Recommendations as Treatments: Debiasing Learning and Evaluation - PowerPoint PPT Presentation

Recommendations as Treatments: Debiasing Learning and Evaluation ICML 2016, NYC Schnabel † , Adith Swaminathan † , Tob obias Sc Ashudeep Singh † , Navin Chandak § , Thorsten Joachims † † Cornell University, § Google Funded in part through NSF Awards IIS-1247637, IIS-1217686, IIS-1513692.

Recommendations as Treatments: Debiasing Learning and Evaluation Movie recommendation Horror Romance Drama 5 5 1 3 5 1 3 5 5 Horror Lovers 5 5 1 3 5 5 5 3 O 5 5 1 3 3 Observed Y/N 1 1 5 5 3 3 1 5 3 5 5 Romance Lovers 5 5 5 5 3 Y 1 5 5 3 True Rating ⇒ Data is Missing Not At Random (MNAR) Example adapted from (Steck et al., 2010) 2

Recommendations as Treatments: Debiasing Learning and Evaluation Selection Bias in Recommendation  Why is there selection bias? o User-induced bias (e.g., browsing) o System-induced bias (e.g., advertising)  Question: What happens if we ignore selection bias? (Marlin et al., 2007; Steck, 2011; Hernándandez-Lobato et al., 2014) 3

Recommendations as Treatments: Debiasing Learning and Evaluation Evaluating Recommendations under Selection Bias Horror Romance Drama ෠ 𝑍 5 5 5 5 1 1 3 3 Recommend 5 1 3 5 5 5 5 Horror Lovers 5 5 1 3 5 5 1 3 5 5 5 5 5 5 3 3 O 5 5 5 5 1 1 3 3 3 3 Observed Y/N 1 1 5 5 3 3 1 1 5 5 3 3 1 5 3 5 5 5 5 Romance Lovers 5 5 3 3 5 5 5 3 5 5 5 3 Y 1 1 5 5 5 5 3 3 True Rating ⇒ Observed ratings are misleading due to selection bias 4

Recommendations as Treatments: Debiasing Learning and Evaluation Evaluating Predicted Ratings under Selection Bias ෠ ෠ 𝑍 𝑍 1 2 Pred Ratings (worse) Pred Ratings (better) Horror Romance Drama Horror Romance Drama 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 5 5 5 5 1 1 1 1 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 1 5 5 5 5 5 5 1 3 5 1 3 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 5 5 5 5 1 1 1 1 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 1 5 5 5 5 5 Horror Horror Lovers Lovers 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 5 5 5 5 1 1 1 1 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 5 5 5 5 1 1 1 1 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 5 5 5 5 1 1 1 1 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 1 1 1 1 1 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 1 5 3 1 5 3 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 1 1 1 1 1 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 Romance Romance Lovers Lovers 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 1 1 1 1 1 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 1 1 1 1 1 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 1 1 1 1 1 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5

Recommendations as Treatments: Debiasing Learning and Evaluation Evaluating Predicted Ratings under Selection Bias ෠ ෠ 𝑍 𝑍 1 2 Pred Ratings (worse) Pred Ratings (better) Horror Romance Drama Horror Romance Drama 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 3 3 3 3 3 5 5 5 5 5 1 1 1 1 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 1 5 5 5 5 5 5 1 3 5 1 3 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 3 3 3 3 3 5 5 5 5 5 1 1 1 1 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 1 5 5 5 5 5 Horror Lovers Horror Lovers 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 3 3 3 3 3 5 5 5 5 5 1 1 1 1 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 1 5 5 5 5 5 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 3 3 3 3 3 5 5 5 5 5 1 1 1 1 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 1 5 5 5 5 5 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 3 3 3 3 3 5 5 5 5 5 1 1 1 1 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 1 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 3 3 3 3 3 1 1 1 1 1 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 1 5 3 1 5 3 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 3 3 3 3 3 1 1 1 1 1 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 Romance Romance Lovers Lovers 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 3 3 3 3 3 1 1 1 1 1 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 3 3 3 3 3 1 1 1 1 1 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 3 3 3 3 3 1 1 1 1 1 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 6

