Probabilistic Clustering-Projection Model for Discrete Data Shipeng Yu 1,2, Kai Yu 2, Volker Tresp 2, Hans-Peter Kriegel 1 1 Institute for Computer Science,

Probabilistic Clustering-Projection Model

for Discrete Data

Shipeng Yu1,2, Kai Yu2, Volker Tresp2, Hans-Peter Kriegel1

1Institute for Computer Science, University of Munich2Siemens Corporate Technology, Munich, Germany

October 2005

2

Outline

Motivation Previous Work The PCP Model Learning in PCP Model Experiments Conclusion and Future Work

3

Motivation

We model discrete data in this work Fundamental problem for data mining and machine learning In “bag-of-words” document modelling: document-word pairs In collaborative filtering: item-rating pairs

Properties The data can be described as a big matrix with integer entries The data matrix is normally very sparse (>90% are zeros)

w1 w2 ¢¢¢ wVd1 2 0 ¢¢¢ 1d2 0 1 ¢¢¢ 2...

......

......

dD 1 1 ¢¢¢ 0

Documents

Words

Occurrences

4

Data Clustering

Goal: Group similar documents together For continuous data: Distance-based similarity (k-means)

Iteratively minimize a distance-based cost function Equivalent to a Gaussian mixture model

For discrete data: Occurrence-based similarity

Similar documents should have similar occurrences of words No Gaussianity holds for discrete data

w1 w2 ¢¢¢ wVd1 2 0 ¢¢¢ 1d2 0 1 ¢¢¢ 2...

......

......

dD 1 1 ¢¢¢ 0

5

Data Projection

Goal: Find a low-dimensional feature mapping For continuous data: Principal Component Analysis

Find orthogonal dimensions to explain data covariance For discrete data: Topic detection

Topics explain the co-occurrences of words Topics are not orthogonal, but independent

w1 w2 ¢¢¢ wVd1 2 0 ¢¢¢ 1d2 0 1 ¢¢¢ 2...

......

......

dD 1 1 ¢¢¢ 0

z1 ¢¢¢ zK

6

Projection versus Clustering

They are normally modelled separately But why not jointly?

More informative projection better document clusters

Better clustering structure better projection for words

There should be a stable situation

And how? PCP Model Well-defined generative model for the data Standard ways for learning and inference Generalizable to new data

w1 w2 ¢¢¢ wVd1 2 0 ¢¢¢ 1d2 0 1 ¢¢¢ 2...

......

......

dD 1 1 ¢¢¢ 0

z1 ¢¢¢ zK

7

Two-sided clustering [Hofmann & Puzicha 98]: Same problem as PLSI Discrete-PCA [Buntine & Perttu 03]: Similar to LDA in spirit TTMM [Keller & Bengio 04]: Lack a full Bayesian explanation

Previous Work for Discrete Data

PLSI [Hofmann 99] First topic model Not well-defined generative model

LDA [Blei et al 03] State-of-the-art topic model Generalize PLSI with Dirichlet prior No clustering effect is modelled

NMF [Lee & Seung 99] Factorize the data matrix Can be explained as a cluster

ing model No projection of words is dire

ctly modelled

Projection model Clustering model

w1 w2 ¢¢¢ wVd1 2 0 ¢¢¢ 1d2 0 1 ¢¢¢ 2...

......

......

dD 1 1 ¢¢¢ 0

z1 ¢¢¢ zKµw1 w2 ¢¢¢ wV

d1 2 0 ¢¢¢ 1d2 0 1 ¢¢¢ 2...

......

......

dD 1 1 ¢¢¢ 0

Joint Projection-Clustering model

8

PCP Model: Overview

Probabilistic Clustering-Projection Model A probabilistic model for discrete data A clustering model using projected features A projection model with structural data

Learning in PCP model: Variational EM Exactly equivalent to iteratively performing cluste

ring and projection operations Guaranteed convergence

9

PCP Model: Sampling Process

…... …... …... …...

Clustering Projection

clusterµ1

Weights ¼

Clustercenters

µ

Projection

¯

documentwD

word

…...

word wD ;N D

wD ;N 1

word

…...

word

w1;1

w1;N 1

clusterµD

topic

topic zD ;N D

zD ;N 1

…...

topic

topic

z1;1

z1;N 1

…...

Clustering model using projected featuresProjection model with structural dataD documents M clusters K topics V words

®

¸

Dirichlet

Dirichlet

Multinomial

Multinomial Multinomial

documentw1

10

PCP Model: Plate Model

Likelihood

Model Parameters

Latent VariablesObservations

Clustering Model

Projection Model

11

Learning in PCP Model

We are interested in the posterior distribution

The integral is intractable Variational EM learning

Approximate the posterior with a variational distribution

Minimize the KL-divergence Variational E-step: Minimize w.r.t. variational parameters Variational M-step: Minimize w.r.t. model parameters

Iterate until convergence

DK L (qjjp̂)

Variational Parameters

Dirichlet Dirichlet Multinomial Multinomial

12

Update Equations

Equations can be separated to clustering updates and projection updates

Variational EM learning corresponds to iteratively performing clustering and projection until convergence

Clustering Updates

Projection Updates

13

Clustering UpdatesUpdate soft cluster assignments, P (cd =m)

Update cluster centers Update cluster weights

Prior term

Likelihood term

Prior termLikelihood term

Sufficient Projection term

14

Projection UpdatesUpdate word projection, P (zd;n = k)

Update projection matrix Empirical estimate

Sufficient Clustering term

15

PCP Learning Algorithm

Clustering Updates

Projection Updates

Sufficient Clustering term

Sufficient Projection term

16

Experiments

Methodology Document Modelling: Compare model generalization Word Projection: Evaluate topic space Document Clustering: Evaluate clustering results

Data sets 5 categories in Reuters-21578: 3948 docs, 7665 words 4 categories in 20Newsgroup: 3888 docs, 8396 words

Preprocessing Stemming and stop-word removing Pick up words that occur at least in 5 documents

17

Case Study Run on a 4-group subset of 20Newsgroup data

Car

Bike

Baseball

Hockey

18

Exp1: Document Modelling

Goal: Evaluate generalization performance Methods to compare

PLSI: A “pseudo” form for generalization LDA: State-of-the-art method

Metric: Perplexity

90% for training and 10% for testing

Perp(Dtest) = exp(¡P

d lnp(wd)=P

d jwdj)

19

Exp2: Word Projection

Goal: Evaluate the projection matrix Methods to compare: PLSI, LDA We train SVMs on the 10-dimensional space after projection Test classification accuracy on leave-out data

¯

Reuters Newsgroup

20

Exp3: Document Clustering

Goal: Evaluate clustering for documents Methods to compare

NMF: Do factorization for clustering LDA+k-means: Do clustering on the projected space

Metric: normalized mutual information

21

Conclusion

PCP is a well-defined generative model PCP models clustering and projection jointly Learning in PCP corresponds to an iterative

process of clustering and projection PCP learning guarantees convergence Future work

Large scale experiments Build a probabilistic model with more factors

Thank you!

Questions?