aaron sherman

Clustering Categorical Data: An Approach Based on Dynamical Systems (1998)

David Gibson, Jon Kleinberg, Prabhakar Raghavan VLDB Journal: Very Large Data Bases

Aaron Sherman

Presentation

What is this presentation about? Definitions and Algorithms Evaluations with Generated Data Real World test Conclusions + Q&A

Categorize this!

Categorizing int’s are easy, but what about words like “red,” “blue,” “august,” and “Moorthy?”

STIRR – Sieving Through iterated Relational Reinforcement

Why is STIRR Better?

No a Priori Quantization Correlation vs. Categorical Similarity New Methods for Hypergraph Clustering

Definitions

Table of Relational Data – Set T of Tuples– Set of K Fields – many possible values (Columns)

– Abstract Node – each possible field

– Г є T – consists of one node from each field

Configuration – weight wv to each node v –w N(w) – Normalization Function – rescale all

weights so their squares add up to 1 Dynamical System – repeated application of f Fixed Point – point u where f(u) = u

Where is all this going?

Weighting Scheme

To update the weight wv:

– For each tuple Г = {v,u1,…uk-1} containing v• X Г § (u1,…uk-1 )

– Wv Σ Г X Г

N() f(w)

Combining Operator П

Product Operator П: §(w1…wk ) = w1 w2… wk

Non-linear term – encode co-occurrence strongly

Does not converge Relatively small # of large basins Very useful data in early iterations

Combining Operator +

Addition Operator +: §(w1…wk ) = w1 +w2+…+

wk

Linear Does a good job converging

Combining Operator Sp

Sp – Combining Rule: §(w1…wk ) =

Non-linear term – encode co-occurrence strongly

Does a good job converging

Combining Operator Sω

Sω – Limiting version of Sp

Take the largest value among the weights Easy to compute, sum like properties Converges the best of all options shown

Initial Configuration

Uniform Initialization – all weights = 1 Random Initialization – independently

choose o1 for each weight then normalize– Some operators more sensitive to initial

configurations then others

Masking / Modification – specific rule for certain nodes to set to higher or lower value

Run Time - Linear

Quasi-Random Input

Create semi random data, and then add tuples to the data to create artificial clusters– Use this to test whether STIRR works

Questions• # of iterations

• Density of cluster to background

How well does STIRR distil a cluster in nodes with above average co-occurrence

# of iterations Purity

How well does STIRR separate distinct planted clusters?Will the data partition?

How long to partition?

S(A,B) = (|a0 – b0| + |a1 –b1| ) / total nodesClusters A,B, a0 nodes from cluster, and a1nodes at other end

How well does STIRR cope with clusters in a few columns with the rest random?

Want to mask irrelevant factors (columns)

Effect of Convergence Operator Max function is

the best Product rule

does not converge

Sum rule is good, but slow

Real World Data

Papers on theory and Database Systems– (Author 1, Author 2, Journal Year)– The two sets of papers were clearly separated in

the STIRR representation– Done using Sp– Grouped most theoretical papers around 1976

Login Data from IBM Servers

Masked one user who logged in / out very frequently

4 highest weight (similar) users – root, help, 2 administrators names

8pm-12am very similar

Conclusion

Powerful technique to categorize data Relatively fast algorithm O(n) Questions?

Additional References

Data Clustering Techniques - Qualifying Oral Examination Paper - Periklis Andritsos– http://www.cs.toronto.edu/~periklis/p

ubs/depth.pdf