Recommendations as Treatments: Debiasing Learning and Evaluation Evaluating Predicted Ratings under Selection Bias ෠ ෠ 𝑍 𝑍 1 2 Pred Ratings (worse) Pred Ratings (better) Horror Romance Drama Horror Romance Drama 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 3 3 3 3 3 5 5 5 5 5 1 1 1 1 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 1 5 5 5 5 5 5 1 3 5 1 3 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 3 3 3 3 3 5 5 5 5 5 1 1 1 1 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 1 5 5 5 5 5 Horror Horror Lovers Lovers 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 3 3 3 3 3 5 5 5 5 5 1 1 1 1 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 3 3 3 3 3 5 5 5 5 5 1 1 1 1 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 3 3 3 3 3 5 5 5 5 5 1 1 1 1 5 5 5 5 5 1 1 1 1 1 5 5 5 5 5 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 1 1 1 1 1 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 1 5 3 1 5 3 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 1 1 1 1 1 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 Romance 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 Romance Lovers Lovers 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 1 1 1 1 1 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 1 1 1 1 1 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 3 3 3 3 5 3 3 3 3 3 1 1 1 1 1 5 5 5 5 1 1 1 1 1 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 ⇒ Observed losses are misleading due to selection bias 7

Recommendations as Treatments: Debiasing Learning and Evaluation Recommendations as Treatments  Question : How can we fix the effects of selection bias? o Connection to potential outcomes framework Counterfactual Outcomes 𝑍 Observed Outcomes ෨ 𝑍 treatments movies treatments 5 5 1 3 5 5 5 1 3 5 5 1 3 5 5 5 3 patients users 5 5 1 3 3 patients 1 1 5 5 3 3 5 5 1 5 3 5 5 5 5 3 1 5 5 3 ⇒ Understand assignment mechanism (Imbens & Ruben, 2015) 8

Recommendations as Treatments: Debiasing Learning and Evaluation Debiasing Evaluation Propensities P Horror Romance Drama  Assignment mechansim for recommendation: o 𝑄 𝑣,𝑗 = 𝑄 𝑃 𝑣,𝑗 = 1 p p/10 p/2  Use Inverse-Propensity-Scoring Estimator (IPS) to obtain unbiased estimate: p/10 p p/2 1 1 2 𝑆 𝐽𝑄𝑇 ෠ ෠ 𝑣,𝑗 − ෠ 𝑍|𝑄 = ෍ 𝑍 𝑍 𝑣,𝑗 𝑉 ⋅ 𝐽 𝑄 𝑣,𝑗 𝑣,𝑗 :𝑃 𝑣𝑗 =1 (Little & Rubin, 2002; Cortes et al., 2008; Bickel et al., 2009; Sugiyama & Kawanabe, 2012). 9

Recommendations as Treatments: Debiasing Learning and Evaluation Propensity estimation  Two settings: o Experimental ̶- Propensities are under our control; known by design (e.g., ad placement) o Observational ̶- Users self-select; need to Observations O estimate 𝑄 𝑣,𝑗 Horror Romance Drama 1 0 1 0 0 1 0 0 0 0 0 0 1 0 0  Estimate parameter of binary random 0 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 1 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 1 0 0 variables : 0 1 0 1 0 0 0 0 1 0 1 0 0 0 1 1 0 0 0 1 1 0 0 1 0 0 1 0 0 1 𝑄 𝑣,𝑗 = 𝑄 𝑃 𝑣,𝑗 = 1 | 𝑌, ෨ 𝑍 0 0 0 0 0 0 0 1 0 1 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 1 1 0 0 0 0  Variety of models: Logistic Regression, 1 0 0 0 0 0 1 1 0 0 0 0 0 1 0 Naïve Bayes, etc. 10

