SLIDE 1
PRIVACY THROUGH SOLIDARITY
Asia J. Biega Rishiraj Saha Roy Gerhard Weikum
YOUR PROFILE MEANS … PERSONALIZATION
python programming, java, inheritance, nltk
SEARCH RECOMMENDATIONS … …
Asia J. Biega jbiega@mpii.de
PRIVACY THROUGH SOLIDARITY Asia J. Biega Rishiraj Saha Roy - - PowerPoint PPT Presentation
PRIVACY THROUGH SOLIDARITY Asia J. Biega Rishiraj Saha Roy - - PowerPoint PPT Presentation
PRIVACY THROUGH SOLIDARITY Asia J. Biega Rishiraj Saha Roy Gerhard Weikum YOUR PROFILE MEANS PERSONALIZATION SEARCH RECOMMENDATIONS python programming, java, inheritance, nltk Asia J. Biega jbiega@mpii.de YOUR PROFILE
SLIDE 2
SLIDE 3
YOUR PROFILE MEANS … YOU
SEARCH RECOMMENDATIONS
programming ethno jazz mediation back exercise pottery tokyo in august Python Programming Mein Kampf The Holy Bible The Modern Libertarian
Asia J. Biega jbiega@mpii.de
SLIDE 4
YOU CAN’T HAVE YOUR CAKE AND (HAVE YOUR SERVICE PROVIDER) EAT IT TOO …
OR
Asia J. Biega jbiega@mpii.de
SLIDE 5
… OR CAN YOU?
programming ethno jazz mediation back exercise pottery tokyo in august python
Asia J. Biega jbiega@mpii.de
SLIDE 6
… OR CAN YOU?
programming ethno jazz mediation back exercise pottery tokyo in august python
Asia J. Biega jbiega@mpii.de
SLIDE 7
… OR CAN YOU?
programming ethno jazz mediation back exercise pottery tokyo in august programming ethno jazz mediation back exercise pottery tokyo in august
USER UTILITY: COHERENT CHUNKS PRIVACY: PROFILE FRAGMENTATION
Asia J. Biega jbiega@mpii.de
SLIDE 8
FRAGMENTATION: PRACTICALLY
Service Provider Local user profiles ?
Asia J. Biega jbiega@mpii.de
SLIDE 9
MEDIATOR ACCOUNTS PROXY
Service Provider Mediator Accounts Item: Items Result: Results Local user profiles query rating ranking predicted ratings Store no links
Asia J. Biega jbiega@mpii.de
SLIDE 10
PRIVACY VS. UTILITY
PRIVACY profile scrambled USER UTILITY results personalized This work
Asia J. Biega jbiega@mpii.de
SLIDE 11
PRIVACY VS. UTILITY
PRIVACY profile scrambled USER UTILITY coherent context preserved SERVICE PROVIDER UTILITY minimize analytics loss U s u a l f
- c
u s
Asia J. Biega jbiega@mpii.de
SLIDE 12
PRIVACY VS. UTILITY - STRATEGIES
PRIVACY Random shuffling USER UTILITY Results intact
Asia J. Biega jbiega@mpii.de
SLIDE 13
CONTROLLING THE TRADE-OFF
P(m|o) = α · Ppriv(m|o) + (1 − α) · Putil(m|o)
RANDOM
α = 1
COHERENT
α = 0
Profiling-Tradeoff Assignment Entropy Pairwise similarity
Asia J. Biega jbiega@mpii.de
SLIDE 14
MEDIATOR ACCOUNTS: SEARCH
programming pottery tokyo in august
QUERY HISTORY
programming pottery tokyo in august
Language model-based personalization query-doc + doc-mediator
User U Mediator M Similarity: topical
Forwarding
Asia J. Biega jbiega@mpii.de
SLIDE 15
MEDIATOR ACCOUNTS: RECOMMENDERS
Umiera piękno: 4 Koreni: 5 Artpop: 1
RATING HISTORY
Umiera piękno Koreni Artpop
Collaborative filtering
User U Mediator M Similarity: categorical
Aggregation
Asia J. Biega jbiega@mpii.de
SLIDE 16
EVALUATION
~900 User profiles Profiling-Tradeoff Assignment (Queries synthesized from StackExchange) Mediator Accounts
Asia J. Biega jbiega@mpii.de
SLIDE 17
EVALUATION
Q: HOW DOES THE PRIVACY-UTILITY TRADE-OFF LOOK LIKE?
Profiling-Tradeoff Assignment (Queries synthesized from StackExchange) Mediator Accounts ~900 User profiles
Asia J. Biega jbiega@mpii.de
SLIDE 18
EVALUATION: METRICS
MODEL EMPIRICAL
PRIVACY UTILITY
Entropy
- bject-level
KL-divergence topic-level Coherence Kendall’s Tau (search) MSE (recommenders)
:
(
:
(
user u
,
mediator m For each object: Diff* Ranking over StackExchange answers
Asia J. Biega jbiega@mpii.de
SLIDE 19
EVALUATION: OBSERVATIONS
SEARCH SYSTEMS TRADE-OFFS (EMPIRICAL)
Asia J. Biega jbiega@mpii.de
SLIDE 20
EVALUATION: OBSERVATIONS
I d e a l
SEARCH SYSTEMS TRADE-OFFS (EMPIRICAL)
Asia J. Biega jbiega@mpii.de
SLIDE 21
EVALUATION: OBSERVATIONS
I d e a l
SEARCH SYSTEMS TRADE-OFFS (EMPIRICAL)
T r a d e
- f
f
Asia J. Biega jbiega@mpii.de
SLIDE 22
EVALUATION: OBSERVATIONS
SEARCH SYSTEMS TRADE-OFFS (EMPIRICAL)
Utility loss low even for random
Asia J. Biega jbiega@mpii.de
SLIDE 23
EVALUATION: OBSERVATIONS
SEARCH SYSTEMS TRADE-OFFS (EMPIRICAL)
Alpha influences the variance
Asia J. Biega jbiega@mpii.de
SLIDE 24
EVALUATION: OBSERVATIONS
PROFILE DIVERSITY SEARCH SYSTEMS
Asia J. Biega jbiega@mpii.de
SLIDE 25
EVALUATION: OBSERVATIONS
PROFILE DIVERSITY SEARCH SYSTEMS
Coherent Random
Asia J. Biega jbiega@mpii.de
SLIDE 26
EVALUATION: OBSERVATIONS
PROFILE DIVERSITY SEARCH SYSTEMS
Coherent Random high diversity = utility-safe scrambling
Asia J. Biega jbiega@mpii.de
SLIDE 27
EVALUATION: OBSERVATIONS
RECOMMENDER SYSTEMS
Similar observations hold
Asia J. Biega jbiega@mpii.de
SLIDE 28
PRIVACY THROUGH SOLIDARITY THANKS!
Asia J. Biega jbiega@mpii.de
USER UTILITY PROFILING PRIVACY MEDIATOR ACCOUNTS SEARCH RECOMMENDERS