Ifkad 2015 presentation
-
Upload
corrado-mencar -
Category
Internet
-
view
140 -
download
0
Transcript of Ifkad 2015 presentation
Fuzzy Information Filters for User Modeling in Collective Intelligence SystemsG. Castellano, C. Castiello, A.M. Fanelli, M. Lucarelli, C. MencarDept. of Informatics, University of Bari, Italy
Research outline↠ Purpose Define an abstract model for
representing users and resources ↠ Approach Fuzzy Information Filters (FIF).↠ Values Generality, adaptivity, handling imprecise
information, explainability↠ Impact personalized e-learning systems,
recommendation systems, community discovery, etc.
Collective Intelligence↠ Intelligence emerging from the interaction of
many individuals→ collaboration, competition, opinions, messaging, …
↠ Personalized experience in web applications↠ Information filtering
→ Based on data→ Based on a model
Information filtering↠ Fight information overload
→ the difficulty a person can have in making decisions caused by too much information. (Wikipedia)
↠ Deliver only relevant information→ User model
User model
↠ Preference→ What a user likes
↠ Competence→ What a user needs
↠ Knowledge→ What a user knows
Graduality & Granularity↠ Preferences & co. are always expressed to a
degree→ Ranking of objects according to prefs., needs, etc.
↠ Preferences & co. are often imprecise→ Refer to classes of objects instead of single individuals
Fuzzy Set Theory↠ Mathematical model
of granularity and graduality
↠ Extends classical set theory
Horror
Thriller
Drama
Fantasy
Comedy
Humor
FIF sequential composition
FIF parallel composition
Description-based filter↠ An object is represented as a collection of
metadata↠ Each metadata is defined by an attribute and a
fuzzy set of values↠ A description-based filter is defined by an
attribute and a fuzzy set of values
Matching
object
Description-based FIF
1. Given an object o={M1, M2, … Mn}
2. Given a Description-based FIF on attribute A and fuzzy set u
3. Find metadata M=(A,v) in o4. Match fuzzy set according to
possibility measureµ=maxxϵAmin{u(x),v(x),λ}
User model as FIF structure
OWA
(Simplified diagram: not all lines are drawn)
OWA
User likes cheap, lightweight, small cars which have a low-consumption engine and 4-5 doors
Filter learning↠ Filters can be designed by hand, or↠ they could be acquired from past observations
→ sequence of objects observed by a user↠ Theory of Possibility →
Principle of Minimum SpecificityI know John is a tall man (more than about 180cm) ⊢ Now I know John is within about
180-190 cmYou tell me John is not so tall (less than about 190cm)
Learning principles↠ Temporal Locality. If I observe an object, I will observe
the same object in the near future↠ Spatial Locality. If I observe an object, I will observe a
similar object↠ Relevance of knowledge. What I know has some
importance for learning↠ Relevance of observation. What I observe has some
importance for learning
Structural learning1. Given an observed object o and a filter f, a
matching degree d is calculated2. If d > threshold, then f is updated
a. Application of minimum specificity and learning principles3. Else a new filter is added in parallel to f
a. The new filter is a sequence of description based filters corresponding to metadata of o.
Conclusive remarks↠ Representation of complex user profiles↠ Filtering endowed with granularity and
graduality↠ Self-adaption to observed objects
Future research↠ Theory
→ Refinement of learning principles and structural learning→ Extended representation of user models→ Experiments with real-world data
↠ Application→ Integration within the Openness platform→ Service-oriented software system