23
1 Automatic Classification of Bookmarked Web Pages Chris Staff Second Talk February 2007

Automatic Classification of Bookmarked Web Pages

Embed Size (px)

DESCRIPTION

Automatic Classification of Bookmarked Web Pages. Chris Staff Second Talk February 2007. Tasks. Representation of bookmark categories Two clustering/similarity algorithms Extra utility User interface Evaluation Write up report. Overview. Represent bookmark categories - PowerPoint PPT Presentation

Citation preview

Page 1: Automatic Classification of Bookmarked Web Pages

1

Automatic Classification of Bookmarked Web Pages

Chris StaffSecond Talk

February 2007

Page 2: Automatic Classification of Bookmarked Web Pages

2

Tasks

1. Representation of bookmark categories

2. Two clustering/similarity algorithms

3. Extra utility4. User interface5. Evaluation6. Write up report

Page 3: Automatic Classification of Bookmarked Web Pages

3

Overview

• Represent bookmark categories– We’re starting with populated bookmark files, so use ‘How Did I Find That?’ approach

– Plus another, individual approach

• When a page is to be bookmarked– If referrer page is available, identify topic of page

– Otherwise, identify page topic using ‘How Did I Find That?’ approach

• Compare current topic topic to bookmark category representations

Page 4: Automatic Classification of Bookmarked Web Pages

4

Overview

• The representation of the bookmark categories and the representation of the page to be bookmarked need to be compatible, so that we can compare them

• Clustering techniques/Similarity measures will tell us cluster membership/degree of similarity

Page 5: Automatic Classification of Bookmarked Web Pages

5

Clustering/Similarity

• Given a sets of documents, D1…Dn, and an individual document di, how can we tell to which Dj di “belongs”?

• What features of documents in Dj and document Di should we use?

Page 6: Automatic Classification of Bookmarked Web Pages

6

The Vector Space Model of IR

• In VSM, documents are plotted into vector space – the nearest neighbours belong to the same cluster

– a query is plotted into vector space too, and its nearest neighbours are the relevant documents

• Typically, a document is represented by term features

G. Salton and C. Buckley. (1988). Term-weighting approaches in automatic text retrieval. Information Processing & Management, 24(5):513--523.p18-wong (Generalised Vector Space Model).pdf - look at refs [1],[2],[3] for original work

Page 7: Automatic Classification of Bookmarked Web Pages

7

Vector Space Model

• Documents are represented as m-dimensional vectors or “bags of words”

• m is the size of the vocabulary

• wk = 1, indicates term is present in document

• wk = 0, indicates term is absent

• dj = <1,0,0,1,...,0,0>

Page 8: Automatic Classification of Bookmarked Web Pages

8

Vector Space Model

Page 9: Automatic Classification of Bookmarked Web Pages

9

Vector Space Model

• The query is then plotted into m-dimensional space and the nearest neighbours are the most relevant

• However, the results set is usually presented as a list ranked by similarity to the query

Page 10: Automatic Classification of Bookmarked Web Pages

10

Context…

• What’s “180374”?• What’s “Chris”?• What’s “R2D2”?• What’s “B4Y2”?• Why were we able to tell what the first three might be, but not the last one?

Page 11: Automatic Classification of Bookmarked Web Pages

11

Context…

• Information is data in context– What it is

• Knowledge is information in context– How and when to use it

Page 12: Automatic Classification of Bookmarked Web Pages

12

Context and HDIFT

• Web-based documents may contain more than one topic– How can we identify in what the user might be interested?

• We want to identify a document’s context so that we can understand what a user might be interested in– What can provide a context?

Page 13: Automatic Classification of Bookmarked Web Pages

13

How Did I Find That?

• HDIFT is a technique for finding documents related to a user’s bookmark categories– represents each category in a novel way;– extracts a query from the category representation (centroid)

– Submits query to 3rd party search engine– Shows top-10 results to user

• First technique that you must use

Page 14: Automatic Classification of Bookmarked Web Pages

14

How Did I Find That?

• Building a category representation– Instead of using term features from the documents in the category, use terms from parents of documents

– Why?– How many parents should we use?

•If we know parent, then referrer page only

•Otherwise, ?

Page 15: Automatic Classification of Bookmarked Web Pages

15

How Did I Find That?

• HDIFT uses 20 parents obtained using the ‘link:’ operator in Google

• Find “context block” in parent that contains link to bookmarked doc– There may be more than one link!

• Write context block to file• Once obtained all context blocks of all bookmarks in category, use SWISH-E to index them

• Create centroid representation

Page 16: Automatic Classification of Bookmarked Web Pages

16

How Did I Find That?

• Each category will have an HDIFT centroid representation

• When user bookmarks a page, use two approaches to determine the context of the page– One will use the referrer page only– The other will generate a representation of the document, based on its parents (i.e., HDIFT)

Page 17: Automatic Classification of Bookmarked Web Pages

17

How Did I Find That?

• Each representation of the bookmarked page is used to determine category membership– Obviously, for evaluation, you won’t have access to the parent of the page to be bookmarked (classified), as the page has been removed from the category containing it

• What difference, if any, is there between the two document representations?

Page 18: Automatic Classification of Bookmarked Web Pages

18

2nd category representation method

• This is largely up to you to come up with

• What’s important is that you try to identify what is likely to be important to the user– Why did the user put those documents into the same category?

– Why is the user bookmarking the current page?

• Feel free to use any clustering algorithm/similarity measure you like

Page 19: Automatic Classification of Bookmarked Web Pages

19

Judging similarity

• Typically, we use cosine similarity which measures the cosine angle or Euclidean Distance between two documents in vector space

Page 20: Automatic Classification of Bookmarked Web Pages

20

Cosine Similarity Measure

sim(Q,D) =

wqk • wdkk=1

m

(wqk )2 •

k=1

m

∑ (wdk )2

k=1

m

From IR vector space model.pdf

Page 21: Automatic Classification of Bookmarked Web Pages

21

Cosine Similarity Measure

• For the approach based on HDIFT, please use CSM

• For other approach, you can use anything you like, e.g., Information Filtering, Clustering– Including CSM, if you like

Page 22: Automatic Classification of Bookmarked Web Pages

22

Indexing Documents

• In HDIFT, I use SWISH-E to index documents to obtain a cluster centroid and extract top-ranking keywords to form a query

• You have a requirement to impose only a maximum 2 second overhead on average to classify a page that has been bookmarked

• The categories can be indexed, represented, and re-indexed when the cluster membership changes, off-line or in the background

Page 23: Automatic Classification of Bookmarked Web Pages

23

Indexing Documents

• You *do not need* to use SWISH-E, though you may need to use something, and there is nothing to stop you using SWISH-E :-)

• You may develop your own lightweight indexing system; use another 3rd party (ideally, open source) system, etc.