Dealing with heterogeneity in profiles for Personalized Information Retrieval

1
Dealing with heterogeneity in profiles for
Personalized Information Retrieval

• Pavel Serdyukov
• Twente University

Dealing with heterogeneity in profiles for
Personalized Information Retrieval
Pavel Serdyukov
2
Outline

• Personalization basics
• The need for dynamic preferences
• Context-aware user profiles
• Methods
• Summary

3
Personalization paradigm
Faceless user
Query
A
B
C
4
Profiles

• Personality is expressed in profile
• Contains the order in which specific features are
  liked or disliked
• Implicitly built
• Most likely quantitative (e.g. user language
  model)
• Explicitly built
• Most likely qualitative I like A better than B
  
• POS/(Music Style, Techno Brit-pop)
• POS/(Country of Origin, USA France)

Music Style ltBrit-pop 0.13 Techno 0.21
gt Country of Origin ltFrance 0.1, USA 0.4 gt
W. Kießling, VLDB 2002
5
Profile usage

• Content-based
• E.g. using cross-entropy of retrieved Object and
  Profile Language Models
• Collaborative
• Use preferences from similar profiles

last.fm
6
Heterogeneity in profiles

• User preferences are not necessarily static
• For most domains
• Multimedia search music, movies, TV
• Text search
• Product search (u-commerce)
• Preferences should be situational!
• and hence context-aware

7
Example Music preferences

• The number of situations of multimedia search and
  consumption is increased over last decade
• 50 of all personal activities have soundtrack!
• Situational context plays great role in music
  preferences (A. North, Music Perception, 2004)

8
Evolution of context-awareness

• Low-level context
• Spatial location, proximity, speed, body
  position
• Temporal Daytime, Weekday
• Physical weather, temperature, light, humidity,
  noise
• Personal heart beat, blood pressure

• High-level context
• Activity, social intercourse, mood

hands washing
web surfing
whiteboard drawing
time
9
Activity recognition

• using only location, time and duration
• using cameras, microphones
• devices set on
• RFIDs of objects involved!

RFID tags
RFID reader
D. Patterson, D. Kautz, M. Philipose.
Fine-Grained Activity Recognition by Aggregating
Abstract Object Usage. 2005
10
Social awareness

• Vicinity of people is important context
• When I am with girls I prefer jazz
• Through location recognition, or
• mobile interconnections
• Nokia Sensor, 10 meter awareness by means of
  Bluetooth technology

11
Context-aware profiles (1)

• The goal is to find context-aware language model
  of user preferences
• Data is a user context history, consisting of
  pairs
• Context and Object are vectors of attributes
• Clustering is the principal approach
• Hard clustering using object similarities
• Soft clustering using similarities of pairs

12
Context-aware profiles (1)

• Hard clustering based algorithm
• Cluster objects in K clusters using some objects
  similarity function
• Use context variables to describe clusters
• Classify new situation characterized by Context
  to do non-discriminative classification and get
• Get new language model

13
Context-aware profiles (2)

• Soft clustering based algorithm
• Probabilistic Latent Semantic Analysis principle
• The choice of objects is driven by latent
  intentions
• Likelihood of data

14
EM algorithm
15
Important applications

• In mobile multimedia
• Cell phones already contain MP3 players and
  persistent Internet connection
• In desktop search
• Search for saved/browsed documents
• Music recommendation again
• Context (metadata) is
• Time, Location (in case of laptop or pda)
• Running applications
• Opened documents (emails, web-pages, etc.)
• Played MP3 files
• User status in messengers
• Agenda records

16
Context-enriched dataset

• Context-aware IR is desperate for publicly
  available dataset!
• Context acquisition and aggregation
  infrastructures are imperfect
• Additional user interaction at this stage
• Initiative fully belongs to the user
• Initiative partly belongs to the user
• Initiative fully belongs to the system

17
Summary

• Personalization must be situational
• Context cannot be ignored then
• Unsupervised learning of context-aware
  preferences is to be utilized
• Semi-supervised methods are to be studied
• User feedback
• Explicit preferences
• Dataset is the primary stumbling block and our
  short-term goal

18
Questions
?