Transcript Απόκτηση Γνώσης για την

Contents
Introduction
Knowledge discovery from text & links
Knowledge discovery from usage data
Important open issues
1
WWW: the new face of the Net
Once upon a time, the Internet was a forum for
exchanging information. Then …
…came
the Web.
The Web introduced
new capabilities …
…and attracted many more people …
…increasing
commercial interest …
…and turning the Net into a real forum …
2
Information overload
…as more people
started using it ...
…attracting even
more people ...
…increasing
…the quantity
the of
quantity
information
of online
on the
information
Web increased...
further...
…and leading to the overload
of information for the users ...
3
WWW: an expanding forum
The Web is large and volatile:
More than 600.000.000 users online
More than 800.000 sign up every day
More than 9.000.000 Web sites
More than 300.000.000.000 pages online
Less than 50% of Web sites will be there next year
… leading to the abundance problem:
“99% of online information is of no
interest to 99% of the people”
4
Information access services
A number of services aim to help the user gain
access to online information and products ...
… but can they really cope?
5
New requirements
Current indexing does not allow for wide coverage:
Less than 5% of the Web covered by search engines.
What I want is hardly ever ranked high enough.
Product information in catalogues is often biased
towards specific suppliers and outdated.
Product descriptions are incomplete and insufficient
for comparison purposes.
‘E’ in ‘E-commerce’ stands for ‘English’: More than
70% of the Web is English.
… and many more problems lead to the conclusion ...
… that more intelligent solutions are needed!
6
A new generation of services
Some have already made their way to the market…
… many more are being developed as I speak …
7
Approaches to Web mining
Primary data (Web content):
Mainly text,
with some multimedia content (increasing)
and mark-up commands including hyperlinks.
Underlying databases (not directly accessible).
 Knowledge discovery from text and links
Pattern discovery in unstructured textual data.
Pattern discovery in the Web graph / hypertext.
8
Approaches to Web mining
Secondary data (Web usage):
Access logs collected by servers,
potentially using cookies,
and a variety of navigational information collected
by Web clients (mainly JavaScript agents).
 Knowledge Discovery from usage data
Discovery of interesting usage patterns, mainly
from server logs.
Web personalization & Web intelligence.
9
Contents
Introduction
Knowledge discovery from text & links
Introduction
Information filtering and retrieval
Ontology learning
Knowledge discovery from usage data
Important open issues
10
Information access
Goals:
Organize documents into categories.
Assign new documents to the categories.
Retrieve information that matches a user query.
Dominating statistical idea:
TFIDF=term frequency * inverse document frequency
Problems on the Web:
Huge scale and high volatility demand automation.
11
Text mining
Knowledge (pattern) discovery in textual data.
Clarifying common misconceptions:
Text mining is NOT about assigning documents to
thematic categories, but about learning document
classifiers.
Text mining is NOT about extracting information
from text, but about learning information extraction
patterns.
Difficulty: unstructured format of textual data.
12
Approaches to text mining
Combination of language engineering (LE),
machine learning (ML) and statistical methods:
LE
MLStats
MLStats
LE
13
Hyperlink information is useful
Information access can be improved by
identifying: authoritative pages (authorities)
and resource index pages (hubs).
Linked pages often contain complementary
information (e.g. product offers).
Thematically related pages are often linked,
either directly or indirectly.
14
Document category modelling
Training documents (pre-classified)
Stopword removal (and, the, etc.)
Stemming (‘played’  ‘play’)
Pre-processing
Bag-of-words coding
Dimensionality reduction Statistical selection/combination of
characteristic terms (MI, PCA)
Machine Learning
Supervised classifier learning
Category models (classifiers)
15
Document category modelling
Example: Filtering spam email.
Task: classify incoming email as spam
and legitimate (2 document categories).
Simple blacklist and keyword-based
methods have failed.
More intelligent, adaptive approaches
are needed (e.g. naive Bayesian
category modeling).
16
Document category modelling
Step 1 (linguistic pre-processing): Tokenization,
removal of stopwords, stemming/lemmatization.
Step 2 (vector representation): bag-of-words or
n-gram modeling (n=2,3).
Step 3 (feature selection): information gain
evaluation.
Step 4 (machine learning): Bayesian modeling,
using word/n-gram frequency.
17
Link structure analysis
Improve information retrieval by scoring Web
pages according to their importance in the
Web or a thematic sub-domain of it.
Nodes with large fan-in (authorities) provide
high quality information.
Nodes with large fan-out (hubs) are good
starting points.
