter center and the alters were placed around it. To adapt this baseline
Timeline View and group the alters into different connected compo- with ego-network analysis, we incorporated several interactions like
nents. As illustrated in Fig. 1-bc, his ego-network size increases and brushing & linking. The node-link graph was also tailored to meet the
the alters tend to connect as a whole (Q7). We then want to identify requirements of various tasks by emphasizing the focused features.
who are the most constant collaborators of Shum, so we sort his alters For example, we enlarged and highlighted the new alter nodes in all
based on the tie strength. Three long-term collaborators are revealed charts for new alter number comparison tasks, and encoded tie strength
in Fig. 9 (Baining Guo, Jian Sun, and Kun Zhou). After mapping the as the link width when appropriate. Visualizations of both egoSlider
color of the alter bars to publication performance, we observe that all and the baseline in the experiment were interactive as described earlier.
the three collaborators did not publish many papers at the beginning Tasks and procedure. In light of the aforementioned analytical
and they become productive after leaving the collaborative relation. questions (Sec. 3), we developed 2 concrete user tasks for each meso-
This suggests that they grew together in this process with Shum (Q6). scopic level or microscopic level question, 12 in total (Table 1). We
By hovering over those authors, which reveals their connected expected egoSlider to have better performance for all the tasks. We
alters, we find that after 2008, Guo and Zhou continued for their chose not to design tasks for macroscopic questions because they are
collaboration while Sun had no connections with the other two authors related to the Data Overview (not focused in the study) and the base-
(Q8). Does Sun have any constant collaboration with other established line system is not scalable for such tasks (which results in enormous
researchers? To answer this question, we thus display Sun’s Alter node-link diagrams). We also excluded the 2-degree related tasks
Timeline View and filter out the authors who publish papers with because the baseline visualizations could become overwhelming due
Sun less than 10 times. The only other alter with good publication to the network sizes. A within-subject experimental design was used.
performance is Xiaoou Tang as shown in Fig. 10-a with the bar chart Each participant performed 2 blocks × 12 tasks × 2 techniques = 48
based design (Q6). tasks in total. Within each block, the tasks were presented in the same
On the other hand, is Sun himself also Tang’s most constant col- order, and every task was shown in the two techniques (egoSlider and
laborator? We further display Tang’s Alter Timeline View and apply small multiples) one after another with different datasets. 4 predefined
the same filter. As shown in Fig. 10-b, apart from Jian Sun, there ego-network datasets were used and shuffled for each study. The order
are actually other researchers who keep a more constant collaborative of techniques was counter-balanced in the whole experiment. A brief
relation with Tang (Q8). For example, one of his students, Xiaogang tutorial was given to help participants get familiar with our visual
Wang, re-collaborated with Tang in 2010 after leaving in 2006. This design and the experimental system before the study. In the end of
might be explained by that Wang went back to CUHK after his PhD the tutorial, a set of tests are used to confirm the users can understand
study at MIT and worked as a colleague of Tang. Moreover, Tang’s and recognize our visual encodings including glyph size and color
most constant 1-degree alters appear in 2005–2008, indicating more accurately. During the study, task description was presented first, and
persistent collaborations (Q5). This might be because during that time, participants needed to click a start button to reveal the visualization
he worked as the director of the computer vision group at MSRA. and selected the answer from 5 choices. We recorded the response
time and accuracy in the study.
7.2 User Study Results. Table 2 summarizes the results from our experiment,
To quantitatively evaluate egoSlider’s effectiveness, we conducted a where the better performance from the two techniques is highlighted
controlled user study for comparing egoSlider with a baseline visual- (task time is highlighted if the difference is greater than 1s). On
ization. We focused on only evaluating the Summary Timeline View average, egoSlider outperforms the baseline in both accuracy
and Alter Timeline View in this comparative user study, because they (egoSlider: 92.5%, baseline: 83.6%) and time (egoSlider: 16.76s,
are the main visual designs of egoSlider. We recruited 15 students baseline: 19.55s). The results indicate that egoSlider is more accurate
(12 males) from a university with diverse majors, all with normal or in almost all tasks. The reason why egoSlider is less accurate in
corrected-to-normal vision and no color-blindness. Four of them had T9 could possibly be that users were confused about the alter bar
some knowledge of graph visualization. An online interactive system position representing connected components just after performing
was built on a webserver (3.40GHz Intel Core i7 CPU and 32GB mem- the tie strength related tasks (T7 and T8). For task time, we first
ory) for presenting tasks with different visualizations. Participants conducted overall analysis based on the level of tasks. Pairwise t-tests
completed tasks on a client desktop machine (Intel Core i5 CPU, 8GB indicate that there were significant differences in task time for both
memory, and 1920×1080 pixel resolution) with the Chrome browser.
