$30 off During Our Annual Pro Sale. View Details »

Incremental Factorization Machines

Takuya Kitazawa
September 15, 2016

Incremental Factorization Machines

Presented at Workshop on Profiling User Preferences for Dynamic Online and Real-Time Recommendations (RecProfile in conjunction with RecSys 2016)

arXiv paper: https://arxiv.org/abs/1607.02858

Takuya Kitazawa

September 15, 2016
Tweet

More Decks by Takuya Kitazawa

Other Decks in Research

Transcript

  1. Incremental Factorization Machines for
    Persistently Cold-starting
    Online Item Recommendation
    Takuya Kitazawa
    Graduate School of Information Science and Technology, The University of Tokyo
    RecProfile 2016

    View Slide

  2. Incremental Factorization Machines for
    (1) Persistently Cold-starting,
    (2) Online Item Recommendation
    as new
    baseline method
    Factorization Machines
    context-aware
    recommender
    Incremental MF
    online item recommender
    optimized by SGD

    View Slide

  3. Continuously recommend ”new” items to ”unforeseen” users
    Poor recommendation accuracy !
    Yahoo! Labs — online ad [M. Aharon et al. RecSys 2013]
    Booking.com — hotel reservation [L. Bernardi et al. CBRecSys 2015]

    Rakuten — golf package reservation [R. Swezey and Y. Chung. CIKM 2015]
    3
    Challenge I: Persistent cold-start in e-commerce
    users’ rare activity changing persona
    short-lived items changing prices

    View Slide

  4. 4
    Promising solution:
    Context-aware (feature-based) recommendation
    Occupation Browser Age Sex
    [ 0, 0, 1, 0, 0, 0, 0, 1, 22, 1 ]
    [ 0, 1, 1, 0, 500 ]
    Category Price
    Enriched inputs

    View Slide

  5. For
    5
    Factorization Machines (FMs; general predictor) for 

    context-aware recommendation
    global bias
    linear
    interaction
    [S. Rendle. ACM TIST, 3(3), 2012.] (Fig. 1)
    MF
    FMs

    View Slide

  6. For
    6
    Factorization Machines (FMs; general predictor) for 

    context-aware recommendation
    global bias
    linear
    interaction
    [S. Rendle. ACM TIST, 3(3), 2012.] (Fig. 1)
    MF
    FMs
    Persistent cold-start can be seen as:

    concept drift
    online algorithms are more effective

    View Slide

  7. 7
    Challenge II: Online item recommendation
    Incrementally update model based on user-item interactions
    (e.g. click, buy, book; not explicit rating)

    View Slide

  8. 8
    Classical static baseline:
    Matrix Factorization (MF) for item recommendation
    https://en.wikipedia.org/wiki/Collaborative_filtering
    R ≈ P
    users
    QT
    items
    ×
    1 1 1
    1
    1
    1
    1 1
    1
    0
    0 0
    0
    0 0
    0
    0
    0
    0
    0
    1.0 least-squared loss for binary output

    View Slide

  9. 9
    Incremental Matrix Factorization (iMF) for 

    online item recommendation
    iMF
    S can be data stream
    update (user ID)
    update (item ID)
    for each observation
    Where are auxiliary features?
    [J. Vinagre et al. UMAP 2014]
    SGD update

    View Slide

  10. 10
    Overview
    FMs (context-aware
    recommender)
    iMF (online item
    recommender)
    Incremental FMs (iFMs)
    (context-aware online
    item recommender)
    optimized by SGD
    as generic framework
    (i.e. baseline)
    classical online baseline
    modern context-aware baseline

    View Slide

  11. 11
    Using FMs to item recommendation
    Solve MF with unique target value

    [J. Vinagre et al. In Proc. of UMAP 2014, pp. 459–470.]
    Solve FM as regression (least-square loss) with y = 1.0
    1.0
    For
    1.0

    View Slide

  12. 12
    Generalizing iMF to incremental FMs (iFMs)
    SGD update for FMs
    O( #non-zero in x × k )
    S is data stream
    for each observation
    SGD update for MF
    update (user information)
    update (item information)
    update
    update
    update
    SGD update
    SGD update

    View Slide

  13. 13
    Incremental adaptive regularization
    FMs have many hyperparameters:
    Adaptive regularization [S. Rendle. In Proc. of WSDM 2012, pp. 133-142.]
    Incremental adaptive regularization
    (“pseudo validation sample” idea)
    outdated
    immediately

    View Slide

  14. 14
    Evaluation of online item recommender:
    Test-then-learn procedure
    0) pre-train with initial 20% of samples
    [J. Vinagre, et al. In Proc. of REDD 2014.] 

    Evaluation of Recommender Systems in Streaming Environments,
    (u1
    , i1
    ) ɾɾɾ
    timestamped n-samples for evaluation
    ɾɾɾ (uk
    , ik
    ) (uk+3
    , ik+3
    )
    For sample (u, i)
    1) top-N recommendation for u
    2) check if i is in top-N list
    3) compute recall@N in
    4) update model based on (u, i)
    (un
    , in
    )

    View Slide

  15. 15
    Experiment: MovieLens 100k (1/2)
    Real-world movie rating dataset
    ■ Binarized: { 1, 2, 3, 4, 5 } => { 0, 1 }
    ■ Users’ rare activity

    View Slide

  16. 16
    Experiment: MovieLens 100k (2/2)
    MF < FMs & static < incremental

    View Slide

  17. 17
    Experiment: Synthetic click (1/2)
    5 ad variants w/ categories: { ad1, ad2 }, { ad3, ad4 }, { ad5 }

    ■ Rule-based generator from [M. Aharon et al., RecSys2013]



    ■ Item instability
    500k impressions ▶ rule change ▶ 500k impressions

    View Slide

  18. 18
    Experiment: Synthetic click (2/2)
    iFMs show higher adaptivity for changes

    View Slide

  19. 19
    Overall results (5 times w/ different initial params.)
    Trade-off : efficiency vs context-awareness
    FMs are: slower better

    View Slide

  20. 20
    Summary
    ■ Key features: 

    positive-only feedback

    incremental adaptive regularization
    ■ Use your appropriate competitor:
    FMs (context-aware
    recommender)
    iMF (online item
    recommender)
    Incremental FMs (iFMs)
    (context-aware online
    item recommender)
    ID-only Feature-based
    Static MF FMs
    Online iMF iFMs

    View Slide

  21. 21
    Future work
    ▶ More competitors/datasets (time-stamped, rich in contexts)
    ▶ Other approaches for positive-only feedback (e.g. BPR)
    [J. Vinagre, et al. In Proc. of UMAP 2014.]
    (Fig. 1)

    View Slide

  22. Incremental Factorization Machines for
    Persistently Cold-starting
    Online Item Recommendation
    Takuya Kitazawa
    Graduate School of Information Science and Technology, The University of Tokyo
    RecProfile 2016

    View Slide

  23. 25
    500k impressions ▶ rule change ▶ 500k impressions
    3
    [M. Aharon et al., 

    RecSys2013] (Table 1)
    ⾢ most clicked ad rule

    View Slide

  24. 26
    New feature insertion
    e.g. insert new user’s dimension

    View Slide