Uncovering the Latent Structures of Crowd Labeling

Transcription

1 Uncovering the Latent Structures of Crowd Labeling Tian Tian and Jun Zhu Presenter:XXX Tsinghua University 1 / 26

2 Motivation Outline 1 Motivation 2 Related Works 3 Crowdsourcing Latent Class 4 Experiments 5 Conclusion 2 / 26

3 Motivation Motivation: Background Artificial intelligence are relying more and more on large-scale training datasets. Expert labeling are expensive and time-consuming. 3 / 26

4 Motivation Motivation: Crowdsourcing Solution Using multiple web workers to label each item. Recovery the ground truth from the noisy data. Table: Different workers may give inconsistent labels to a same item. worker a worker b worker c worker d item item item item / 26

5 Related Works Outline 1 Motivation 2 Related Works 3 Crowdsourcing Latent Class 4 Experiments 5 Conclusion 5 / 26

6 Related Works Majority Voting Majority Voting Assumption For every worker, the ground truth label is always the most common to be given, and the labels for each item are given independently. P(Y m = d) = (n,m) I δ W nm,d d,(n,m) I δ, m, (1) W nm,d 6 / 26

7 Related Works Dawid-Skene Estimator Dawid-Skene estimator Assumption The performance of a worker is consistent across different items, and his or her behavior can be measured by a confusion matrix. Table: An example of binary classification confusion matrix. label A label B label A 1 0 label B 1/3 2/3 7 / 26

8 Related Works Dawid-Skene Estimator Dawid-Skene estimator Assumption The performance of a worker is consistent across different items, and his or her behavior can be measured by a confusion matrix. P(W q, p) = m d q d n,l p ndl δ W nm,l, (2) p ndl : the probability that worker n labels an item as l when it s ground truth label is d. 8 / 26

9 Crowdsourcing Latent Class Outline 1 Motivation 2 Related Works 3 Crowdsourcing Latent Class 4 Experiments 5 Conclusion 9 / 26

10 Crowdsourcing Latent Class Latent Class Assumptions Latent Class Assumptions Assumption I. Each item belongs to one specific latent class only. Assumption II. Items in the same latent class should be classified into the same category. 10 / 26

11 Crowdsourcing Latent Class Latent Class Assumptions Latent Class Assumptions Each item belongs to one specific latent class only. Items in the same latent class in the same category. Figure: Latent classes of a binary classification task to classify fruit and vegetable. 11 / 26

12 Crowdsourcing Latent Class Latent Class Dawid-Skene Estimator Latent Class Confusion Matrix Extend the original confusion matrix. An entry p nkl in this matrix represents the probability that worker n gives label l to an item which belongs to latent class k. Table: An example of latent class confusion matrix. tomato pumpkin cuke apple Fruit 2/3 0 1/3 1 Vegetable 1/3 1 2/ / 26

13 Crowdsourcing Latent Class Latent Class Dawid-Skene Estimator Chinese Restaurant Process The number of latent classes K is unknown in advance. we choose a nonparametric prior for the latent class confusion matrix to learn the number of classes. Z i is the latent class assignment of item i, with the CRP prior, the probability Old Class : P(Z i = k) n k + α, k = 1 K, (3) New Class : P(Z i = K + 1) α. (4) 13 / 26

14 Crowdsourcing Latent Class Latent Class Dawid-Skene Estimator Nonparametric Dawid-Skene Model Full Bayesian Model: Assign : Z α c CRP(α c ), (5) Entries : p nk α d Dirichlet(α d ), n, k, (6) Obervation : W nm Z, p n Multinomial(A nm ), n, m, (7) Here A nm = {A nm1,, A nmd }, A nmd = K k=1 p nkd δ Zm,k 14 / 26

15 Crowdsourcing Latent Class Latent Class Dawid-Skene Estimator Inference We use Gibbs Sampling to infer the hidden variables. Conditional Distributions: confusion matrix parameter: p nk Z, W Dirichlet(p nk B nk ), n, k, (8) where B nkd = M m=1 δ W nm,dδ Zm,k + α d /D. hidden variables, when k K, P(Z m = k Z m, p, W) n k When generating a new class, P(Z m = k new Z m, p, W) α c N n=1 N n=1 d=1 D δ p Wnm,d nkd, (9) D d=1 Γ(δ W nm,d + α d /D). (10) Γ(1 + α d ) 15 / 26