 T1 How many egos whose 1-degree alter number keeps increasing? only analyze the similarities of ego-networks within one time step.
T2 Which ego has the largest number of constant 1-degree alters at a time step? We can explore the temporal similarities by comparing the dynamic
T3 Which ego has the largest number of 1-degree alters at a time step? ego-network sequences of different egos across time. For example,
T4 Which ego has the smallest percentage of new 1-degree alters at a time step? users may want to analyze if two individuals share a similar evolu-
T5 Which year does the ego has the largest number of 1-degree alters? tionary path. We can adopt time-series analyzing techniques such
T6 How many years does the ego’s 1-degree alter number increase? as dynamic time-warping [16] to identify such patterns. Second,
T7 Which year does a specific alter have the strongest tie strength with the ego? other potential analytical methods can be incorporated to enhance
T8 How many years does a specific alter’s tie strength with the ego increase? egoSlider’s abilities to explore larger dataset. For example, ranking
T9 Which year does the ego-network have the largest connected component number? of egos or other data mining algorithms can be used to help users
T10 How many years does the ego-network’s connected component number increase?
start with exploring the most important people or anormalies in the
T11 Which year does the ego has the smallest percentage of new neighbors?
T12 How many years are the percentage of new neighbors of the ego less than 50%?
data. Third, when users select too many individuals from the Data
Overview, the connecting lines between the same ego across different
Table 1. Experimental tasks: T1-4 are mesoscopic level tasks, and T5- time steps may become cluttered. This problem can be solved partially
12 are microscopic level tasks. by applying edge bundling techniques [35]. Fourth, eye tracking
Task Task Accuracy Task Time (s) Task Time T-test devices can be exploited to further understand detailed user behaviors
egoSlider Baseline egoSlider Baseline d f = 29 when exploring data with egoSlider.
A major consideration when designing egoSlider is scalability.
T1 80% 53.3% 20.07 (6.74) 25.43 (10.15) −4.175, p < .001
T2 100% 90.0% 16.47 (8.69) 26.46 (15.38) −4.321, p < .001
From the algorithm perspective, egoSlider’s data processing module
T3 100% 93.3% 17.58 (7.47) 15.92 (6.04) takes several minutes to extract the ego-network structures from the
T4 90% 63.3% 17.77 (5.61) 25.18 (10.85) −3.571, p < .05 raw DBLP records, compute the similarity metrics, and store them
into MongoDB, which is acceptable as a one-off job. To minimize the
T5 100% 90.0% 16.60 (4.75) 15.19 (7.50)
initial page loading time, data is progressively sent from the server and
T6 86.7% 73.3% 15.83 (4.33) 20.76 (9.59) −2.403, p < .05
T7 100% 100% 12.55 (4.08) 14.45 (4.12) −2.328, p < .05
rendered in the visualization client. From visualization perspective, in
T8 86.7% 70.0% 16.75 (6.73) 16.48 (4.89) order to minimize user’s memory burden, we only used 5 color hues
T9 86.7% 96.7% 17.64 (5.48) 17.15 (5.59) in the visual encodings. Also, the snapshot glyph and transition glyph
T10 96.7% 93.3% 17.86 (6.15) 19.84 (7.20) share the same color for alters of stronger/weaker ties (red and blue).
T11 91.7% 86.7% 16.76 (7.28) 17.12 (7.49) Although users can scroll the visualization, approximately a maximum
T12 96.7% 93.3% 15.29 (4.29) 19.63 (6.37) −3.139, p < .05 of 15 time-steps can be shown in one screen. Thus, visual aggregation
Table 2. Experimental results and task time t-test results. Task time is of many time steps are needed to display really long timelines. In the
shown in avg. (std.). T-values of significant differences are reported. current design, the number of alter bars we can support approximates
the Dunbar’s number (150) [26], i.e., the suggested cognitive limit
mesoscopic level tasks (t119 = −5.5501, p < .05) and microscopic number of people with whom one can maintain stable social relation-
level tasks (t239 = −2.607, p < .05). We then perform pairwise t-tests ships. Most of the top 10 researchers described in Sec. 7.1 have more
for each task, and significant differences were found in T1, T2, T4, than 100 alters in total, which is suitable in our current visualization.
