bayesian adaptive user profiling with explicit & implicit feedback (slides)

Upload: philip-zigoris

Post on 31-May-2018

232 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    1/30

    Bayesian Adaptive User

    Profiling with Explicit & Implicit

    Feedback

    Philip Zigoris, Yi ZhangUniversity of California, Santa Cruz

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    2/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 2

    Obstacles to Personalization

    1. Asking for feedback is intrusive.

    Implicit Feedback - Infer user feedback

    from user behavior.

    Little to no information about new

    users (the cold-start problem).Borrow information from existing users

    via Bayesian hierarchical model.

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    3/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 3

    Outline

    Implicit Feedback

    Hierarchical Bayesian Framework

    Gaussian Network

    Experiments

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    4/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 4

    Implicit Feedback

    A users interaction with a document suggests

    their opinion of it.

    Includes: Keyboard usage Mouse usage Viewing time Eye-tracking

    Appeal: cheap! **Modulo security issues

    Question: Does it have any predictive value?

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    5/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 5

    Is Implicit Feedback Useful?

    Yes Positive correlation between the time spent

    viewing a page and a users opinion of thepage. [Claypool et al., 2001; Fox et al.,2005]

    No (sometimes)

    In Kelly et al., 2004 demonstrates thatcorrelation varies significantly across tasks.

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    6/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 6

    Outline

    Implicit Feedback

    Hierarchical Bayesian Framework

    Gaussian Network

    Experiments

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    7/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 7

    f

    I Like!

    X1 Y1

    X2 Y2

    X3 Y3

    The Task at Hand

    Doc Rating

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    8/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 8

    Cold-Start Problem

    Personalized systems require training

    data.

    Users do not want to endure poorperformance while the system is

    learning.

    Solution: Give the system a head-start with some (Bayesian) prior

    knowledge.

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    9/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 9

    Bayesian Prior over User Profiles

    User-ModelSpace

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    10/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 10

    fu ~ P(f |q)

    y ~ fu(x)

    Parameter describing

    prior distribution

    Hierarchical User ModelGeneric Form

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    11/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 11

    P( fu |q)P( fu |q,(x1,y1))P( fu |q,(x1,y1),(x2,y2 ))P( fu |q,(x1,y1),K ,(x3,y3 ))P( fu |q,(x1,y1),K ,(x4,y4 ))P( fu |q,(x1,y1),K ,(x5,y5 ))

    Refining Beliefs Based on User Data

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    12/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 12

    fMAPu

    = argmaxf

    P(f |q)P(Du

    | f)[ ]

    = argmaf

    x log(P(f |q)) + log(P(Du

    | f))[ ]

    The Posterior Distribution

    P( fu |q,Du) =P(D

    u| f

    u,q)P( f

    u|q)

    P(Du

    |q)

    Influence of

    prior

    Data

    likelihood

    Maximum A Posteriorimodel:

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    13/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 13

    Tradeoff: Prior vs. Data

    More data data likelihood term will

    dominate the objective function.

    log(P(Du | f)) = log P(f(x i) = y i)( )

    = log(P(f(x i) = y i))

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    14/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 14

    Outline

    Implicit Feedback

    Hierarchical Bayesian Framework

    Gaussian Network

    Experiments

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    15/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 15

    Gaussian Network

    wu ~ N(m,S)

    y ~ N(xT wu,ku2)

    wMAPu = argmin

    w

    (w - m)TS- 1(w - m)+ 1ku

    (x iT w - y i)

    2

    i

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    16/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 16

    Learning the Prior from

    Existing Users

    Unbiased estimator for prior parameters:

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    17/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 17

    Outline

    Implicit Feedback

    Hierarchical Bayesian Framework

    Gaussian Network

    Experiments

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    18/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 18

    Experimental Datasets

    Claypool

    75 Student Volunteers, 1823 DocumentsUnguided, unrestricted web browsing for 20-30 minutes

    Every document explicitly rated (Scale 1 to 5)

    Zhang

    15 Users, 4663 Documents

    1 month, users spent >1 hour everyday using system Focused on reading news articlesNot every document is explicitly rated (Scale 1 to 5)Includes other forms of explicit feedback (TBA)

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    19/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 19

    Feature Sets

    Document length

    Speed of host server

    Number of pages linking to host server

    Document

    Mouse usage

    Keyboard usage

    Time spent on page

    Implicit

    Relevance Score

    Readability ScoreExplicit(Zhang)

    ExamplesType

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    20/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 20

    Evaluation Methodology

    Tested four models

    Linear model with (P)rior

    Linear model with (N)o prior

    (G)eneric, user-independent model

    Moving (A)verage

    Original ordering preserved

    Leave-one-user-out Model updated and evaluated (error2) after

    every example

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    21/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 21

    Summary of Results

    Performance (MSE) averaged over time and user.

    Hierarchical model significantly*

    outperformsother methods.

    * (95% Wilcoxon signed rank test)

    Explicit

    ImplicitImplicit &

    Explicit

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    22/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 22

    Performance Over TimeExplicit & Implicit Feedback

    Shifting rating bias

    Personalization seems

    to begin

    Prior getting

    in the way

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    23/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 23

    Performance Over TimeImplicit Feedback Only (Zhang et al)

    Follows moving

    average

    Implicit feedback

    becomes useful

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    24/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 24

    Performance Over TimeImplicit Feedback Only (Claypool et al)

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    25/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 25

    Comparative Performance

    with Different Feature Sets

    Implicit feedback

    hurts performance

    Implicit feedbackbenefits performance

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    26/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 26

    Discussion: The Dynamics of the

    User Model

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    27/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 27

    Discussion: Is Implicit Feedback

    Useful? Why?

    Error can be decomposed into:

    Bias - How closely the learning algorithm can approximate the bestsolution.

    Variance - Sensitivity of learning algorithm to the training sample.

    Noise - Irreducible uncertainty of problem

    Including implicit feedback increases variance. This explains why:Negative impact on early performanceRequires substantial amount of training data to be useful.

    Linear model has strong bias. This may explain why:Implicit feedback does not perform well on its own.

    Di i Shif i U

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    28/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 28

    Discussion: Shifting User

    Behavior

    Implicit assumption of our model:

    User behavior is consistent over time.

    Already observed shifting bias in rating.

    Also, observed shifts in keyboard andmouse usage.

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    29/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 29

    Conclusion

    Implicit feedback (of the type used in our study) has

    marginal predictive value.

    A question ofwhen, not if

    Requires personalization

    Requires lots of data

    Hierarchical model effectively alleviates cold-

    start problem

    Shifting user behavior exists

  • 8/14/2019 Bayesian Adaptive User Profiling with Explicit & Implicit Feedback (slides)

    30/30

    11/8/2006 Bayesian Profiling with Explicit & Implicit Feedback 30

    Thank You!