Share
iCAST 2020 Invited Talk: Fairness in AI Service and Awareness for Computational Social Science
iCAST2020にて基調講演を行った際のスライドになります。
Abstract
This talk aims to present an overview of the common pitfalls for applying machine learning techniques to real-world problems from a perspective of fairness. This talk mainly highlights the importance of diversity of the data and the problem related to algorithmic bias. In the age of information overload, machine learning becomes increasingly important in everyday life. There has been a growing interest in discovering the harmful effect of bias in machine learning and a way to take fairness into service. Based on our research and experience in the industry, we discuss open questions for further application.
Biography
Tomoki Fukuma is the founder and CEO of TDAI Lab, a machine learning AI startup, founded in Tokyo in 2016. He is now a Ph.D. student in the Department of Systems Innovation, School of Engineering, The University of Tokyo. The main area of his interest is learning the true quality of the content in an online platform. This includes topics in unbiased learning-to-rank, recommendation, and AI for Society. He is also a Japanese ballroom dancer, representing Japan from 2016 to 2019.
株式会社TDAI Lab
Transcript
71 slides
-
ɹConfidentialɹ© TDAI Lab All rights reserved. 1 Fairness in AI
Service and Awareness for Computational Social Science Tomoki Fukuma TDAI Lab Co., Ltd. CEO The 11th International Conference on Awareness Science and Technology (iCAST 2020) December 7-9, 2020 -
ɹConfidentialɹ© Japan Data Science Consortium. All right reserved. ɹConfidentialɹ© TDAI
Lab All rights reserved. 2 Agenda • Prologue • Technique to Measure the Quality of the Contents • Pros/Cons of the Techniques • Biased Model Case Study • How to Fight Against Bias • Summary -
ɹConfidentialɹ© TDAI Lab All rights reserved. Information Overload With the
widespread use of the Internet, the opinions of various people can easily be collected in large quantities. https://www.statista.com/chart/19058/how-many-websites-are-there/ -
ɹConfidentialɹ© TDAI Lab All rights reserved. Information Overload https://reputationstacker.com/how-many-reviews-do-you-need/ Users
cannot find or don’t know what they want. Service providers have to systematically select valuable information for users. ɹɹ But How ? -
ɹConfidentialɹ© TDAI Lab All rights reserved. ɹɹ Information Collection from
Pull to Push Pull Push “search, recommendation” Users are moving from traditional pull-type information collection to push-type information collection -
ɹConfidentialɹ© Japan Data Science Consortium. All right reserved. ɹConfidentialɹ© TDAI
Lab All rights reserved. 6 Agenda • Prologue • Technique to Measure the Quality of the Contents • Pros/Cons of the Techniques • Biased Model Case Study • How to Fight Against Bias • Summary -
ɹConfidentialɹ© TDAI Lab All rights reserved. Technique Summary Voting Mechanism
Unsupervised Learning Supervised learning (review, Q&A, SNS) ɾGraph-Based Ranking ɾAnomaly Detection ɾCollaborative Filtering ɾLearning-to-Rank ɾHelpfulness Prediction ɾFake Detection Wisdom of the Crowd Machine Learning -
ɹConfidentialɹ© TDAI Lab All rights reserved. Case 1.1 : Wisdom
of the Crowd ut How? A mechanism that allows users to vote content helpfulness or likeness Twitter StackExchange Amazon -
ɹConfidentialɹ© TDAI Lab All rights reserved. Case 1.1 : Wisdom
of the Crowd ɾFew contents acquires the majority of the votes. ɾMost of the contents are unrated. → We need machine learning to estimate the quality of the unrated contents. -
ɹConfidentialɹ© TDAI Lab All rights reserved. Case 1.2 : Unsupervised