T6, T7, and T12 (6 out of 12). The baseline was faster than egoSlider Though the alter numbers in online social networks might be much
in T3 and T5, however, the differences were not significant. These larger than the ones in traditional social networks [10], the scalability
tasks were both related to identifying the largest 1-degree alters, so it of the Alter Timeline View can be improved by aggregating the alter
may be because the snapshot glyphs of egoSlider are too small. For bars as rectangles and alter curves as flows. Finally, the current Alter
T10, egoSlider was faster but not significant, however, the variance Timeline View design may be cluttered in some extreme cases such
of the baseline was relatively high, indicating that the node-link as there are a lot of transitions between 1-degree alters and 2-degree
visualization is not stable. Users need to manually change the alter alters while it can be solved by interactively filtering these edges.
bar position encoding in T9 and T10. This operation takes time and
can cause confusion, which might be the reason that egoSlider had 9 C ONCLUSION AND F UTURE W ORK
worse performance. As for participants’ feedback, all of them thought We have presented egoSlider, a visualization system for analyzing
egoSlider was more visually attractive, useful, and efficient than the dynamic ego-network data, which incorporates a number of new vi-
baseline. Some mentioned that it took time to remember all the visual sual encodings with three interactive visualization views to address
encodings while the tooltips helped a bit. One participant felt that it ego-network analytical questions across different levels. A rich set
was not easy to scroll the timeline and suggested to widen the slider. of interactions is supported, allowing for flexible visual exploration
through the three views. We have also described a comprehensive
8 D ISCUSSION
usage scenario with real dataset and a controlled study. The results
The usage scenario and the quantitative user study demonstrate the indicate that egoSlider is effective in dynamic ego-network analysis
effectiveness and usefulness of egoSlider. However, there still exist and outperforms the baseline visualization in many analytical tasks.
some limitations. First, there is a learning curve for using egoSlider. In the future, we plan to embed some analytical methods to inspect
Although we intuitively encode important ego-network features in the the similarities between people’s ego-network sequences along time
Summary Timeline View and Alter Timeline View, users need to get and to detect trends and anomalies in ego-network evolution. We
familiar with the visual encodings before starting the exploration. Sec- also want to incorporate multivariate visualization techniques to show
ond, egoSlider encodes the alter bar position based on the tie strength more ego and alter attributes, as well as several extensions of the
of alters by default and the alter bars can be reordered in groups based current system discussed in Sec. 8. Moreover, we aim to conduct more
on their connected components by clicking a button. Although this realistic case studies and user studies with a variety of ego-network
interaction enriches the system’s functionality, users may be confused datasets to further examine the effectiveness of egoSlider.
when switching between the two options, as indicated from the results
of the user study. More distinct visual cues about those two view 10 ACKNOWLEDGMENTS
statuses are needed for improving the usability of egoSlider. Third, We thank all the anonymous reviewers for their valuable comments.
due to the limitation of MDS layout, two similar egos might be This research was supported in part by HK RGC GRF 618313, the
located at very different positions at two time steps, because the layout National Basic Research Program of China (973 Program) under Grant
optimizes the similarities locally in each MDS plot. This introduces No. 2014CB340304, and a grant from Huawei Co. Ltd.
inconsistency in the Data Overview.
In addition, there are also many interesting perspectives that can R EFERENCES
be extended from the current egoSlider system. First, currently we [1] The dblp dataset. http://dblp.dagstuhl.de/xml/.
 [2] A. Abbasi, K. S. K. Chung, and L. Hossain. Egocentric analysis of [28] M. Farrugia, N. Hurley, and A. Quigley. Exploring temporal ego networks
co-authorship network structure, position and performance. Information using small multiples and tree-ring layouts. In Proc. of the SIGCHI Conf.
Processing and Management, 48(4):671–679, 2012. on Human Factors in Computing Systems, pages 79–88, 2011.
[3] J.-w. Ahn, C. Plaisant, and B. Shneiderman. A task taxonomy for network [29] P. Federico, J. Pfeffer, W. Aigner, S. Miksch, and L. Zenk. Visual analysis
evolution analysis. IEEE Trans. on Visualization and Computer Graphics, of dynamic networks using change centrality. In Proc. of the International
20(3):365–376, 2014. Conf. Advances in Social Networks Analysis and Mining (ASONAM),
[4] G. Ahuja. Collaboration networks, structural holes, and innovation: A pages 179–183, 2012.
longitudinal study. Administrative Science Quarterly, 45(3):425–455, [30] D. Fisher. Using egocentric networks to understand communication.