Recommendations as Treatments: Debiasing Learning and Evaluation Debiasing Evaluation  Robustness to selection bias: Severity of Selection Bias Severity of Selection Bias 11

Recommendations as Treatments: Debiasing Learning and Evaluation Debiasing Evaluation  Robustness to inaccurate propensities: IPS-est More accurate propensities More accurate propensities 12

Recommendations as Treatments: Debiasing Learning and Evaluation Debiasing Learning  Empirical Risk Minimization (ERM) successful in many settings (Cortes & Vapnik, 1995)  Use ERM together with Inverse-Propensity-Scoring Estimator (IPS) 𝑍 𝐹𝑆𝑁 = argmin ෠ 𝑆 𝐽𝑄𝑇 ෠ ෠ 𝑍 | 𝑄 ෠ 𝑍∈ℋ  For matrix factorization with MSE loss: 1 2 + 𝜇 2 + 𝑋 𝐺 𝑍 𝐹𝑆𝑁 = argmin ෠ 2 ෍ 𝑍 𝑣,𝑗 − 𝑊 𝑣 𝑋 𝑊 𝐺 𝑗 𝑄 𝑣,𝑗 𝑊,𝑋 𝑃 𝑣,𝑗 =1 propensity weight 13

Recommendations as Treatments: Debiasing Learning and Evaluation - PowerPoint PPT Presentation

Recommendations as Treatments: Debiasing Learning and Evaluation ICML 2016, NYC Schnabel , Adith Swaminathan , Tob obias Sc Ashudeep Singh , Navin Chandak , Thorsten Joachims Cornell University, Google Funded in part

We need drugs that New treatments, new mechanisms of action Cathepsin K inhibitors

When should disease-modifying treatments be discontinued in patients with multiple sclerosis? An

IMRT treatments in Belgium: IMRT treatments in Belgium: an update an update S. Vy Vyn nckier

GWG Recommendations Vice President, Portfolio Development and Review California Institute for

GWG Recommendations Vice President, Portfolio Development and Review California Institute for

FNFNES Recommendations Recommendations from the FNFNES are currently being finalized,

4/17/20 Design, Analysis, and Assessment of Learning Workgroup Recommendations Recommendations

OUT OF THE PIPELINE: NOVEL TARGETS AND TREATMENTS FOR SCHIZOPHRENIA Learning Objectives

Acute Burn Healed in 11 Days Four Treatments Acute Burn Healed in 11 Days Four Treatments 4 FSM

Traditional Anoxia treatments Pest control for artistic and cultural heritage conservation is a

Implementation Plan for ALL (27) Recommendations, including 2 IOC-specific RECOMMENDATIONS ON

DDSN Recommendations 1 Department of Disabilities and Special Needs Recommendations to the House

1 2 BUSINESS POTENTIAL Aero Treatments Private Limited is a Independent Third party treatment

Coding RT Treatments: Breast Cancer NAACCR OCTOBER 3, 2019 WILSON APOLLO, CTR, RTT, MS 1 2

Heat, Cold, and Contrast Treatments 27b Hydrotherapy: Heat, Cold, and Contrast Treatments

An Ontological Analysis of the Notion of Treatment Paul Johannesson and Erik Perjons, Stockholm

Section O Special Treatments, Procedures, and Programs Minimum Data Set (MDS) 3.0 Section X 1

The Council should consider these recommendations and data from the resident surveys as it

CONDOROIL in short CONDOROIL Group focuses on surface treatments for more than 40 years and

Actually, heat treatments comes with high costs in manpower, logistics, energy consumpiton... Heat

MenACWY-TT (MenQuadfi): Evidence to Recommendations Framework (EtR), Grading of Recommendations,

Visualiza(on of Unique Temporal Sequences of Treatments and

The Potential for New Treatments for Alzheimers Disease: Where Are We Headed ? Dr.

Heat Treatments Paul Fields Cereal Research Centre, Winnipeg pfields@agr.gc.ca