A MultiAgent System for Personalized Press Reviews

1
A MultiAgent System for Personalized Press Reviews

• A. Addis, G. Cherchi,A. Manconi, and E. Vargiu
• Intelligent Agents and Soft-Computing Group
• Dept. of Electrical and Electronic Engineering
• University of Cagliari
• (Italy)

2
Outline

• Introduction
• The Proposed MAS
• The Abstract Architecture
• The Concrete Architecture
• Experimental Results
• Conclusions and Future Work

3
Introduction
4
Motivations

• Internet offers a growing amount of data spread
  on heterogeneous sources (web services,
  distributed databases, web pages, etc.)
• Very difficult for the users to select contents
  according to their personal interests

5
Motivations
I'm interested on concert musics in Sardinia
6
Motivations

• Support the user through an automated system,
  able to
• Retrieve and extract information from
  heterogeneous sources
• Select the contents really deemed relevant for
  the user, according to her/his personal interests.

7
The Proposed Approach

• A multiagent system able to
• take into account users needs and
  preferences(Personalization)
• adapt to changes occurring in the
  environment(Adaptation)
• interact with other agents and the
  user(Cooperation)

8
The Proposed Approach

• Why a multiagent system?
• a centralized classification system may be
  quickly overwhelmed by a large and dynamic
  document stream (like daily-updated news)
• internet is intrinsicly large and distributed,
  thus it gives the opportunity to take advantage
  of distributed computing paradigms and resources

9
The Proposed MAS

• The Abstract Architecture

10
The AbstractArchitecture

• Information Extraction
• Text Categorization
• Users Feedback

11
Information Extraction

• The information extraction module extracts data
  from web sources through specialized wrappers
• Two kinds of wrappers have been implemented
• HTML/XHTML wrapper
• RSS wrapper

12
Information Extractionthe HTML wrapper

• The process of extracting data from HTML
  typically consists of two steps
• learning page structure to detect the tags
  containing object
• performing structured data extraction applying
  the mapping function to populate the
  corresponding data repository

13
Information Extractionthe RSS wrapper

• Extracts information from online newpapers
  articles in RSS format
• Being RSS a well-structured format, it is very
  simple to process RSS pages

14
Text Categorization

• The Text Categorization module progressively
  filters information flowing from external sources
  to the end user
• high level text categorization using a
  high-level taxonomy
• personalized text categorization considering the
  user's needs and preferences

15
High LevelText Categorization

• A fragment of the adopted taxonomy()

• economia, affari e finanza
• agricoltura
• industria ed energia
• computer e information technology
• affari e servizi finanziari
• macroeconomia
• borsa e mercati finanziari
• trasporti
• politica
• difesa
• elezioni e referendum
• governo
• parlamento
• partiti e movimenti
• costituzione
• politica interna

() a subset of the one proposed by the
International Press Telecommunications Council
16
High LevelText Categorization

• Classifiers devised to perform text
  categorization have been implemented by using kNN
  and w-kNN algorithms
• These techniques do not require specific training
  and are very robust wrt the impact of noisy data

17
High LevelText Categorization

• A suitable encoding has been adopted
• all non informative words are removed using a
  stop word list
• a standard stemming algorithm removes the most
  common morphological and inflexional suffixes
• for each class of the taxonomy, features
  selection, based on the information gain
  statistics, has been enforced

18
PersonalizedText Categorization

• A set of arguments of interest for the user can
  be obtained by composing generic topics with
  suitable logical operators (i.e., and, or, not)
• For instance, a user might be interested in being
  kept informed about all articles that involve
  both defense and government

19
PersonalizedText Categorization

• Two ways of combining classifiers
• Horizontal composition (combination)
• Vertical composition (pipeline)

20
PersonalizedText Categorization

• The combination is evaluated using P-norms

21
PersonalizedText Categorization

• The composition exploits a pipeline of
  classifiers

C1
C2
Cn
22
PersonalizedText Categorization

• Particular care has been taken in limiting the
  phenomenon of false negatives
• The impact of false positives is reduced at the
  composition level exponentially depending on the
  number of the involved agents

23
Users Feedback

• The user's feedback module is aimed at dealing
  with any feedback provided by the user
• Two solutions have been implemented
• training an ANN
• using a kNN classifier

24
The Proposed MAS

• The Concrete Architecture

25
The PACMAS Architecture

• A multiagent architecture designed to support the
  development of applications aimed at
• Retrieving heterogeneous data spread among
  different sources
• Filtering and organizing them to personal
  interests explicitly stated by each user
• Providing adaptation techniques to improve and
  refine user profile

26
The PACMAS Architecture
Information Sources
Mid-span Levels
27
Implementation
28
Information Level

• Information agents are aimed at performing
  information extraction
• a set of agents wraps italian online newspapers
  containing articles in RSS and HTML format
• an agent wraps the adopted generic taxonomy

29
Filter Level

• Filter agents are aimed at performing text
  categorization
• removing all non-informative words
• removing the most common morphological and
  inflexional suffixes
• weighting terms using the TF-IDF
• selecting the relevant features
• generating for each document a feature vector

30
Task Level

• Task agents are aimed at performing text
  categorization.
• Agents perform
• High level text categorization
• embody a kNN classifier
• measure the classification accuracy
• Personalized text categorization
• take into account users preferences automatically
  composing topics

31
Interface Level

• Interface agents are aimed at interacting with
  the user and performing users' feedback

32
Experimental Results
33
Experimental Results

• Task Agents have been trained by a set of
  newspaper articles previously classified by
  experts of the domain
• For each item of the taxonomy
• a set of 200 documents has been selected
• KNN algorithm (with k 7) has been adopted
• 150 features have been used
• random datasets have been generated

34
Experimental Results

• For each item of the taxonomy, several tests have
  been conducted
• a set of 500 documents has been selected
• KNN algorithm (with k 7) has been adopted
• random datasets have been generated
• The average accuracy of the system is 80.05

35
Experimental Results
36
Conclusions andFuture Work
37
Conclusions

• We presented a system aimed at
• retrieving articles from italian online
  newspapers
• classifying them using suitable machine learning
  techniques
• The system has been built upon PACMAS, a support
  for implementing Personalized, Adaptive, and
  Cooperative MultiAgent Systems

38
Future Work

• A new release of the system will be improved by
  adopting different classifier algorithms
• We are investigating how to enhance the
  feedback-related functionalities according to an
  evolutionary computation framework

39
Thats all folks!