Understanding and Organizing User Generated Data Methods and Applications.

Understanding and Organizing User Generated Data

Methods and Applications

2

August 16, 1977 June 25, 2009

3

August 16, 1977

4

5

June 25, 2009

6

7

officially pronounced dead

8

Media

Social

9

Part 2: Similarity

Part 1: Direct LinksThis talk:

Results that are directly applicable in end-user services

10

Part 1: Direct Links

11

Probability that two of my friends are (becoming) friends themselves is high!

high clustering

12

VENETA: Friend Finding

13

I want to meet people!

privacy preserving!

same contact = friend of a friend

14

Cluestr: Contact Recommendation

15

Clustering Survey:Communities are often addressed as groups!

„There‘s no training tonight!“

„Let‘s have a BBQ tomorrow!“

„Our next meeting is at 2pm!“

16

ClusteringRecommend contacts from clusters of already

selected contacts

Communities can be identified using

clustering algorithm

17

recommended contacts

Group

(i.e. „invited“ contacts)

updated group

new recommendations

Considerable time savings possible!

18

Part 2: Similarity

19

20

Academic Conferences

21

conference

publication

author

Similarity between Scientific Conferences

22

• Confsearch (Screenshot)• Highlight Ratings• Highlight Related Conference Search

23

Music Similarity

24

How similar is Michael Jackson to Elvis Presley?

25

Audio Analysis Usage Data

26

ba

cBAsim

),(

#common users (co-occurrences)

Occurrences of song A Occurrences of song B

„Users who listen to Elvis also listen to ...“

Problem: Only pairwise similarity, but no global view!

27

Getting a global view...

d = ?

pairwise similarities1

graph for all-pairsdistances (shortest path)

2

MDS to embed graph into Euclidean spacewhile approximately preserving distances (=> global view)

3

28

• Principal Component Analysis (PCA): – Project on hyperplane that maximizes variance.– Computed by solving an eigenvalue problem.

• Basic idea of MDS:– Assume that the exact positions y1,...,yN in a high-dimensional space

are given.– It can be shown that knowing only the distances d(yi, yj) between

points we can calculate the same result as applying PCA to y1,...,yN.

• Problem: Complexity O(n2 log n) – use approximation: LMDS [da Silva and Tenenbaum, 2002]

Classical Multidimensional Scaling (MDS)

29

Example: Kleinberg graph (20x20 grid with random edges)

Original embedding(spring embedder) After 6 rounds After 12 rounds After 30 rounds

Problem:

Some links erroneously shortcut certain paths

Use embedding as estimator for distance:

Remove edges that get

stretched most and re-embed

30

31

skip =

listen =

Play „random“ songs that match my mood!

32

After only few skips, we know pretty well which songs match the user‘s mood

Realization using our map?

33

„In my shelf AC/DC isnext to the ZZ Top...“

Browsing Covers

34

„from users for users“

Conclusion

35

Thank you

37

List of PublicationsSocial Audio Features for Advanced Music Retrieval InterfacesM. Kuhn, R. Wattenhofer, S. WeltenMultimedia 2010

Visually and Acoustically Exploring the High-Dimensional Space of MusicL. Bossard, M. Kuhn, R. WattenhoferSocialCom 2009

Cluestr: Mobile Social Networking for Enhanced Group CommunicationR. Grob, M. Kuhn, R. Wattenhofer, M. WirzGROUP 2009

From Web to Map: Exploring the World of MusicO. Goussevskaia, M. Kuhn, M. Lorenzi, R. WattenhoferWI 2008

VENETA: Serverless Friend-of-Friend Detection in Mobile Social NetworkingM. von Arb, M. Bader, M. Kuhn, R. WattenhoferWiMob 2008

Exploring Music Collections on Mobile DevicesO. Goussevskaia, M. Kuhn, R. WattenhoferMobileHCI 2008

The Layered World of Scientific ConferencesM. Kuhn and R. WattenhoferAPWeb 2008

The Theoretic Center of Computer ScienceM. Kuhn and R. Wattenhofer. (Invited paper)SIGACT News, December 2007

Layers and Hierarchies in Real Virtual NetworksO. Goussevskaia, M. Kuhn, R. WattenhoferWI 2007