2000. IEEE Internet Computing, 9(5):20–28, 2005.
[5] D. Archambault, H. C. Purchase, and B. Pinaud. Animation, small [31] C. Friedrich and P. Eades. Graph drawing in motion. Journal of Graph
multiples, and the effect of mental map preservation in dynamic graphs. Algorithms and Applications, 6(3):353–370, 2002.
IEEE Trans. on Visualization and Computer Graphics, 17(4):539–552, [32] M. Greilich, M. Burch, and S. Diehl. Visualizing the evolution of
2011. compound digraphs with timearctrees. In Computer Graphics Forum,
[6] V. Arnaboldi, M. Conti, A. Passarella, and R. Dunbar. Dynamics of volume 28, pages 975–982, 2009.
personal social relationships in online social networks: A study on twitter. [33] M. Grossetti. Where do social relations come from? Social Networks,
In Proc. of the ACM Conf. on Online Social Networks, pages 15–26, 2013. 27(4):289–300, 2005.
[7] V. Arnaboldi, A. Passarella, M. Tesconi, and D. Gazzè. Towards a [34] M. Hlawatsch, M. Burch, and D. Weiskopf. Visual adjacency lists for
characterization of egocentric networks in online social networks. In On dynamic graphs. IEEE Trans. on Visualization and Computer Graphics,
the Move to Meaningful Internet Systems: OTM 2011 Workshops, pages 20(11):1590–1603, 2014.
524–533, 2011. [35] D. Holten. Hierarchical edge bundles: Visualization of adjacency
[8] B. Bach, E. Pietriga, and J.-D. Fekete. Graphdiaries: Animated relations in hierarchical data. IEEE Trans. on Visualization and Computer
transitions and temporal navigation for dynamic networks. IEEE Trans. Graphics, 12(5):741–748, Sept 2006.
on Visualization and Computer Graphics, 20(5):740–754, 2014. [36] D. J. Hruschka. Friendship: Development, ecology, and evolution of a
[9] B. Bach, E. Pietriga, and J.-D. Fekete. Visualizing dynamic networks relationship, volume 5. University of California Press, 2010.
with matrix cubes. In Proc. of the 32nd Annual ACM Conf. on Human [37] D. Hunter, P. Smyth, D. Q. Vu, and A. U. Asuncion. Dynamic egocentric
Factors in Computing systems, pages 877–886, 2014. models for citation networks. In Proc. of the 28th International Conf. on
[10] L. Backstrom, P. Boldi, M. Rosa, J. Ugander, and S. Vigna. Four degrees Machine Learning, pages 857–864, 2011.
of separation. In Proc. of the 4th Annual ACM Web Science Conference, [38] M. Itoh, M. Toyoda, and M. Kitsuregawa. An interactive visualization
pages 33–42, 2012. framework for time-series of web graphs in a 3d environment. In
[11] L. Backstrom and J. Kleinberg. Romantic partnerships and the dispersion International Conf. on Information Visualisation, pages 54–60, 2010.
of social ties: A network analysis of relationship status on facebook. In [39] S. L. Jarvenpaa and A. Majchrzak. Knowledge collaboration among
Proc. of the 17th ACM Conf. on Computer Supported Cooperative Work professionals protecting national security: Role of transactive memories
& Social Computing, pages 831–841, 2014. in ego-centered knowledge networks. Organization Science, 19(2):260–
[12] F. Beck, M. Burch, S. Diehl, and D. Weiskopf. The state of the art in 276, 2008.
visualizing dynamic graphs. In EuroVis - STARs, 2014. [40] J. B. Kruskal. Multidimensional scaling by optimizing goodness of fit to
[13] F. Beck, M. Burch, C. Vehlow, S. Diehl, and D. Weiskopf. Rapid serial a nonmetric hypothesis. Psychometrika, 29(1):1–27, 1964.
visual presentation in dynamic graph visualization. In IEEE Symp. on [41] M. J. Lubbers, J. L. Molina, J. Lerner, U. Brandes, J. Ávila, and
Visual Languages and Human-Centric Computing, pages 185–192, 2012. C. McCarty. Longitudinal analysis of personal networks. the case of
[14] E. Bellotti. What are friends for? elective communities of single people. argentinean migrants in spain. Social Networks, 32(1):91–104, 2010.