Machine Learning Google: PageRank ranks web pages according to “importance” REV2 is an anomaly detection method based on rating behavior, detecting “fraudulent” users. – Every webpage is a node, and every web-link is an edge. – Every reviewer is a node, and every rating is an edge. REV2: Fraudulent User Prediction in Rating Platforms -
ɹConfidentialɹ© TDAI Lab All rights reserved. Case 1.3 : Unsupervised
Machine Learning ut How? Youtube Spotify Netflix Collaborative Filtering is based on "If the interests of a user A and user B are similar, user A will behave similarly to user B in the future". Create different rankings from people to people. -
ɹConfidentialɹ© TDAI Lab All rights reserved. Case 1.4 : Supervised
Machine Learning Prediction Airbnb: Improving Deep Learning For Airbnb Search Helpfulness Prediction aims to predict helpfulness score based on review text, metadata. -
ɹConfidentialɹ© TDAI Lab All rights reserved. Case 1.5 : Supervised
Machine Learning Facebook: Deepfake Detection Fake Detection KDD 2020 Fake News Detection Assessing Human Value Hiring Credibility Insurance fee Future Crime -
ɹConfidentialɹ© TDAI Lab All rights reserved. Are They Truly Estimating
the Quality of the Contents? -
ɹConfidentialɹ© Japan Data Science Consortium. All right reserved. ɹConfidentialɹ© TDAI
Lab All rights reserved. 15 Agenda • Prologue • Technique to Measure the Quality of the Contents • Pros/Cons of the Techniques • Biased Model Case Study • How to Fight Against Bias • Summary -
ɹConfidentialɹ© TDAI Lab All rights reserved. Technique Summary Voting Mechanism
Unsupervised Learning Supervised learning (review, Q&A, SNS) ɾGraph-Based Ranking ɾAnomaly Detection ɾCollaborative Filtering ɾLearning-to-Rank ɾHelpfulness Prediction ɾFake Detection Wisdom of the Crowd Machine Learning -
ɹConfidentialɹ© TDAI Lab All rights reserved. Wisdom of the Crowd
Pros. Cons. ɾEasy to implement ɾMost of the contents are unrated. ɾPopularity ≠ Quality ※Wisdom of the crowd assumes the independence of the decision. -
ɹConfidentialɹ© TDAI Lab All rights reserved. Cognitive Bias ɾ Content
topic describes what the content is about; ɾ Content quality refers to what the voter assesses the quality of the content to be; ɾ Presentation order captures the position of the content in a list-style webpage; ɾ Social influence in terms of prior votes as seen by the voter; ɾ Author reputation reflects the online reputation of the user who created the content. Burghardt et al.* suggests a number of factors that may affect votes on content: *1 https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0173610 -
ɹConfidentialɹ© TDAI Lab All rights reserved. Cognitive Bias ɾ Content
topic describes what the content is about; ɾ Content quality refers to what the voter assesses the quality of the content to be; ɾ Presentation order captures the position of the content in a list-style webpage; ɾ Social influence in terms of prior votes as seen by the voter; ɾ Author reputation reflects the online reputation of the user who created the content. Burghardt et al.* suggests a number of factors that may affect votes on content: *1 https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0173610 ɹɹ Independent judges are violated → Decrease in diversity of opinions → Less quality of wisdom of the crowd -
ɹConfidentialɹ© TDAI Lab All rights reserved. Secondary Bias Winner-cycle Bias
Early Bird Bias ɹɹ popularity and quality doesn’t have high correlation -
ɹConfidentialɹ© TDAI Lab All rights reserved. Technique Summary Voting Mechanism
Unsupervised Learning Supervised learning (review, Q&A, SNS) ɾGraph-Based Ranking ɾAnomaly Detection ɾCollaborative Filtering ɾLearning-to-Rank ɾHelpfulness Prediction ɾFake Detection Wisdom of the Crowd Machine Learning -
ɹConfidentialɹ© TDAI Lab All rights reserved. Unsupervised Learning Pros. ɾCertain
