Get Another Label? Improving Data Quality and Data Mining Using Multiple, Noisy Labelers Victor Sheng Foster Provost Panos Ipeirotis New York University Stern School Outsourcing KDD preprocessing Traditionally, data mining teams have invested substantial internal resources in data formulation, information extraction, cleaning, and other preprocessing – Raghu from his Innovation Lecture “the best you can expect are noisy labels” 2 Now, we can outsource preprocessing tasks, such as labeling, feature extraction, verifying information extraction, etc. – using Mechanical Turk, Rent-a-Coder, etc. – quality may be lower than expert labeling (much?) – but low costs can allow massive scale The ideas may apply also to focusing user-generated tagging, crowdsourcing, etc. ESP Game (by Luis von Ahn) 3 Other “free” labeling schemes Open Mind initiative (www.openmind.org) Other gwap games – – – Tag a Tune Verbosity (tag words) Matchin (image ranking) Web 2.0 systems? – Can/should tagging be directed? Noisy labels can be problematic Many tasks rely on high-quality labels for objects: – – – – – 5 learning predictive models searching for relevant information finding duplicate database records image recognition/labeling song categorization Noisy labels can lead to degraded task performance Here, labels are values for target variable Quality and Classification Performance Labeling quality increases classification quality increases P = 1.0 100 P = 0.8 Accuracy 90 80 P = 0.6 70 60 P = 0.5 50 6 80 10 0 12 0 14 0 16 0 18 0 20 0 22 0 24 0 26 0 28 0 30 0 60 40 20 1 40 Number of examples (Mushroom) Summary of results Repeated labeling can improve data quality and model quality (but not always) When labels are noisy, repeated labeling can be preferable to single labeling even when labels aren’t particularly cheap When labels are relatively cheap, repeated labeling can do much better (omitted) Round-robin repeated labeling does well Selective repeated labeling improves substantially Majority Voting and Label Quality Ask multiple labelers, keep majority label as “true” label Quality is probability of being correct 1 P is probability of individual labeler being correct Integrated quality 0.9 P=1.0 P=0.9 0.8 P=0.8 0.7 P=0.7 0.6 P=0.6 0.5 P=0.5 0.4 P=0.4 0.3 0.2 9 1 3 5 7 9 Number of labelers 11 13 Tradeoffs for Modeling Get more labels Improve label quality Improve classification Get more examples Improve classification P = 1.0 100 P = 0.8 Accuracy 90 80 P = 0.6 70 60 P = 0.5 50 10 80 10 0 12 0 14 0 16 0 18 0 20 0 22 0 24 0 26 0 28 0 30 0 60 40 20 1 40 Number of examples (Mushroom) Basic Labeling Strategies Single Labeling – Get as many data points as possible – one label each Round-robin Repeated Labeling – Fixed Round Robin (FRR) – 11 keep labeling the same set of points Generalized Round Robin (GRR) repeatedly-label data points, giving next label to point with fewest so far Fixed Round Robin vs. Single Labeling FRR (100 examples) SL p= 0.6, labeling quality #examples =100 12 With high noise, repeated labeling better than single labeling Fixed Round Robin vs. Single Labeling Single FRR (50 examples) p= 0.8, labeling quality #examples =50 13 With low noise, more (single labeled) examples better Gen. Round Robin vs. Single Labeling 100 P=0.6, k=5 Accuracy 90 P: labeling quality k: #labels GRR 80 SL 70 60 Repeated labeling is better than single labeling 50 80 1680 3280 4880 6480 Data acquisition cost (mushroom, p=0.6) Tradeoffs for Modeling Get more labels Improve label quality Improve classification Get more examples Improve classification P = 1.0 100 P = 0.8 Accuracy 90 80 P = 0.6 70 60 P = 0.5 50 15 80 10 0 12 0 14 0 16 0 18 0 20 0 22 0 24 0 26 0 28 0 30 0 60 40 20 1 40 Number of examples (Mushroom) Selective Repeated-Labeling We have seen: – With enough examples and noisy labels, getting multiple labels is better than