Social Networks, 30(4):318–329, 2008. [42] U. Matzat and C. Snijders. The online measurement of ego centered
[15] M. Berlingerio, D. Koutra, T. Eliassi-Rad, and C. Faloutsos. Netsimile: online social networks. Social Networks, 32:105–111, 2010.
A scalable approach to size-independent network similarity. ACM [43] J. Mcauley and J. Leskovec. Discovering social circles in ego networks.
Computing Research Repository, abs/1209.2684, 2012. ACM Trans. on Knowledge Discovery from Data, 8(1):4, 2014.
[16] D. J. Berndt and J. Clifford. Using dynamic time warping to find patterns [44] A. J. Omalley, S. Arbesman, D. M. Steiger, J. H. Fowler, and N. A.
in time series. In KDD workshop, volume 10, pages 359–370, 1994. Christakis. Egocentric social network structure, health, and pro-social
[17] C. Bidart and D. Lavenu. Evolutions of personal networks and life events. behaviors in a national panel study of americans. PLoS One, 7(5):e36250,
Social Networks, 27(4):359–376, 2005. 2012.
[18] S. P. Borgatti and D. S. Halgin. An introduction to personal network [45] C. Prell. Social network analysis: History, theory and methodology.
analysis and tie churn statistics using e-net. Connections, 32(1):37–48, Sage, 2011.
2012. [46] F. Reitz. A framework for an ego-centered and time-aware visualization
[19] U. Brandes, M. Hoefer, and C. Pich. Affiliation dynamics with an of relations in arbitrary data repositories. Computing Research Reposi-
application to movie-actor biographies. In EuroVis, pages 179–186, 2006. tory, abs/1009.5183, 2010.
[20] U. Brandes and B. Nick. Asymmetric relations in longitudinal social [47] N. H. Riche, Y. Riche, N. Roussel, S. Carpendale, T. Madhyastha, and
networks. IEEE Trans. on Visualization and Computer Graphics, T. J. Grabowski. Linkwave: A visual adjacency list for dynamic weighted
17(12):2283–2290, 2011. networks. In Proc. of the 26th Conf. on l’Interaction Homme-Machine,
[21] U. Brandes and D. Wagner. Analysis and visualization of social networks. pages 113–122. ACM, 2014.
In Graph Drawing Software, pages 321–340. Springer, 2004. [48] S. G. Roberts, R. I. Dunbar, T. V. Pollet, and T. Kuppens. Exploring
[22] M. Burch, B. Schmidt, and D. Weiskopf. A matrix-based visualization variation in active network size: Constraints and ego characteristics.
for exploring dynamic compound digraphs. In Proc. of the International Social Networks, 31(2):138–146, 2009.
Conf. on Information Visualisation, pages 66–73, 2013. [49] S. Rufiange and M. J. McGuffin. Diffani: Visualizing dynamic graphs
[23] M. Burch, C. Vehlow, F. Beck, S. Diehl, and D. Weiskopf. Parallel with a hybrid of difference maps and animation. IEEE Trans. on
edge splatting for scalable dynamic graph visualization. IEEE Trans. on Visualization and Computer Graphics, 19(12):2556–2565, 2013.
Visualization and Computer Graphics, 17(12):2344–2353, 2011. [50] L. Shi, C. Wang, and Z. Wen. Dynamic network visualization in 1.5 d. In
[24] R. S. Burt. Structural holes versus network closure as social capital. Proc. of the IEEE Pacific Visualization Symp., pages 179–186, 2011.
Social Capital: Theory and Research, pages 31–56, 2001. [51] E. R. Tufte. Envisioning information. Optometry & Vision Science,
[25] A. Degenne and M.-O. Lebeaux. The dynamics of personal networks at 68(4):322–324, 1991.
the time of entry into adult life. Social Networks, 27(4):337–358, 2005. [52] S. Uddin and L. Hossain. Time scale degree centrality: A time-variant
[26] R. I. Dunbar. Neocortex size as a constraint on group size in primates. approach to degree centrality measures. In Proc. of the International
Journal of Human Evolution, 22(6):469–493, 1992. Conf. Advances in Social Networks Analysis and Mining (ASONAM),
[27] P. Eades and M. L. Huang. Navigating clustered graphs using pages 520–524, 2011.
force-directed methods. Journal of Graph Algorithms and Applications, [53] S. van den Elzen, D. Holten, J. Blaas, and J. J. van Wijk. Dynamic
4(3):157–181, 2000. network visualization withextended massive sequence views. IEEE
Trans. on Visualization and Computer Graphics, 20(8):1087–1099, 2014.

View publication stats