degree of performance even without annotated data ※ if the assumption is valid Cons. ɾRequires domain specific knowledge ɾad hoc maintenance for improving performance -
ɹConfidentialɹ© TDAI Lab All rights reserved. The limitation of Unsupervised
learning ex.) K-Means Clustering : assumes features mapped nearly have similar pattern → easy to fail when modeling high-dimensional data PageRank: pages that are linked to important pages are important → winner-cycle bias ɹɹ Assumption of algorithm internally contains the bias of developer → If assumptions are different from real-world behavior, the algorithm sometimes behaves unfairly. -
ɹConfidentialɹ© TDAI Lab All rights reserved. Technique Summary Voting Mechanism
Unsupervised Learning Supervised learning (review, Q&A, SNS) ɾGraph-Based Ranking ɾAnomaly Detection ɾCollaborative Filtering ɾLearning-to-Rank ɾHelpfulness Prediction ɾFake Detection Wisdom of the Crowd Machine Learning -
ɹConfidentialɹ© TDAI Lab All rights reserved. Supervised Learning Pros. ɾDomain-invariant
universal technique (no need domain specific knowledge) ɾAdaptive to user behavior Cons. ɾRequires sufficient annotated data ɾvulnerable to “Data Bias” -
ɹConfidentialɹ© TDAI Lab All rights reserved. What is Learning-to Rank
Traditionally, Learning to Rank (LTR) is supervised through annotated datasets: Relevance annotations for query-document pairs provided by human judges. ɾQueries, representing queries users will issue, ɾDocuments, per query a preselected set of documents to be ranked, ɾRelevance Labels, indicating relevance/preference per document-query pair. -
ɹConfidentialɹ© TDAI Lab All rights reserved. Supervised Learning-to Rank However,
over time several limitations of this approach have become apparent. ɾexpensive to make (Qin and Liu, 2013; Chapelle and Chang, 2011). ɾunethical to create in privacy-sensitive settings (Wang et al., 2016a). ɾimpossible for small scale problems, e.g., personalization. ɾstationary, cannot capture future changes in relevancy (Lefortier et al., 2014). ɾnot necessarily aligned with actual user preferences (Sanderson, 2010), i.e., annotators and users often disagree -
ɹConfidentialɹ© TDAI Lab All rights reserved. Supervised Learning-to Rank Traditionally,
learning to rank is supervised through annotated datasets: Relevance annotations for query-document pairs provided by human judges. However, over time several limitations of this approach have become apparent. ɾexpensive to make (Qin and Liu, 2013; Chapelle and Chang, 2011). ɾunethical to create in privacy-sensitive settings (Wang et al., 2016a). ɾimpossible for small scale problems, e.g., personalization. ɾstationary, cannot capture future changes in relevancy (Lefortier et al., 2014). ɾnot necessarily aligned with actual user preferences (Sanderson, 2010), i.e., annotators and users often disagree ɹɹ Can we naively use click as annotation? ɹɹ The model will be biased if you are not careful -
ɹConfidentialɹ© Japan Data Science Consortium. All right reserved. ɹConfidentialɹ© TDAI
Lab All rights reserved. 29 Agenda • Prologue • Technique to Measure the Quality of the Contents • Pros/Cons of the Techniques • Biased Model Case Study • How to Fight Against Bias • Summary -
ɹConfidentialɹ© TDAI Lab All rights reserved. Bias In Bias Out
Trained w/ Biased Data Clean Algorithm Biased Algorithm BERT image is cited from https://arxiv.org/pdf/2004.06660.pdf -
ɹConfidentialɹ© Japan Data Science Consortium. All right reserved. ɹConfidentialɹ© TDAI
Lab All rights reserved. 31 Agenda • Prologue • Technique to Measure the Quality of the Contents • Pros/Cons of the Techniques • Biased Model Case Study • (Case1) Machine Learning against human • (Case2) Learning to Rank • (Case3) Recommendation • How to Fight Against Bias • Summary -
