lda training system xueminzhao@tencent.com 8/22/2012

Post on 17-Dec-2015

214 Views

Category:

Documents

1 Downloads

Preview:

Click to see full reader

TRANSCRIPT

LDA Training System

xueminzhao@tencent.com8/22/2012

Outline

• Introduction

• SparseLDA

• Rethinking LDA: Why Priors Matter

• LDA Training System Design: MapReduce-LDA

Outline

• Introduction

• SparseLDA

• Rethinking LDA: Why Priors Matter

• LDA Training System Design: MapReduce-LDA

Problem – Text Relevance

• Q1: apple pie• Q2: iphone crack

• Doc1: Apple Computer Inc. is a well known company located in California, USA.

• Doc2: The apple is the pomaceous fruit of the apple tree, spcies Malus domestica in the rose.

Topic Models

Topic Model – Generative Process

Topic Model - Inference

Latent Dirichlet Allocation

Outline

• Introduction

• SparseLDA

• Rethinking LDA: Why Priors Matter

• LDA Training System Design: MapReduce-LDA

Gibbs Sampling for LDA

Gibbs Sampling for LDA

Document-Topic Statistics

Topic-Word Statistics

For each token,

For each token,

For each token,

For each token,

For each token,

Sample a new topic

For each token,

Summary so far

The normalizing constant

The normalizing constant

The normalizing constant

Statistics are sparse

Summary so far

Huge savings: time and memory

Outline

• Introduction

• SparseLDA

• Rethinking LDA: Why Priors Matter

• LDA Training System Design: MapReduce-LDA

Priors for LDA

Priors for LDA

Priors for LDA

Priors for LDA

Priors for LDA

Comparing Priors for LDA

Optimizing m

Selecting T

Outline

• Introduction

• SparseLDA

• Rethinking LDA: Why Priors Matter

• LDA Training System Design: MapReduce-LDA

Overview

MapReduce Jobs

Scalability

• Hypothesis- memory 40GB per machine;- 5 words per doc.

• Scalability- if #<docs> <= 1,000,000,000, no #<topics> limit;- if #<topics> < 14,000, no #<docs> limit.

Experiment for Correctness Validation

References• D. Blei, Andrew Ng, and M. Jordan, Latent Dirichlet Allocation, JMLR2003.• Thomas L. Griffiths, and Mark Steyvers, Finding scientific topics, PNAS2004.• Gregor Heinrich, Parameter estimation for text analysis, Technical Report, 2009.• Limin Yao, David Mimno, and Andrew McCallum. Efficient Methods for Topic

Model Inference on StreamingDocument Collections. KDD'09.• Hanna M. Wallach, David Mimno, and Andrew McCallum, Rethinking LDA: Why

Priors Matter, NIPS2009.• David Newman, Arthur Asuncion, Padhraic Smyth, and Max Welling, Distributed

Inference for Latent Dirichlet Allocation, NIPS2007.• Yi Wang, Hongjie Bai, Matt Stanton, Wen-Yen Chen, and Edward Y. Chang, PLDA:

Parallel Latent Dirichlet Allocation for Large-scale Applications, AAIM2009.• Xueminzhao. LDA design doc. http://x.x.x.x/~

xueminzhao/html_docs/internal/modules/lda.html.

Thanks!

top related