Collective Collaborative Tagging System

1
Collective Collaborative Tagging System

• Jong Y. Choi, Joshua Rosen, Siddharth Maini,
  Marlon E. Pierce, and Geoffrey C. Fox
• Community Grids Laboratory
• Indiana University

2
People-Powered Knowledge

• Delicious example

Bookmark
Tags
Social Networks
People-generated
3
People-Powered Knowledge

• Collaborative Tagging
• Online bookmarking with annotations
• Create social networks
• Utilize power of peoples knowledge
• Pros and cons
• High-quality classifier by using human
  intelligence
• But lack of control or authority

4
Motivations

• Distributed and fragmented knowledge
• ? Need an unified data set
• ? More accurate and richer information
• No flexibility in choosing different information
  retrieval (IR) algorithms
• ? Need a playground to do experiment with
  various IR techniques
• ? Help to discover hidden knowledge

5
Proposed System
Collective Collaborative Tagging (CCT) System
CCT System
Data Importer
RDF RSS Atom HTML
Data Coordinator
Distributed Tagging Data
Populate Bookmarks/ tags
Repository
Query with various options
User Service
Search Result SOAP, REST,
6
Development Plan and Progress

• 1st - Service and algorithm development
• Identify services and algorithms
• 2nd - Interface development
• Web2.o style interface
• REST, SOAP,
• 3rd Export/import service development
• Merging distributed data sets
• Export data to build mesh-up sites
• So far, we are mainly in 1st stage and do some
  experiments in 2nd stage

7
Prototype
Different Data Sources
Various IR algorithms
Flexible Options
Result Comparison
8
Service Types and Algorithms
Service
Description
Algorithm
Type
Searching
Given input tags, returning the most relevant X
(X URLs, tags, or users)
Latent Semantic Indexing (LSI), FolkRank
I
Recommendation
Indirect input tags, returning undiscovered X
II
Clustering
Community discovering. Finding a group or a
community with similar interests
K-Means, Deterministic Annealing Clustering
III
Trend detection
Analysis the tagging activities in time-series
manner and detect abnormality
Time Series Analysis
IV
9
Data Models (I)

• Vector-space model (bag-of-words model)
• Assume n URLs and q tags
• A URL can be represented by q-dimension vector,
  di (t1, t2, , tq)
• A total data set can be represented by n-by-q
  matrix
• Pairwise Dissimilarity Matrix
• n-by-n symmetric matrix
• Distance (Euclidean, Manhattan, )
• Angles, cosine, sine,
• O(n2) complexity

10
Data Models (II)

• Graph model
• Building a graph with nodes and edges
• Edges are indicating relationship
• Becoming complex networks (tag graph)
• Dissimilarity
• Related with path distance
• Finding path is important (Shortest path
  problem)
• Naive approach O(n3) complexity

(Source MSI-CIEC)
11
Searching

• Latent Semantic Indexing
• Using vector-space model, find the most similar
  URLs with users query tags
• Dimension reduction from high q to low d (q gtgt d)
• Removing noisy terms, extracting latent concepts

Ideal Line
Recall
2 terms4 terms8 terms20 dim. reductionNone
Precision
12
Clustering

• Discover the group structures of URLs
• Non-parametric learning algorithm
• Non-trivial optimization problem
• Should avoid local minima/maxima solution

13
Deterministic Annealing Clustering

• Deterministically avoid local minima
• Tracing global solution by changing level of
  energy
• Analogy to physical annealing process (High ? Low)

14
More Machine Learning Algorithms

• Classification
• To response more quickly to users requests
• Training data based on users input and answering
  questions based on the training results
• Artificial Neural Network, Support Vector
  Machine,
• Trend Detection
• Can be used for prediction/forecasting
• Time-series analysis of tagging activities
• Markov chain model, Fourier transform,

15
Conclusion

• The goal of our Collective Collaborative Tagging
  (CCT) system
• Utilize various data sets
• Provide various information retrieval (IR)
  algorithms
• Help to utilize people-powered knowledge
• Currently various models and algorithms are
  being investigated
• Service interfaces and import/export function
  will be added soon

16
Thank you!! Questions?
jychoi_at_cs.indiana.edu
17
Vector-space Vs. Graph
Vector-space Model
Graph Model
-. q-dimensional vector -. q-by-n matrix
Represen-tation
-. G(V, E) -. V URL, tags, users
-. Distances, cosine, -. O(N2) complexity
Dis-similarity
-. Paths, hops, connectivity, -. O(N3)
complexity
-. Latent Semantic Indexing -. Dimension
reduction schemes -. PCA
Algorithm
-. PageRank, FolkRank, -. Pairwise
clustering -. MDS
18
Pairwise Dissimilarity

• Pairwise clustering
• Input from vector-based model vs. graph model
• How to avoid local minima/maxima? (e.g, K-Means)

Vector-space model
Graph model