Citation Network Analysis of a Computer Science Conference

Clement Lee (Lancaster University)

2020-02-28 (Fri)

Background

Previously statistics postdoc at Newcastle University
Open Lab: openlab.ncl.ac.uk
World leading research group in
- Human Computer Interaction (HCI)
- Digital Civics
- Ubiquitous Computing

Themes

Digital Civics

Community health: Feed Finder, App Movement
Education: Learning Circle
Social care, public health, local democracy…

Human-Computer Interaction (HCI)

Designing digital technology with people in mind
ACM CHI Conference on Human Factors in Computing Systems

Digital library

dl.acm.org
All publications of ACM (journals, conference proceedings …)
What if we scrape and analyse the data?

(Source: xkcd.com/1683/)

Old wine in a new bottle

Bibliometrics is old
Works within CHI

Authors	Year	Digital Library ID
Bartneck and Hu	2009	1518810
Kaye	2009	1520364
Matejka et al.	2012	2212796
Liu et al.	2014	2556969
Padilla et al.	2014	2578867
Pohl and Mottelson	2019	3310429
Lee et al.	2019	3312776

Current approaches

Text approach
- Changing meanings of words
- Some words dominating the field, e.g. “design”, “user”, etc.

Network approach
- Focus usually on collaboration networks
Availability of data
- Other databases not very helpful
- Or require quite some manual scraping

Scope

Gather the references of all CHI papers from 1981 to 2019
- Digital Library ID useful for identifying papers
Construct citation network within ~7000 papers
- Well defined boundary
Cluster the papers
- Using methods for social networks
Examine the groups
- Do they represent different topics/themes?
- How is the trend over the years?

Scope

Gather the references of all CHI papers from 1981 to 2019
- Digital Library ID useful for identifying papers
Construct citation network within ~7000 papers
- Well defined boundary
Cluster the papers
- Using methods for social networks
Examine the groups
- Do they represent different topics/themes?
- How is the trend over the years?

Community detection

Algorithms for clustering relational data
Principles based on heuristics
Pros
1. Quick
Cons
1. Huge inequality in group sizes
2. Variation of results between algorithms
3. Number of groups?
4. Resolution limit

Stochastic block models

Statistical models for clustering relational data
Properly defined by model assumptions
Pros
1. One definitive set of results (theoretically)
2. Can resolve issues with community detection
Cons
1. Not off the shelf
2. Slow / difficult to scale

Difference

Community detection
- Nodes (papers) are tightly connected within each group
- Low connectivity between groups
Stochastic block models
- Same connectivity for any pair of nodes in two groups
- Not guaranteed to be high or low
- Existence of miscellaneous group

Statistical models

Slightly different way of looking at the data
- Computational/physical: input the parameters, output simulated data
- Statistical: input real data, output (estimates of) the parameters
Linear regression
- Parameters: slope and intercept
Epidemic modelling
- Parameters: transmission rate, basic reproduction number, …
- Say something about when individuals get infected

Inferring hidden structure

… is not trivial
Always comes with some degree of uncertainty
Label switching
- (A, B, C) in group 1; (D, E) in group 2
- (D, E) in group 1; (A, B, C) in group 2

Utilising extra information

Publication years available
Forcing the “mean” year to be increasing for groups
- Mean of group 1 \(<\) mean of group 2 \(<\cdots\)
- Smaller groups (labels, not sizes) represent earlier topics
- The group labels are actually interpretable

Results

Results (more interesting ones)

A simple web app

A new direction

Directed Acyclic Graph (DAG)
- No cycles of any length
- Start with any node A and travel along the edges
- Impossible to go back to node A

(Source: en.wikipedia.org/wiki/Directed_acyclic_graph)

Common as models

Bayesian networks
Probabilistic graphical models
Examples
- Kocabas, V. and Dragicevic, S. (2007), Enhancing a GIS Cellular Automata Model of Land Use Change: Bayesian Networks, Influence Diagrams and Causality. Transactions in GIS, 11: 681-702.
- Roos, J., Gavin, G., and Bonnevay, S. (2017), A dynamic Bayesian network approach to forecast short-term urban rail passenger flows with incomplete data. Transportation Research Procedia, 26: 53-61.

(Source: Kocabas and Dragicevic (2007))

(Source: Kocabas and Dragicevic (2007))

Not so common as data being modelled

Examples
- Cells in a spreadsheet

(Source: lynda.com/Excel-Mac-tutorials/Track-cell-dependencies-locate-formula-sources/431189/498037-4.html)

Not so common as data being modelled

Examples
- Cells in a spreadsheet
- River networks

(Source: topotoolbox.wordpress.com/2017/03/13/)

Not so common as data being modelled

Examples
- Cells in a spreadsheet
- River networks
- File/project/software dependencies

(Source: javadepend.com/Blog/?p=2328)

What next?

Incorporating the fact that “the data is a DAG” into the model
- Not as trivial as incorporating the publication years
- How to order the groups? What meanings do they have?
Incorporating the text and/or collaboration network into the model
- Same ultimate goal of clustering
- Potentially reduces statistical uncertainty

Thank you!

Questions?