L3S Overview Visit in Sweden

1
Approach for Personalization in Open Environment
Usig Metadata and Collaborative Filtering
Techniques
2
Personalization in Open Environment

• Standards based
• Standard metadata to describe learning materials
  and user profiles (LOM, PAPI)
• Common ontology to classify content
• Distributed content
• Distributed metadata
• Distributed navigation information
• Adaptive functionalities implemented as queries
  that take into account the user profile

3
Recommender Systems

• Content-based recommendations
• Metadata about resources is
• Created (overload)
• Extracted
• Metadata is matched against the user profile to
  make personalized recommendations.
• Needed
• Mapping between different metadata schema
• Matching user profile with learning objects
  descriptions
• Our approach until now
• Based on metadata
• Unit level personalization
• Need rich metadata set (Ex dcsubject)
• Common ontology for modeling knowledge domain and
  for user knowledge

4
Recommender Systems

• Collaborative filtering
• User groups consisting of users that have similar
  profiles. Recommendations for users that belong
  to a group are based on results from other
  members of the group
• Based on implicit/explicit ratings from other
  users
• Drawbacks
• When there are constantly new resources added to
  the system, because they have never been rated by
  any other user before
• High item-to-user ratio (the number of resources
  is very big related to the number of users)
• Explicit ratings from other users are not easily
  obtained since the user has no direct reward
• Hybrid systems

5
Additional Techniques

• Reputation-based retrieval
• Different weights for ratings (the opinion of an
  expert is more important)
• Best-seller retrieval
• Based on the most popular resources from a given
  community (best seller lists).
• http//www.amazon.com
• Citation-based retrieval
• Explicit references to other resources are used
  (bibliography entries, links, etc)
• http//citeseer.nj.nec.com/cs
• http//www.google.com/

6
Different Scenarios Different Needs

• Q Which is the optimal metadata set?
• Overload for creating metadata
• Costs
• Does it worth?
• User privacy
• A Efficient personalization is possible with
  respect to the context (the learning service that
  is offered)
• Lecture (more close to reality)
• Main topic, some secondary topics from a common
  otology
• User assessments in the ontext of the lecture
• No possible interconnection of parts from several
  other lectures (alternative pages explicitly
  suggested by teachers, etc)
• ULI context
• For students, that want a qualification, for a
  specific subject, qualification that is
  recognized by their university
• Metadata set does not so rich
• Virtual University
• Personalization at unit level, or smaller
• Learning paths can span around multiple lectures,
  etc
• Reach metadata set to make user assessments
• Recommender system

7
User Modeling

• Behaviour-based model (binary model representing
  what users find interesting and uninteresting,
  machine-learning techniques)
• Knowledge-based model (domain knowledge is used
  when representing user profiles)
• User Modelling in Dialog Systems Potentials and
  Hazards, A. Kobsa

8
Experiments with the Quickstep recommender system

• Capturing knowledge of user preferences
  ontologies in recommender systems S.E.
  Middleton et al
• Multi-class behavioral model (based on paper
  topics)
• Vector terms for paper representation, term
  frequency number of terms
• Inductive learning techniques (nearest neighbour
  technique) to classify papers
• 4 level is-a ontology for computer science
• 2 trials (14, 24 reaserchers and professors)to
  compare a group of users using an ontology-based
  labeling strategy with a group of users using a
  flat labeling strategy
• 103 research papers for the first trial, covering
  17 topics and 135 papers for the second trial,
  covering 23 topics
• Conclusions
• Ontology users were 7-15 happier overall than
  the other users
• The is-a hierarchy produces a more complete
  profile, by including super class topics when a
  specific topic is browsed by the user

9
Issues to investigate

• Use ULI context (http//www.uli-campus.de) for
  personalization
• How collaborative filtering techniques can be
  used for making recommendations in this context?
• Create expressive user profiles and stil
  maintaing user privacy.
• Which would be the optimal metadata set for
  personalization in this context?

10
Approach

• Formulate approach for personalization based
  while investigating issues
• (use the MINERVA system for implementing and
  testing)