Achieving Private SVD-based Recommendations on Inconsistently Masked Data

1
Achieving Private SVD-based Recommendations on
Inconsistently Masked Data

• Ibrahim Yakut and Huseyin Polat
• iyakut,polath_at_anadolu.edu.tr
• Department of Computer Engineering
• Anadolu University, Turkey

2
Collaborative Filtering (CF)
Problem Information Overload
Solution Collaborative Filtering (CF)
3
Collaborative Filtering (CF)

• Recent technique for filtering and recommendation
• Relatively new concept very popular
• Used to cope with information overload
• Widely used technique by online vendors
• Many important applications in
• E-commerce
• Search engines
• Direct recommendations (books, movies, CDs, news,
  etc.)

4
Collaborative Filtering Process
Item for which prediction is sought
i1
i2
iq
im
u1
u2
Prediction
ua
Active user
un
Paq Prediction on item q for active user
5
SVD-based CF

• Singular value decomposition (SVD) based schemes
  are proposed

6
SVD-based CF
m x n
Stored and used as a model
7
Motivation

• Privacy-preserving CF schemes proposed so far are
  based on consistently masked data
• However, privacy concerns differs from user to
  user
• Users might decide to mask their data differently
  to achieve required privacy levels
• Can we still achieve CF services on
  inconsistently masked data?

8
Proposed Scheme

• Use randomized perturbation techniques (RPT) to
  inconsistently mask users data.
• Employ users mean in normalizing ratings through
  z-scores

q
u
9
Randomized Perturbation Techniques (RPT)
Collaborative Filtering
Central Database
Rn-1
Rn
R1
R2
Usern-1
Usern
User1
User2
10
Data Disguising Ways

• Users are divided into two groups
• Having privacy concerns
• Having no concerns to divulge private data
• Considering users who have privacy concerns, they
  decide how and how much data they disguise
• Uniform or Gaussian perturbing data
• Different s over (0, ?) per user
• Disguising how many cell

11
Inconsistently Masked User-Item Matrix
12
Estimation From Inconsistently Masked Data (A)

• SVD-based CF is based on scalar product and sum
• We should show how to estimate SVD from
  differently perturbed data
• To get rid of the contribution of random numbers
  in diagonal entries of ATA, we need
• Average s of random numbers
• Number of disguised cells

13
Estimation From Inconsistently Masked Data
A not disguised, B disguised by V, B' B V
14
Estimation From Inconsistently Masked Data
A masked, uniform B masked, Gaussian
Sum (A') and Sum (B') can be similarly estimated
15
Estimation From Inconsistently Masked Data
A B masked, different numbers of cells
ds of commonly filled cells
16
Experiments

• Data Sets
• Jester is a web-based joke data
• 17,988 users, 100 jokes
• Ratings over a range (-10,10),continuos
• 50 of all ratings are present
• MovieLens Public (MLP)
• 943 users, 1,682 movies
• Ratings range from 1 to 5, discrete
• 100,000 ratings, totally, present

17
Results
Percentage of disguising users
18
Results
Changing level of perturbation
19
Results
Percentage of gaussian disguised
20
Results
Percentage of disguised cells
21
Conclusion

• We showed that how to achieve CF tasks using
  SVD-based algorithms on inconsistently masked
  data
• Future work
• How to extend our schemes to other CF algorithms
• How to increase accuracy when aggregate
  information disclosed

22
THANKS FOR YOUR INTEREST...QUESTIONS?