ɹConfidentialɹ© TDAI Lab All rights reserved. Case2.1 Income predictions ɾtraining
a basic classifier that can predict whether or not a person's income is larger than $50K a year. ɾIt is important to note sensitive attributes(race and sex) are not part of the features used for training the model. https://godatadriven.com/blog/towards-fairness-in-ml-with-adversarial-networks/ -
ɹConfidentialɹ© TDAI Lab All rights reserved. Case2.1 Income predictions ɾtraining
a basic classifier that can predict whether or not a person's income is larger than $50K a year. ɾIt is important to note sensitive attributes(race and sex) are not part of the features used for training the model. ɹɹ Our classifier did not have access to the race and sex attributes. But still we have model biased against women and black people. ɹɹ What caused our classifier to behave this way? -
ɹConfidentialɹ© TDAI Lab All rights reserved. Difficulty of removing bias
ɾModel can indirectly learn these biases, for example, through other characteristics (e.g., education level, zip-code of residence, religion, career, and preference of music). ɾAs a result, we end-up with the unfair predictions observed in previous section, even after having removed the race and sex attributes. -
ɹConfidentialɹ© Japan Data Science Consortium. All right reserved. ɹConfidentialɹ© TDAI
Lab All rights reserved. 37 Agenda • Prologue • Technique to Measure the Quality of the Contents • Pros/Cons of the Techniques • Biased Model Case Study • (Case1) Machine Learning against human • (Case2) Learning to Rank • (Case3) Recommendation • How to Fight Against Bias • Summary -
ɹConfidentialɹ© TDAI Lab All rights reserved. Supervised Learning-to Rank Traditionally,
learning to rank is supervised through annotated datasets: Relevance annotations for query-document pairs provided by human judges. However, over time several limitations of this approach have become apparent. ɾexpensive to make (Qin and Liu, 2013; Chapelle and Chang, 2011). ɾunethical to create in privacy-sensitive settings (Wang et al., 2016a). ɾimpossible for small scale problems, e.g., personalization. ɾstationary, cannot capture future changes in relevancy (Lefortier et al., 2014). ɾnot necessarily aligned with actual user preferences (Sanderson, 2010), i.e., annotators and users often disagree ɹɹ Can we naively use click as annotation? ɹɹ The model will be biased if you are not careful -
ɹConfidentialɹ© TDAI Lab All rights reserved. Presentation Bias Bias on
the Web | June 2018 | Communications of the ACM -
ɹConfidentialɹ© TDAI Lab All rights reserved. Presentation Bias Bias on
the Web | June 2018 | Communications of the ACM -
ɹConfidentialɹ© TDAI Lab All rights reserved. Position Based Model (PBM)
ℙ(C = 1 ∣ q, d, k) Click = ℙ(E = 1 ∣ k) Examination ⋅ ℙ(R = 1 ∣ q, d) Relevance Examination 👀 Relevance💰 Click ✅ O=1 R=1 Y=1 O=1 R=0 Y=0 — — — O=0 R=1 Y=0 O=0 R=1 Y=0 O=0 R=0 Y=0 We often do not know the true relevance labels , but can only observe clicks: y (di) Unbiased Learning-to-Rank with Biased Feedback -
ɹConfidentialɹ© TDAI Lab All rights reserved. Position Based Model (PBM)
ℙ(C = 1 ∣ q, d, k) Click = ℙ(E = 1 ∣ k) Examination ⋅ ℙ(R = 1 ∣ q, d) Relevance Examination 👀 Relevance💰 Click ✅ O=1 R=1 Y=1 O=1 R=0 Y=0 — — — O=0 R=1 Y=0 O=0 R=1 Y=0 O=0 R=0 Y=0 ɹɹ • A click on document is a biased and noisy indicator that is relevant • A missing click does not necessarily indicate non-relevance. di di -
ɹConfidentialɹ© Japan Data Science Consortium. All right reserved. ɹConfidentialɹ© TDAI
Lab All rights reserved. 43 Agenda • Prologue • Technique to Measure the Quality of the Contents • Pros/Cons of the Techniques • Biased Model Case Study • (Case1) Machine Learning against human • (Case2) Learning to Rank • (Case3) Recommendation • How to Fight Against Bias • Summary -
