CYCLADES The Personalization Service

1
CYCLADESThe Personalization Service

• Fabrizio Sebastiani
• fabrizio_at_iei.pi.cnr.it
• IEI-CNR (Italy)

2
Main goal

• The Personalization Service (PS) is responsible
  for allowing a user to interact with the system
  in a flexible and highly personalized way.

3
Personalization as a topic-driven activity

• Personalization is viewed as a content-based
  notion i.e. the interaction with the user must
  take into account the users interests.
• The users interests are viewed as a
  hierarchically structured set of topics, each of
  which represents the users subjective view of a
  discipline which is of interest to her.
• Personalization is achieved by
• Creating and maintaining a set of topic profiles
  (i.e. internal representations of topics)
• Letting the information seeking/presenting
  behaviour of the system be influenced by these
  profiles.

4
User topics and community topics

• Topics (and their profiles) may belong to
• a particular user (user-specific topics).
• a particular community (community-specific
  topics). Any user has a default subscription to
  the topics belonging to a community she is a
  member of, and may selectively cancel these
  subscriptions.
• The topics that pertain to a user are thus the
  union of
• the topics that belong to the user as an
  individual
• the topics that belong to the communities the
  user is a member of.

5
An example topic hierarchy
Information Retrieval
Fuzzy Logic
Digital Libraries
Umbertos CYCLADES folder
Possibilistic Logics
Logics for Uncertainty and Imprecision
Logics of subjective probability
Probabilistic Logics
Logics of objective probability
6
Topic-based personalization

• Personalization influences the behaviour of the
  system in
• Seeking information
• Presenting the retrieved information to the user
• both when it operates
• in ad hoc mode
• in on-demand mode

7
Personalization in the ad hoc mode

• Ad hoc (pull) mode the system delivers to the
  user information it deems relevant to a specific
  request she has just explicitly issued.
• This modality is apt to serve user information
  needs of a contingent (temporally local) nature.
• For personalization, the topic profiles of the
  user and of the communities to which the user
  belongs are to be used globally (thus forming a
  user profile) as "background information" that
  allow the request to be understood in the context
  of the users long-term interests.

8
Personalization in the on demand mode

• On demand mode the user asks the system to
  deliver to her any information that might have
  recently become available and that is relevant to
  her own interests.
• This modality is used to satisfy information
  needs of a permanent or semi-permanent
  (temporally global) nature.
• For personalization, each of these profiles is to
  be used as a standing (i.e. permanent) query.

9
Classification of results

• The PS further provides personalized automatic
  classification of the records resulting from the
  (either ad hoc or on-demand) information-providing
  activity, into a set of hierarchically organized
  folders, each corresponding to a topic. A topic
  profile is thus (the declarative part of) a
  classifier, i.e. a tool that can automatically
  decide whether an item is relevant or not to the
  topic
• Any automatic classifier is going to have a
  non-null error rate the user is thus allowed to
  correct the classifiers decisions that she
  perceives as wrong

10
How are classifiers built and updated?

• The classifier is automatically built by learning
  from items the user has classified manually.
• The classifier is automatically updated by
  interpreting implicit user feedback, consisting
  of
• moving an item from folder A to folder B
• deleting an item from folder A
• aliasing an item from folder A to folder B
• Updates reflect
• errors by the classifier
• semantic shifts in the meaning of the involved
  topics

11
Dynamic topic hierarchies

• The user (or the community owner) may decide
  herself how to structure the topic hierarchy
  initially a single-node hierarchy is the default
• The topic hierarchy is dynamic, as the user (or
  the system, after user approval) may restructure
  it by
• Splitting a topic into several subtopics
• Collapsing the subtopics into a single topic
• Deleting a topic

12
Some technical issues

• Topic profiles will be inductively built and
  updated by supervised learning techniques we
  plan to use incremental classification techniques
  (e.g. Balanced Winnow)
• The hierarchical topic structure will be
  dynamically modified by unsupervised learning
  techniques (i.e. clustering)
• The classifiers will be evaluated and optimized
  using user-dependent utility measures (e.g.
  giving different penalties to errors of
  commission and errors of omission).

13
The PS and the Mediator Service

• When the MS interacts with user X, the PS feeds
  to the MS the topic profiles (resp. user profile)
  of X that the QBS uses to issue on-demand X
  queries (resp. to personalize ad hoc X queries)
• The QBS feeds to the PS the results of these
  queries that the PS classifies in the topic
  hierarchy specific to the user.
• The MS informs the PS of user actions (such as
  moving, deleting or aliasing records across
  folders) that the PS may use for updating topic
  profiles and/or restructuring (upon user
  approval) the structure of the topic hierarchy.