Transcript PhD Direction

Adaptive User Profiling
Carolina Bailey
([email protected])
1
User Profiling: Areas
 Information Retrieval
 Personalised Search
 Personalised TV Listings
 Recommendation Systems
 Expert Finding Systems – profile matching
 Information Filtering
 Spam Filters
 News Filtering
 E-Learning
 Learner Profiles
 Intelligent Environments
 Intelligent Agents
 Behaviour Prediction
2
User Profiling
 A user profile can be used like a filter on a set of data, with
various sets of data:
 Search Engine results
 Environment variables such as lighting settings and temperature
settings
 Recipes
 News feeds… etc.
 … any collection of data items that could be personalised
 Any information available about the user can be incorporated
into the profile
 Likes, dislikes – specified or implied
 Various histories e.g. past behaviour, purchasing, browsing, TV
watching, bookmarks
 Disabilities and/or medical details
 Future data sources
3
User Profiling: Steps
Sources of Data
Decide what source to
use.
Collect Relevant Data
Go get it
Raw Data
Got it!
Processing Method/
Algorithm
Get some meaning out
of it.
Storage and
Representation
Save it in a consistent
format or structure.
User Profile
Call it a User Profile.
Application
Use it.
4
Building a User Profile
Various Data Source Applications
Examples of Data Sources
HTML file (e.g. bookmarks)
XML files
Text files (e.g. rules)
Various Methods of Processing
Various Representations of Profiles
Some Example Personalised Applications
and Data Sources…
5
Data Source Applications
Search Engine
6
Data Source Applications
Question Answering
System
7
Data Source Applications
Personalised TV
8
Data Source Applications
Intelligent Environment
9
Data Sources - XML Recipe
Ref: Recipe example from http://www.brics.dk/~amoeller/XML/xml/example.html
10
Data Sources – Fuzzy Logic Example
11
Methods of processing data
Processing
Method/
Algorithm
From raw data to
profile representation...
Link
Analysis
Categorisation
Similarity
Comparison
Perceptron
HITS
Algorithm
Text
Mining
Text
Filtering
Rocchio
Algorithm
Genetic
Algorithms
Naive
Bayesian
Classification
Generalisation
Word-Sense
Disambiguation
Thesauri
Term to
Concept
Mapping
12
Representations of Profiles
A user profile can be
represented as...
Ontology
Vector(s)
Prototype
Matrix
Keywords
Representation
Bag of
Words
Rules
Concept
Hierarchy
Linear
Model
Boolean
13
A Global, Unified Profile
 Potentially, one single profile could be used
anywhere, for any application.
 Currently, the common theme in previous
research, is that there is no common theme!
Different data storage methods, data processing
methods and algorithms, representation of profiles etc.
 What is the most efficient of these different
methods and processes?
 Can a user profile from one application be used
within another application?
14
A Global, Unified Profile
15
A Global, Unified Profile
16
Global Profile - Considerations
 Mapping and categorising items to the Global Profile
E.g. a generic term for temperature, heating, radiators
etc.
 Extensible way to add new data (and data sources) to
the profile
Textual data
Fuzzy data
Future data items and sources – e.g. SatNav
 Data storage choices
Main Server
Distributed
 Transparency of the profile
 Updating and synchronising
17
Security, Privacy and Legal Implications
 The User must be in ultimate control!
 What data should be used in a profile? Purchasing
history? Criminal record?
 Who and what should be allowed access to a profile?
The Police? The Government? Could it be used
against their wishes?
 Fine balance between what is good-intentioned
personalisation and what is a complete loss of privacy
 As people lose more and more control of what
information is stored about them, their personal
freedom may feel encroached upon, resulting in a
strong resistance to further developments towards
user profiling
18
The End
To be continued…
[email protected]
19