16 Crowdsourcing Latent Class Latent Class Minimax Entropy Estimator Latent Class Minimax Entropy Estimator We also extend the minimax entropy estimator. Idea: change the category constraints. New objective: min Z max p nmd log p nmd α m τmd 2 p,τ,σ 2 n,m,d m,d n,m,d ( ) pnmd δ Wnm,d = τmd, m, d, β n σ 2 ndk 2 m s.t. n ( ) pnmd δ Wnm,d δzm,k = σ ndk, n, k, p nmd = 1, n, m. d (11) 16 / 26

17 Crowdsourcing Latent Class Category Recovery Category Recovery Regard items in a same class as one imaginary item, here we call it a hyper-item. We use a generalized Dawid-Skene estimator with hyper-items to estimate the category assignments. P(W q, p) = n q d p nkd ndl, (12) k d n,l where n nkd = m δ W nm,dδ Zm,k is the count of labels that worker n gives to hyper-item k. 17 / 26

18 Experiments Outline 1 Motivation 2 Related Works 3 Crowdsourcing Latent Class 4 Experiments 5 Conclusion 18 / 26

19 Experiments Synthetic Dataset 4 latent classes, 40 items parameters for each latent class. 2 types of workers, 20 workers of each type. K α =0.1 c α =0.2 c α =0.8 c α c =0.4 α =1.6 c Error Rate (%) # Iterations (a) α c Figure: (a) shows the numbers of latent classes found by NDS with different color. (b) shows the average category recovery error rates. (b) 19 / 26

20 Experiments Real Flower Dataset 4 flower species, 50 pics for each. 2 species for each category. 36 workers, 2366 labels in total. (a) NDS (b) LC-ME (c) NDS+CR (d) LC-ME+CR Figure: entry corresponding to image, column corresponding to flower species. (a)(b): color denotes latent class. (c)(d): denotes category. 20 / 26

21 Experiments Real Flower Dataset 4 flower species, 50 pics for each. 2 species for each category. 36 workers, 2366 labels in total. Figure: Representative pictures for different latent classes.(best viewed in color). 21 / 26

22 Experiments Category Recovery Table: Performance of several models on flowers dataset. the average error rate of 10 trials, together with standard deviation, are presented. # MV ± ± ± ± DS ± ± ± ± NDS ± ± ± ± ME ± ± ± ± LC-ME ± ± ± ± The first line means the number of workers we used on estimation tasks. 22 / 26

23 Conclusion Outline 1 Motivation 2 Related Works 3 Crowdsourcing Latent Class 4 Experiments 5 Conclusion 23 / 26

24 Conclusion Conclusion We show a latent class structures in crowdsourcing. We propose algorithms to disclose meaningful latent classes. We show that latent structures can help to achieve higher accuracies on category recovery tasks. 24 / 26

25 References Outline 1 Motivation 2 Related Works 3 Crowdsourcing Latent Class 4 Experiments 5 Conclusion 25 / 26

26 References Selected Reference Dawid, A.P., and Skene, A.M.: Maximum likelihood estimation of observer error-rates using the em algorithm. Applied statistics (1979), Dempster, A.P., Laird, N.M., Rubin, D.B., etal.: Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal statistical Society 39, 1 (1977), Li, H., Yu, B., and Zhou, D.: Error rate analysis of labeling by crowdsourcing. Raykar, V.C., Yu, S., Zhao, L.H., Valadez, G.H., Florin, C., Bogoni, L., and Moy, L.: Learning from crowds. The Journal of Machine Learning Research 11 (2010), Sheshadri, A., and Lease, M.: Square: A benchmark for research on computing crowd consensus. In First AAAI Conference on Human Computation and Crowdsourcing (2013). Tian, Y., and Zhu, J.: Learning from crowds in the presence of schools of thought. In Proceedings of the 18th ACM SIGKDD international conference on Knowledge discovery and data mining (2012), ACM, Welinder, P., Branson, S., Belongie, S., and Perona, P.: The multidimensional wisdom of crowds. In Neural Information Processing Systems (2010), vol.10, Zhou, D., Platt, J.C., Basu, S., and Mao, Y.: Learning from the wisdom of crowds by minimax entropy. In Neural Information Processing Systems (2012), / 26