Enhanced topic distillation using text, markup tags, and hyperlinks

1
Enhanced topic distillation using text, markup
tags, and hyperlinks

• Soumen ChakrabartiMukul JoshiVivek Tawde
• www.cse.iitb.ac.in/soumen

2
Topic distillation
Keyword query

• Given a query or some example URLs
• Collect a relevant subgraph (community) of the
  Web
• Bipartite reinforcement between hubs and
  authorities
• Prototypes
• HITS, Clever, SALSA
• Bharat and Henzinger

Searchengine
Base set
Expanded set
Root set
3
Two issues

• How to collect the base set
• Radius-1 expansion is arbitrary
• Content relevance must play a role
• How to spread prestige along links
• Instability of HITS (Borodin, Lempel Zheng)
• Stability of PageRank (Zheng)
• Stochastic variants of HITS (Lempel)
• Need better recall collecting base graph
• Need accurate boundaries around it

4
Challenges and limitations

• Topic distillation results deteriorating
• Web authoring style in flux since 1996
• Complex pages, templates, cloaks
• File or page boundary less meaningful
• Clique attacksrampant multi-host nepotism
  via rings, ads, banner exchanges
• Models too simplistic
• Hub and authority symmetry is illusory
• Coarse-grain hub model leaks authority
• Ad-hoc linear segmentation not content-aware

5
Clique attacks!
Irrelevantlinks formpseudo-community
Relevant regionsthat lead to inclusionof page
in base set
6
Benign drift and generalization
Remainingsectionsgeneralize and/or drift
This sectionspecializes onShakespeare
7
A fine-grained hypertext model
lthtmlgtltbodygt lttable gt lttrgtlttdgt lttable gt
lttrgtlttdgtlta hrefhttp//art.qaz.comgtartlt/agtlt/td
gtlt/trgt lttrgtlttdgtlta hrefhttp//ski.qaz.comgtsk
ilt/agtlt/tdgtlt/trgt lt/tablegt lt/tdgtlt/trgt lttrgtlttdgt
ltulgt ltligtlta hrefhttp//www.fromages.com
gtFromages.comlt/agt French cheeselt/ligt
ltligtlta hrefhttp//www.teddingtoncheese.co.ukgtTe
ddingtonlt/agt Buy onlinelt/ligt
lt/ulgt lt/tdgtlt/trgt lt/tablegt lt/bodygtlt/htmlgt
8
Preliminary approaches

• Apply HITS to fine-grained base graph
• Blocked reinforcement
• Model DOM trees as resistance or flow networks
• Ad-hoc decay factors
• Apply BH outlier elimination to every DOM node
• Hot absorbs cold, includes drift-enhancing links

Cold
Warm enoughto figure asone hub
Hot
9
Generative model for hub text
Global termdistribution ?0

• Global hub text distribution ?0 relevant to given
  query
• Authors use internal DOM nodes to hierarchically
  specialize ?0 into ?I
• At a certain frontier, local models are frozen
  and text generated

Progressivedistortion
Modelfrontier
?I
Other pages
10
Examples using the binary model

• Binary model
• Code length for document d
• Cost for specializing a term distribution

11
Discovering the frontier
Referencedistribution?0
Cumulative distortion cost KL(?0??u)
KL(?u??v)

• Use u to directly generate text snippets in the
  subtree rooted at u
• Expand to children v and use different params for
  each tree
• Greedily pick better local choice

u
v
Dv
12
Exploiting co-citation in our model
1
2
Initial values ofleaf hub scores target auth
scores
Segment treeusing hub scores
Have reasonto believethese could be good too
0.10
0.20
0.01
0.06
0.05
0.13
3
4
Aggregate hubscores are copiedback to leaves
0.12
Knownauthorities
0.10
0.20
0.13
0.10
0.20
0.12
0.12
0.12
0.13
Frontier microhubsaccumulate scores
Non-linear transform, unlike HITS
13
Complete algorithm

• Collect root set and base set
• Pre-segment using text and mark relevant
  micro-hubs to be pruned
• Assign only root set authority scores to 1s
• Iterate
• Transfer from authority to hub leaves
• Re-segment hub DOM trees using link text
• Smooth and redistribute hub scores
• Transfer from hub leaves to authority roots
• Report top authority and hot microhubs

14
Experimental setup

• Large data sets
• 28 queries from Clever, gt20 topics from Dmoz
• Collect 200010000 pages per query/topic
• Several million DOM nodes and fine links
• Find top authorities using various algos
• Measurements anecdotes
• For ad-hoc query, measure cosine similarity of
  authorities with root-set centroid in vector
  space
• Compare HITS, DOM, DOMText

15
Avoiding topic drift via micro-hubs
Query cyclingNo danger of topic drift
Query affirmative actionTopic drift from
software sites
16
Empirical convergence

• Convergence for all queries within 20 iterations
• Faster convergence for drift-free graphs, slower
  for graphs that posed a danger of topic drift
• Very important to not set all auth scores to gt 0

17
Results for the Clever benchmark

• Take top 40 auths
• Find average cosine similarity to root set
  centroid
• HITS lt DOMText lt DOM similarity
• DOM alone cannot prune well enough most top
  auths from root set
• HITS drifts often

18
Dmoz experiments and results

• 223 topics from http//dmoz.org
• Sample root set URLs from a class c
• Top authorities not in root set submitted to
  Rainbow classifier
• ?d Pr(c d) is the expected number of relevant
  documents
• DOMText best

DMoz
Rainbowclassifier
Train
Sample
Test
Expanded set
Music
Root set
Top authority
19
Anecdotes

• amusement parks http//www.411fun.com/THEMEPARK
  Sleaks authority via nepotistic links to
  www.411florists.com, www.411fashion.com,
  www.411eshopping.com, etc.
• New algorithm reduces drift
• Mixed hubs accurately segmented, e.g. amusement
  parks, classical guitar, Shakespeare and sushi
• Mixed hubs and clique attacks rampant

20
Application surfing like humans
Focused Crawling Train a topic classifierInitiali
ze priority queue to a few sample URLs about
a topicAssume they have relevance 1Repeat
Fetch page most relevant to topic Estimate
relevance R using classifier Guess that all
outlinks have relevance R Add outlinks to
priority queue
?
!
?

• Problem average out-degree is too high (10)
• Discovering irrelevance after 10X more work
• Can we use DOM and text to bias the walk?

21
Preliminary results
Relevance R1
Featurescollected fromsource pageDOM
Relevance R2
Promising andunpromisingclicks
Feedback
Standardfocusedcrawler
Meta-learner
22
Summary

• Hypertext shows complex idioms, missed by
  coarse-grained graph model
• Enhanced fine-grained distillation
• Identifies content-bearing hot micro-hubs
• Disaggregates hub scores
• Reduces topic drift via mixed hubs and
  pseudo-communities
• Application online reinforcement learning
• Need probabilistic combination of evidence from
  text and links