18
Link structure analysis
The HITS algorithm [Kleinberg, ACM Journal 1999]:
Given a set of Web pages, e.g. as generated by a query,
expand the base set by including pages that are linked to by
the ones in the initial set or link to them,
assign a hub and an authority weight to each page,
initialised to 1,
update the authority weight of page p according to the hub
weights of the pages that link to it:
ap 

q |q  p
hq
update the hub weight of page p according to the authority
weights of the pages that it links to:
hp  
aq
q| p  q
repeat the weight update for a given number of times,
return a list of the pages ranked by their weights.
19
Link structure analysis
Interesting issues:
Does the social network hypothesis hold, i.e.,
“authorities are highly cited”? This may be
unrealistic in competitive commercial domains.
What happens if link structure adapts to the
method, e.g. unrelated pages link to each other to
increase their rating?
What about interesting new pages? How will
people get to them?
20
Focused crawling & spidering
Crawling/Spidering: Automatic navigation
through the Web by robots with the aim of
indexing the Web.
Crawling v. Spidering (subjective): inter-site v.
intra-site navigation.
Focused crawling/spidering: Efficient,
thematic indexing of relevant Web pages, e.g.
maintenance of a thematic portal.
Underlying assumption similar to HITS:
thematically similar pages are linked.
21
Focused crawling
Focused crawling [Chakrabarti et al., WWW 1999]:
Given an initial set of Web pages about a topic, e.g. as
found in a Web directory,
use document category modelling to build a topic classifier,
extract the hyperlinks within the initial set of pages and add
them to a queue of pages to be visited,
retrieve pages from the queue,
use the classifier to assess the relevance of retrieved pages,
use a variant of HITS to assign a hub score to pages and the
hyperlinks in the queue,
re-sort the links in the queue according to their hub score,
continue the retrieval of new pages, periodically updating
the score of hyperlinks in the queue.
22
Focused crawling & spidering
Domain-specific spidering:
Goal: retrieve interesting pages, without traversing
the whole site.
Differences from crawling:
The site is much more restricted in size and thematic
diversity than the whole of the Web.
Social network analysis is less relevant within a site (no
hubs and authorities).
Requirement: link scoring using local features, e.g.
the anchor text and the textual context.
23
Information extraction
Goals:
Identify interesting “events” in unstructured text.
Extract information related to the events and store
it in structured templates.
Typical application:
Information extraction from newsfeeds.
Difficulties:
Deals with unstructured or semi-structured text.
Identification of entities and relations.
Usually requires some understanding of the text.
24
A typical extraction system
Morphology
Syntax
Semantics
Discourse
Unstructured text and database
schema (event templates)
Lemmatization (‘said’  ‘say’),
Sentence and word separation.
Part-of-speech tagging, etc.
Shallow syntactic parsing.
Named-entity recognition.
Co-reference resolution.
Sense disambiguation.
Pattern matching.
Structured data (filled templates)
25
Wrappers/fact extraction
Simplified information extraction:
Extract interesting facts from Web documents.
Assumes structure in the documents (usually
dynamically generated from databases).
Reduced demand for pre-processing and LE.
Typical application:
Product comparison services (price, availability, …).
Difficulties:
Semi-structured data.
Different underlying database schemata and
presentation formats.
26
Wrappers/fact extraction
Example:
<HTML><TITLE> Some Country Codes </TITLE>
<BODY><B> Some Country Codes </B> <P>
<B> Congo </B> <I> 242 </I>
<B> Egypt </B> <I> 20 </I>
<B> Greece </B> <I> 30 </I>
<B> Spain </B> <I> 34 </I>
<HR> <B> End </B> </BODY> </HTML>
Wrapper (page P)
Skip past first occurrence of <P> in P
While (next <B> is before next <HR> in P)
For each <l, r>  { (<B>, </B>}) , (<I>, </I>) }
Extract the text between l and r
return <country, code > extracted pairs
Country
Code
Congo
242
Egypt
20
Greece
30
Spain
34
27
Wrapper induction
Training documents (semistructured)
Data pre-processing
Abstraction of mark-up structure
(often omitted)
Database schema (interesting facts)
Machine Learning
Structural/sequence learning
Fact extraction patterns (wrapper)
28
Ontology learning
Training documents (unclassified)
Stopword removal (and, the, etc.)
Stemming (‘played’  ‘play’)
Syntactic/Semantic analysis
Pre-processing
Bag-of-words coding
Dimensionality reduction Hand-made thesauri (Wordnet)
Term co-occurrence (LSI)
Machine Learning Unsupervised learning (clustering
and association discovery)
Ontologies
29
Ontology learning
Hierarchical clustering is most suitable:
Agglomerative clustering
Conceptual clustering (COBWEB)
Model-based clustering (EM-type: MCLUST)
… but flat clustering can also be adapted:
K-means and its variants
Bayesian clustering (Autoclass)
Neural networks (self-organizing maps)
Association discovery (e.g. Apriori) for nontaxonomic relations.
