Top-N Recommendation for Shared Accounts

Koen Verstrepen, Bart Goethals

University of Antwerp, Belgium

RecSys'15

Background

  • Multiple users often share a single account.
  • Often no contextual information is available to distinguish different users in a single account.
  • It yields 3 problems.
  • Currently no prior research on this topic.

Prob1: Dominance Problem

  • All recommendations are relevant to only some of the users that share the account and at least one user does not get any relevant recommendation.
  •  We say that these few users dominate the account.

Prob2: Generality Problem

  • The recommendations are only a little bit relevant to all users in the shared account, but are not really appealing to any of them.

Prob3: Presentation Problem

  • Since all users share a single account, how does every user know which recommendation is meant for her?

Problem Definition

               : The set of users that share acct a.
               : The acct that user u belongs to. (exactly 1)
                      : user-rating-matrix

           : RecSys on T, item i is called relevant to user u if i is in the top-N recommendations for u 

Problem Definition

  • Unfortunately, T is unknown.
  • We are only given:                                                                                 : acct-rating-matrix
               : RecSys for shared acct on R, which computes top       recommendations for account a
N_a
NaN_a

Problem Definition

  • Goal:                                                                           
    (1) Maximize top                     for every user in the acct a
               Dominance problem + Generality problem
    (2) Make it clear for a user in the user set of a which items in the top       are meant for him or her.
               Presentation problem
N_a
NaN_a

The Reference RecSys

  • An Item-based CF by Deshpande et al.
  •                                                                                            


    where

The Reference RecSys

  • Choice for                 : cosine similarity
  • Normalized cosine similarity by Deshpande et al.
  •                                                                                            
     
sim(j,i)
sim(j,i)sim(j,i)

Problem Illustration

few large contributions

many small contributions

Problem Illustration

  • Yields Generality Problem:
    It does not discriminate between a recommendation score that is the sum of a few large contributions and a recommendation score that is the sum of many small contributions.

Problem Illustration

2:3

Problem Illustration

  • Yields Dominance Problem:
    It does not take into account that        has more known preferences than       .
  • Both of the 2 examples yields Presentation problem.
u_a
uau_a
u_b
ubu_b

Solve Generality Problem

  • length adjusted item-based recommendation score:
  •                                                                                            
     
  • By choosing p > 0, we create a bias in favor of the sum of a few large similarities. (p↑ => bias↑)
  • Similar to the concept of normalization.

Solve Generality Problem

  • Formally, we want to rank all items i according to:            
     
  • Since we only have I(a), we approximate it with an upper bound:
     
  • This is because every set of items I(u) for which u∈U(a) is also an element of the power set of I(a), which is        .
  • The subset of I(a) which corresponds to users are more likely to result into the highest recommendation scores than the subset consists of random items.
2^{I(a)}
2I(a)2^{I(a)}

Solve Generality Problem

  • The disambiguating item-based (DAMIB) recommender system:


     
  • When                    is high, we expect that      corresponds well to an individual user.
  • When                    is low, there is no user in the shared account for whom i is a strong recommendation and we expect       to be a random subset.
S_i^*
Si∗S_i^*
S_i^*
Si∗S_i^*

Efficient Computation

  •      can be computed in                  time instead of exponential time since:                                                                         
S_i^*
Si∗S_i^*
O(nlogn)
O(nlogn)O(nlogn)

Efficient Computation

  • Proof:                                                                         

Solve Dominance Problem

  • Use the DAMIB scores to find the       highest scoring candidate recommendations.
  • Avoid recommending the items whose explanation         is not sufficiently different from the explanations of the higher ranked recommendations.
  • The difference measuring metric:
S_i^*
Si∗S_i^*
N_a
NaN_a
|S_i^*\ \backslash\ C(a)|\geq1
∣Si∗ \ C(a)∣≥1|S_i^*\ \backslash\ C(a)|\geq1

All highly relevant recommendations to individual users have already been recommended => Clean the set C(a)

Still need to recommend the items which fail to meet the red conditions since Na recommendations are required.

|S_i^*\ \backslash\ C(a)|\geq1
∣Si∗ \ C(a)∣≥1|S_i^*\ \backslash\ C(a)|\geq1

Solve Presentation Problem

  • Present every recommendation                             together with its explanation      .
  • e.g.
    Recommend item r with its similar items s1, s2 and s3. A user will recognize s1, s2 and s3 as her preferences, and know that r is recommended to her.
i \in top-N_a
i∈top−Nai \in top-N_a
S_i^*
Si∗S_i^*

Experiments

  • Source code available on https://bitbucket.org/BlindReview/rsa (No permission)
  • Datasets with shared account info are rarely seen.
  • However, we can create synthetic shared accounts by randomly grouping users in groups of 2, 3 or 4.               
  • Datasets:
    (1) Yahoo!Music
    (2) Movielens1M
    (3) Book-Crossing
    (4) Wiki10+

Performance Evaluation

  • Performance evaluation metrics:
    (1) Recall of user u:


    (2) The fraction of users that does not get any relevant recs:
     

where

Performance Evaluation

  • Competitors:
    (1) IB: The original item-based Rec Sys (Baseline)
    (2) IB-COVER: COVER(a,IB)
    (3) DAMIB-COVER with p = 0.5
    (4) DAMIB-COVER with p = 0.75
rec(562)=\frac{3}{5}
rec(562)=35rec(562)=\frac{3}{5}
rec(562)=\frac{0}{5}=0
rec(562)=05=0rec(562)=\frac{0}{5}=0

Q: DAMIB-COVER has a bad performance when accounts are not shared?

  • The author claims that DAMIB-COVER algorithm still performs well when no accounts are shared.
  • By Deshpande et al.
    Randomly choose 1 preference of every user to be the test preference       for that user.
  • Performance evaluation metric: Hit rate at 5
h_u
huh_u

Performance Evaluation

POP: The baseline-algorithm which is a non-personalized algorithm that ranks all items according to their popularity, i.e. the number of users in the training set that prefer the item.

  • Performance evaluation metric for presentation problem:

     
  • The DAMIB-COVER explanations are superior to the
    IB-COVER explanations.

Performance Evaluation

Related Work

  • [TKDE'08][Using context to improve predictive modeling of customers in personalization applications][C. Palmisano, A. Tuzhilin, and M. Gorgolione]
  • [WebMine'06][Contextual recommendation][S. Anand and B. Mobasher]
  • [UAI'12][Guess who rated this movie: Identifying users through subspace clustering][A. Zhang, N. Fawaz, S. Ioannidis, and A. Montanari]
    (Synthetic Account Evaluation Approach)
  • [ICDE'09][Recommendation diversification using explanations][C. Yu, L. Lakshmanan, and S. Amer-Yahia]
    (Diversity Enhancement, Solution similar to IB-COVER)

Conclusions and Future Work

  • The widely used item-based Rec Sys fails when it makes recommendations for shared accounts.
  • The DAMIB-COVER algorithm overcome the challenge of recommendation for shared accounts in the absence of contextual information. 
  • The computation of a recommendation score cost 
                   instead of exponential time.
  • Planning to generalize the proposed solution to a wider range of Rec Sys.
O(nlogn)
O(nlogn)O(nlogn)
Made with Slides.com