single-labeling – When we consider costly preprocessing, the benefit is magnified (omitted -- see paper) Can we do better than the basic strategies? Key observation: we have additional information to guide selection of data for repeated labeling – 16 the current multiset of labels Example: {+,-,+,+,-,+} vs. {+,+,+,+} Natural Candidate: Entropy Entropy is a natural measure of label uncertainty: | S | | S | | S | | S | E(S ) log2 log2 |S| |S| |S| |S| E({+,+,+,+,+,+})=0 E({+,-, +,-, +,- })=1 | S |: positive | S |: negative Strategy: Get more labels for examples with highentropy label multisets 17 What Not to Do: Use Entropy 0.95 Improves at first, hurts in long run Labeling quality 0.9 0.85 0.8 0.75 0.7 ENT ROPY GRR 0.65 0.6 0 400 800 1200 1600 Number of labels (waveform, p=0.6) 18 2000 Why not Entropy In the presence of noise, entropy will be high even with many labels Entropy is scale invariant (3+ , 2-) has same entropy as (600+ , 400-) 19 Estimating Label Uncertainty (LU) Observe +’s and –’s and compute Pr{+|obs} and Pr{-|obs} Label uncertainty = tail of beta distribution Beta probability density function SLU 20 0.0 0.5 1.0 Label Uncertainty 21 p=0.7 5 labels (3+, 2-) Entropy ~ 0.97 CDFb=0.34 Label Uncertainty 22 p=0.7 10 labels (7+, 3-) Entropy ~ 0.88 CDFb=0.11 Label Uncertainty 23 p=0.7 20 labels (14+, 6-) Entropy ~ 0.88 CDFb=0.04 Labeling quality Label Uncertainty vs. Round Robin 1 0.9 0.8 GRR LU 0.7 0.6 0 24 400 800 1200 1600 2000 Number of labels (waveform, p=0.6) similar results across a dozen data sets Recall: Gen. Round Robin vs. Single Labeling 100 P=0.6, k=5 Accuracy 90 P: labeling quality k: #labels GRR 80 SL 70 60 Multi-labeling is better than single labeling 50 80 1680 3280 4880 6480 Data acquisition cost (mushroom, p=0.6) Labeling quality Label Uncertainty vs. Round Robin 1 0.9 0.8 GRR LU 0.7 0.6 0 26 400 800 1200 1600 2000 Number of labels (waveform, p=0.6) similar results across a dozen data sets Another strategy: Model Uncertainty (MU) Learning a model of the data provides an alternative source of information about label certainty Model uncertainty: get more labels for instances that cannot be modeled well ? - -- + + + ++ - - - - - + + -- + ++ Intuition? + - -- - + + + + + - - ---– for data quality, low-certainty “regions” may + + -+ be due to incorrect labeling of corresponding -+instances 27 – for modeling: why improve training data quality if model already is certain there? ? Yet another strategy: Label & Model Uncertainty (LMU) Label and model uncertainty (LMU): avoid examples where either strategy is certain S LMU 28 S LU S MU Comparison Model Uncertainty Label & Model alone also improves Label Uncertainty quality Uncertainty 1 Labeling quality 0.95 0.9 0.85 0.8 GRR 0.75 GRR MU LU LMU 0.7 0.65 0.6 0 29 400 800 1200 Number of labels (waveform, p=0.6) 1600 2000 Across 12 domains, LMU is always better than GRR. LMU is statistically significantly better than LU and MU. Comparison: Model Quality Label & Model Uncertainty 90 Accuracy 85 80 75 GRR MU LU LMU 70 65 60 30 0 400 800 1200 Number of labels (spambase, p=0.6) 1600 2000 Summary of results Micro-task outsourcing (e.g., MTurk, RentaCoder ESP game) has changed the landscape for data formulation Repeated labeling can improve data quality and model quality (but not always) When labels are noisy, repeated labeling can be preferable to single labeling even when labels aren’t particularly cheap When labels are relatively cheap, repeated labeling can do much better (omitted) Round-robin repeated labeling can do well Selective repeated labeling improves substantially Opens up many new directions… 32 Strategies using “learning-curve gradient” Estimating the quality of each labeler Example-conditional quality Increased compensation vs. labeler quality Multiple “real” labels Truly “soft” labels Selective repeated tagging Thanks! Q & A?