30
Ontology learning
Example: Acquisition of an ontology for tourist
information. [based on Maedche & Staab, ECAI 2000]
31
Ontology learning
Source data: Web pages of tourist sites.
Background knowledge: generic and domain-specific
ontologies.
Target users: Tourist directories, large travel agencies.
Goals:
Identify types of page (e.g. room descriptions) and
terms/entities inside pages (e.g. hotel addresses).
Identify taxonomic relations between concepts (e.g.
accommodation – hotel).
Identify non-taxonomic relations between concepts (e.g.
accommodation – area).
32
Ontology learning
Heavy linguistic pre-processing:
Syntactic analysis,
e.g. verb subcategorization frames:
verb(arrive) -> prep(at), dir_obj(Torino).
Semantic analysis,
e.g. named entity recognition:
‘Via Lagrange’ -> Street name
e.g. special dependency relations:
‘Hotel Concord in Torino’
33
Contents
Introduction
Knowledge discovery from text & links
Knowledge discovery from usage data
Personalization on the Web
Data collection and preparation issues
Personalized assistants
Discovering generic user models
Sequential pattern discovery
Knowledge discovery in action
Important open issues
34
Personalized information access
sources
personalization
server
receivers
35
Personalization v. intelligence
Better service for the user:
Reduction of the information overload.
More accurate information retrieval and
extraction.
Recommendation and guidance.
36
Personalized assistants
Personalized crawling [Liebermann et al., ACM Comm., 2000]:
The system knows the user (log-in).
It uses heuristics to extract “important” terms from the Web pages
that the user visits and add them to thematic profiles.
Each time the user views a page, the system:
searches the Web for related pages,
filters them according to the relevant thematic profile,
and constructs a list of recommended links for the user.
The Letizia version of the system searches the Web locally,
following outgoing links from the current page.
The Powerscout version uses a search engine to explore the Web.
37
Personalized assistants
Adaptive Web interfaces [Jörding, UM 1999]:
The TELLIM system collects user information, (e.g. a
selection of a link) using a Java applet .
User information is used as training data in order to create
generic models reflecting the users’ interest in different
products.
The system creates short-term personal models using the
generic models and the current user’s behavior.
Web pages containing more detailed information about these
products, together with multimedia content and VRML
presentations are created dynamically and presented to the
users.
38
User modelling
Basic elements:
Constructing models that can be used to adapt the
system to the user’s requirements.
Different types of requirement: interests (sports
and finance news), knowledge level (novice expert), preferences (no-frame GUI), etc.
Different types of model: personal – generic.
Knowledge discovery facilitates the
acquisition of user models from data.
39
User Models
User model (type A):
[PERSONAL]
User x -> sports, stock market
User model (type B):
[PERSONAL]
User x, Age 26, Male -> sports, stock market
User community:
[GENERIC]
Users {x,y,z} -> sports, stock market
User stereotype:
[GENERIC]
Users {x,y,z}, Age [20..30], Male -> sports, stock
market
40
Generic user models
Stereotypes: Models that represent a type
of user, associating personal characteristics
with parameters of the system,
e.g. Male users of age 20-30 are interested in sports
and politics.
Communities: Models that represent a
group of users with common preferences,
e.g. Users that are interested in sports and politics.
41
Learning user models
Community 2
Community 1
User 1
User 2
User 3
User 4
User 5
User
communities
User
models
Observation of the users interacting with the system.
42
Knowledge discovery process
Data collection
Data pre-processing
Pattern discovery
Collection of usage data by
the server and the client.
Data cleaning, user identification,
session identification
Construction of user models
Knowledge post-processing Report generation, visualization,
personalization module.
43
Pre-processing usage data
Cleaning:
Log entries that correspond to error responses.
Trails of robots.
Pages that have not been requested explicitly by the user
(mainly image files, loaded automatically). Should be
domain-specific.
User identification:
Identification by log-in.
Cookies and Javascript.
Extended Log Format (browser and OS version).
Bookmark user-specific URL.
Various other heuristics.
44
Pre-processing usage data
User session/Transaction identification in log files:
Time-based methods, e.g. 30 min silence interval. Problems
with cache. Partial solutions: special HTTP headers, Java
agents.
Context-based methods: e.g. separate pages into
navigational and content and impose heuristics on the type
of page that a user session may consist of.
User sessions can be subdivided into smaller transaction
sequences, e.g. by identifying a “backward reference” in the
sequence of requests.
Encoding of training data:
Bag-of-pages representation of sessions/transactions.
Transition-based representation of sessions/transactions.
