Maps Daily Visualizing Media Trends

1
Maps Daily - Visualizing Media Trends

• Daisuke Mashima
• (Georgia Tech, GA)
• Dr. Stephen Kobourov
• Dr. Yifan Hu
• Dr. Emden R. Gansner
• (Information Visualization Group at ATT Labs, NJ)

2
Goal of Our Project

• Visualize dynamic data by using a map metaphor

3
last.fm

• http//www.last.fm
• UK-based Internet radio and music community site
• Over 30 million active users from all over the
  world

4
last.fm
5
last.fm API

• REST API to access last.fm database
• Bindings for major programming languages

6
Why Maps?

• A picture is worth a thousand words
• Graphs are often good for visualizing data.

7
How to Draw a Map
Collect relational Data (artists, similarities)
1 label"Queen", listeners"1637182",
playcount"52401547", tag"classic rock",
url"http//www.last.fm/music/Queen" 2
label"The Rolling Stones", listeners"1516977",
playcount"40697311", tag"classic rock",
url"http//www.last.fm/music/TheRollingStones"
9999 -- 9997similarity99.65 9999 --
9994similarity48.41
(DOT Format)
8
How to Draw a Map
Embedding into 2D (MDS, Force-directed layout)
9
How to Draw a Map
Clustering (Modularity-based etc.)
10
How to Draw a Map
Draw a map (Country-like boundaries, colors)
11
Challenges

• Dynamic Layout vs Mental Map
• Tradeoff between readability and data distortion
• Trend visualization
• How changes can be emphasized

12
Approach

• Step1 Create a big canonical map
• Step2 Extract HOT artists
• Step3 Visualize Trends

13
Step 1 Create Canonical Map

• Need as many artists as possible
• Just map 20,000 artists

14
Step 1 Create Canonical Map

• Layout and clustering done separately
• Combine them
• Cluster nodes with raw edge weights
• Adjust edge weights based on clustering
• What is distorted?
• Distance (similarity) among artists

15
Step 1 Create Canonical Map
16
Step2 Extract HOT Artists

• Metric
• of listeners? Playcount?
• Cumulative? Difference?
• How many?
• Need to fit a computer screen
• Top 250 in terms of difference in Playcount

17
Step 3 Visualize Trends

• Pick Top 250 artists and map them
• Set font sizes according to diff sizes
• FontSize BaseSize Range F(diff)
• where
• F(diff) (diff MEAN(diff))/((MAX(diff)
  MEAN(diff))
• Label overlap removal

18
Step 3 Visualize Trends
Looks OK.
19
Step 3 Visualize Trends Animated Map
Not easy to follow. Is it a Map?
20
Step 3 Visualize Trends More Stable
Animation

• Define a time window to be animated
• Extract all artists that appeared in top 250
  during the time window
• Map them

21
Step 3 Visualize Trends More Stable
Animation
Only changing font size is not enough.
22
Step 3 Visualize Trends

• Heat map?
• Use the same module to draw a heat map
• Re-cluster artists based on diff sizes (log scale
  etc.)

23
Step 3 Visualize Trends Heat Map
Country boundaries are lost.
24
Our Final Product
25
Behind The Scene
26
Software

• GMap by ATT InfoVis Group (Map drawling)
• Shell Script Cron (Automation)
• Java Application (Crawling etc.)
• Graphviz (Layout, overlap removal)
• LinLog Layout (Clustering)
• ImageMagick (Conversion from PS to GIF)
• Gifsicle (Animation GIF)

27
Future Work

• Consider other metrics
• Interactive user interface
• Improve visualization
• Put it online
• Black box it
• Evaluation

28
Thank you very much.???????????????