Inferring Demographic Attributes of Anonymous Internet Users

1
Topic
Inferring Demographic Attributes of Anonymous
Internet Users
Web Mining Seminar Qian, Jun
2
Structure

• 1 Abstract
• 2 Introduction
• 3 Approach
• 4 Conclusion

3
Abstract

• Anonymous internet uses
• Advertisement demographic attributes
• Usage information
• Latent Semantic Analysis
• Neural Model

4
Structure

• 1 Abstract
• 2 Introduction
• 3 Approach
• 4 Conclusion

5
The Problem

• Web advertisers want to target customers with
  certain demographic attributes
• Most Internet users are anonymous

6
The Solution

• 1 Post the ad on relevant web-sites
• 2 Wait for the search term of the users
• 3 Make survey
• Anyway...

7
Target Of This Research

• Build a high-quality database to establish the
  possibility of inferring up to 6 demographic
  factors to those whose demographic information is
  not otherwise available

8
Methodology

• Collect usage information
• Prepare usage information-LSA
• Create a neural model

9
LSA Overview

• Information retrieval technique to create vector
• Like create a single vector representing an
  internet user of interest
• Combination of vectors and a vector

10
Vector-space Information Retrieval

• Documents are vectors of terms d(t1,t2,tn)
• A query is a vector of terms as well
  q(t1,t2,tn)
• Term-by-Document matrix/Row-by-Column
  

11
The Singular Value Decomposition (SVD)

• Decompose txd term-by-document matrix A , A
  TSDt, into
• a txk matrix T of term vectors
• the transpose of a dxk matrix of document vectors
• a kxk diagonal matrix S of singular value ,
  define 100ltklt300

12
Structure

• 1 Abstract
• 2 Introduction
• 3 Approach
• 4 Conclusion

13
Collect Background Information

• Target----- a collection of documents consisting
  of popular web pages accessed by internet users
• Procedure--a web-crawler was used,web pages with
  less than 4k bytes in size were accessed

14
Create A Term By Document Matrix

• Target------ Create term-by-document matrix from
  the document collection as input
• Procedure--SMART software from Cornell University

15
Perform A SVD On The Term-document Matrix

• Target------an LSA vector representing all the
  usage data associated with each Internet user of
  interest
• Procedure--Compute the sum of the vectors in the
  matrix T, scale the resulting vector by the
  inverse of the matrix S, add the document vectors
  representing the web pages accessed by the
  Internet user to the pseudo-document vector
  created in the previous step

16
Create A Neural Model To Test The Hypothesis

• Model----- 3-layer neural model
• Training--- independent dependent variables
• Number----40000 observations for training, 20000
  observations for validation

17
Variables
Variables Gender Age Under 18 Age 55 Income
Under 50000 Marital Status Some College
Education Children in the Home
Possible Values male, female true,false true,false
true,false single,married true,false true,false
18
Training

• Training Data contain equal proportions of the
  values of the dependent variable under
  consideration
• Validation Datacontain true proportions of the
  values of the dependent variable under
  consideration

19
Structure

• 1 Abstract
• 2 Introduction
• 3 Approach
• 4 Conclusion

20
Conclusion

• It is really possible to make demographic
  inferences about Internet users for whom
  information is not otherwise available
• Privacy concern