Manually determined features of interest.
45
Collaborative filtering
Information filtering according to the
choices of similar users.
Avoids semantic content analysis.
Cold-start problem with new users.
Approaches:
memory-based learning,
model-based clustering,
item-based recommendation.
46
Memory-based learning
Nearest-neighbour approach:
Construct a model for each user. Often use explicit
user ratings for each item.
Index the user in the space of system parameters,
e.g. item ratings.
For each new user,
index the user in the same space, and
find the k closest neighbours.
Simple metrics to measure the similarity between users,
e.g. Pearson correlation.
Recommend the items that the new user has not
seen and are popular among the neighbours.
47
Model-based clustering
Clustering users into communities.
Methods used:
Conceptual clustering (COBWEB).
Graph-based clustering (Cluster mining).
Statistical clustering (Autoclass).
Neural Networks (Self-Organising Maps).
Model-based clustering (EM-type).
BIRCH.
Community models: cluster descriptions.
48
Model-based clustering
0,9
0,9
0,9
0,8
0,4
0,1
0,1
0,5
0,5
49
Item-based recommendation
Focus on item usage in the profiles, instead
of the users themselves.
Practically useful in e-commerce, e.g. crosssell recommendations.
Simple modification to the clique-based
clustering method: graph of items instead of
graph of users.
Related to frequent itemset discovery in
association rule mining.
50
Item-based recommendation
Sports
0,9
Politics
0,9
0,9
0,8
0,4
Finance
0,1
0,1
World
0,5
0,5
51
Contents
Introduction
Knowledge discovery from text & links
Knowledge discovery from usage data
Personalization on the Web
Data collection and preparation issues
Personalized assistants
Discovering generic user models
Sequential pattern discovery
Knowledge discovery in action
Important open issues
52
Sequential pattern discovery
Identifying navigational patterns, rather than
“bag-of-page” models.
Methods:
Clustering transitions between pages.
First-order Markov models.
Probabilistic grammar induction.
Association-rule sequence mining.
Path traversal through graphs.
Personal and community navigation models.
53
Sequential pattern discovery
Clique-based transition clustering; small modification of
the model-based item clustering approach: an item is a
transition between pages.
Sports
->Politics
0,9
0,9
0,9
0,8
0,4
Sports
->Finance
Finance
->Politics
0,1
0,1
0,5
0,5
Finance
->Sports
54
References
J. Borges and M. Levene, Data mining of user navigation patterns. Proceedings of Workshop on Web Usage Analysis
and User Profiling (WEBKDD), in conjunction with ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining. San Diego, CA., pp. 31-36.
S. Chakrabarti, M. H. van den Berg, B. E. Dom, Focused Crawling: a new approach to topic-specific Web resource
discovery, Proceedings of the Eighth International World Wide Web Conference (WWW), Toronto, Canada, May 1999.
T. Jörding, T, A Temporary User Modeling Approach for Adaptive Shopping on the We`, In Proceedings of the 2nd
Workshop on Adaptive Systems and User Modeling on the WWW, UM'99, Banff, Canada, 1999.
J. Kleinberg. Authoritative sources in a hyperlinked environment. Journal of the ACM, v. 46, 1999.
H. Lieberman, C. Fry and L. Weitzman. Exploring the Web with Reconnaissance Agents, Communications of the ACM,
August 2001, pp. 69-75.
A. Maedche, S. Staab. Discovering Conceptual Relations from Text. In: W.Horn (ed.): ECAI 2000. Proceedings of the
14th European Conference on Artificial Intelligence (ECAI), Berlin, August 21-25, 2000.
A. McCallum, D. Freitag and F. Pereira, Maximum Entropy Markov Models for Information Extraction and
Segmentation, Proceedings of the International Conference on Machine Learning (ICML), Stanford, CA, 2000, pp.
591-598.
I. Muslea , S. Minton and C. Knoblock , STALKER: Learning extraction rules for semistructured Web-based
information sources. Proceedings of the National Conference on Artificial Intelligence (AAAI), Madison, Wisconsin,
1998.
C. Nédellec, Corpus-based learning of semantic relations by the ILP system, Asium, Learning Language in Logic,
Cussens J. and Dzeroski S. (Eds.), Springer Verlag, September 2000.
J. Rennie and A. McCallum. Efficient Web Spidering with Reinforcement Learning. Proceedings of the International
Conference on Machine Learning (ICML), 1999.
E. I. Schwartz. Webonomics. New York: Broadway books, 1997.
E. Schwarzkopf, An adaptive Web site for the UM2001 conference. Proceedings of the Workshop on Machine Learning
for User Modeling, in conjunction with the International Conference on User modelling (UM), pp 77-86, 2001.
55