ɹConfidentialɹ© TDAI Lab All rights reserved. Recommendation https://medium.com/@lz2576/a-first-look-at-recommendation-system-with-matrix-factorization-and-neural-nets-7e21e54295c ɾCollaborative Filtering
is based on "If the interests of a user A and user B are similar, user A will behave similarly to user B in the future". ɾTrying to predict the missing value in the matrix -
ɹConfidentialɹ© TDAI Lab All rights reserved. Exposure Model P (Yu,i
= 1) click prob = P (Ou,i = 1) exposure prob ⋅ P (Ru,i = 1) relevance level Recommendation Exposure Relevance Click Yes O=1 R=1 Y=1 Yes O=1 R=0 Y=0 — — — — No O=0 R=1 Y=0 No O=0 R=1 Y=0 No O=0 R=0 Y=0 Modeling User Exposure in Recommendation Liang et al. WWW 2016 -
ɹConfidentialɹ© TDAI Lab All rights reserved. Exposure Model Liang et
al. WWW 2016 P (Yu,i = 1) click prob = P (Ou,i = 1) exposure prob ⋅ P (Ru,i = 1) relevance level Recommendation Exposure Relevance Click Yes O=1 R=1 Y=1 Yes O=1 R=0 Y=0 — — — — No O=0 R=1 Y=0 No O=0 R=1 Y=0 No O=0 R=0 Y=0 ɹɹ • A click on item is a biased and noisy indicator that is relevant • A missing click does not necessarily indicate non-relevance. i i -
ɹConfidentialɹ© TDAI Lab All rights reserved. Missing Not At Random
ɾThe observations are biased by self-selection bias. → User like to click the items they like, not the other way around -
ɹConfidentialɹ© TDAI Lab All rights reserved. Algorithms amplify the bias
Y-axis: Cumulative % of exposure X-axis: Cumulative % of producers (ranked in increasing exposure) Google Local: Top 20 % businesses got ~80 % of total exposure Lastfm: Bottom 60 % artists got ~20 % of total exposure -
ɹConfidentialɹ© TDAI Lab All rights reserved. Filter Bubble Users are
unknowingly only exposed to information that they already believed and like, as if they were trapped in a "bubble “ →less diversity -
ɹConfidentialɹ© TDAI Lab All rights reserved. Recap ɾBias In Bias
Out ɾMost web systems are trained with user implicit feedback. ɹˠ Click can only be done on things that are shown ɾML learn to reinforce their own bias ɹˠ less diversity -
ɹConfidentialɹ© Japan Data Science Consortium. All right reserved. ɹConfidentialɹ© TDAI
Lab All rights reserved. 51 Agenda • Prologue • Technique to Measure the Quality of the Contents • Pros/Cons of the Techniques • Biased Model Case Study • How to Fight Against Bias • Summary -
ɹConfidentialɹ© TDAI Lab All rights reserved. Bias In Bias Out
Trained w/ Biased Data Clean Algorithm BERT image is cited from https://arxiv.org/pdf/2004.06660.pdf Debiasing Algorithm Clean Algorithm -
ɹConfidentialɹ© Japan Data Science Consortium. All right reserved. ɹConfidentialɹ© TDAI
Lab All rights reserved. 53 Agenda • Prologue • Technique to Measure the Quality of the Contents • Pros/Cons of the Techniques • Biased Model Case Study • How to Fight Against Bias • IPS • Adversarial Training • Summary -
ɹConfidentialɹ© TDAI Lab All rights reserved. Inverse Propensity Scoring If
no click noise is present, this provides an unbiased estimate: P (Ru,i = 1) = P (Yu,i = 1) P (Ou,i = 1) P(R = 1 ∣ q, d) = P(C = 1 ∣ q, d, k) P(E = 1 ∣ k) Search engine(PBM) : Recommendation(Exposure Model): -
ɹConfidentialɹ© TDAI Lab All rights reserved. Inverse Propensity Scoring Handling
Position Bias for Unbiased Learning to Rank in Hotels Search P(R = 1 ∣ q, d) = P(C = 1 ∣ q, d, k) P(E = 1 ∣ k) Difficult to estimate Requiring randomized experiments -
ɹConfidentialɹ© TDAI Lab All rights reserved. Inverse Propensity Scoring Handling
Position Bias for Unbiased Learning to Rank in Hotels Search P(R = 1 ∣ q, d) = P(C = 1 ∣ q, d, k) P(E = 1 ∣ k) Difficult to estimate Requiring randomized experiments If a site is clicked 5 times less visible, let the model learn 5 times more weightedɹ -
ɹConfidentialɹ© TDAI Lab All rights reserved. Inverse Propensity Scoring P
(Ru,i = 1) = P (Yu,i = 1) P (Ou,i = 1)?? Sometimes substituted by propensity score amortized over user or item: P (Ou,* = 1), P (O*,i = 1) P(R = 1 ∣ q, d) = P(C = 1 ∣ q, d, k) P(E = 1 ∣ k) Almost impossible to estimate user-wise propensity score PBM also makes strong assumptions that “examination propensity only depends on position” →ɹSubstantial parameter reduction P (E = 1 ∣ u, k, d) Human Bias? -
ɹConfidentialɹ© Japan Data Science Consortium. All right reserved. ɹConfidentialɹ© TDAI
Lab All rights reserved. 58 Agenda • Prologue • Technique to Measure the Quality of the Contents • Pros/Cons of the Techniques • Biased Model Case Study • How to Fight Against Bias • IPS • Adversarial Training • Summary -
ɹConfidentialɹ© Japan Data Science Consortium. All right reserved. ɹConfidentialɹ© TDAI
Lab All rights reserved. 59 Generative adversarial network -
ɹConfidentialɹ© TDAI Lab All rights reserved. Recap: Generative Adversarial Network
https://poloclub.github.io/ganlab/ -
ɹConfidentialɹ© TDAI Lab All rights reserved. Case1.3 Income predictions ɾtraining
a basic classifier that can predict whether or not a person's income is larger than $50K a year. ɾIt is important to note sensitive attributes(race and sex) are not part of the features used for training the model. -
ɹConfidentialɹ© TDAI Lab All rights reserved. Typical Machine Learning ɾTask
specific model are trained. -
ɹConfidentialɹ© TDAI Lab All rights reserved. Fairness with adversarial training
Based on Learning to Pivot with Adversarial Networks FRAGE: Frequency-Agnostic Word Representation Task specific model + Discriminator ɾDiscriminator tries to predict sensitive features. ɾTask specific model need to fool the discriminator -
ɹConfidentialɹ© TDAI Lab All rights reserved. Fair income predictions Task
specific model + Discriminator ɾDiscriminator tries to predict sensitive features. ɾTask specific model need to fool the discriminator -
ɹConfidentialɹ© Japan Data Science Consortium. All right reserved. ɹConfidentialɹ© TDAI
Lab All rights reserved. 65 Agenda • Prologue • Technique to Measure the Quality of the Contents • Pros/Cons of the Techniques • Biased Model Case Study • How to Fight Against Bias • Summary -
ɹConfidentialɹ© TDAI Lab All rights reserved. Trusted AI ML is
necessary in the age of information overload. Developers must build technology people can trust. → using human bias to “reduce” bias Demand for: ɾFairness ɾAccountability ɾTransparency ɾRobustness -
ɹConfidentialɹ© TDAI Lab All rights reserved. Trusted AI ← Explainability
/ Fairness ← Fairness ← Robustness ← Trust ← Trust ← Fairness -
ɹConfidentialɹ© TDAI Lab All rights reserved. Case Study:Babysitter Risk Rating
Common Pitfalls for Studying the Human Side of Machine Learning Neurips 2018 Xcorp launches a new service that uses social media data to predict whether a babysitter candidate is likely to abuse drugs or exhibit other undesirable tendencies (e.g. aggressiveness, disrespectfulness, etc.) Using computational techniques, Xcorp will produce a score to rate the riskiness of the candidates. Candidates must opt in to being scored when asked by a potential employer. This product produces a rating of the quality of the babysitter candidate from 1-5 and displays this to the hiring parent. What would it mean for this system to be fair? ɹ -
ɹConfidentialɹ© TDAI Lab All rights reserved. Pretrained Model Age ɾRecent
NLP progress(2018~) is based on pretrained models trained with huge corpus. ɾThe risk of using and trusting those models blindly is increasing ɾWhat if they are biased ??