IdeasonMachineLearningInterpretability

PatrickHall,WenPhan,SriSatish AmbatiandtheH2O.ai team

BigIdeas

Learning from data …

Adaptedfrom:LearningfromData.https://work.caltech.edu/textbook.html

EXPLAINHYPOTHESIS

h ≈ g, βj g(x(i)j), g(x(i)

(-j))

(explainpredictionswithreasoncodes)

Learning from data …transparently.

Adaptedfrom:LearningfromData.https://work.caltech.edu/textbook.html

Increasingfairness,accountability,andtrustbydecreasingunwantedsociologicalbiases

Source:http://money.cnn.com/,AppleComputers

Increasingtrustbyquantifyingpredictionvariance

Source:http://www.vias.org/tmdatanaleng/

7

AframeworkforinterpretabilityComplexity of learned functions:• Linear, monotonic• Nonlinear, monotonic• Nonlinear, non-monotonic

(~ Number of parameters/VC dimension)

Enhancing trust and understanding: the mechanisms and results of an interpretable model should be both transparent AND dependable.

Understanding ~ transparency Trust ~ fairness and accountability

Scope of interpretability:Global vs. local

Application domain:Model-agnostic vs. model-specific

BigChallenges

LinearModelsStrongmodellocality

Usuallystablemodelsandexplanations

MachineLearningWeakmodellocality

Sometimesunstablemodelsandexplanations(a.k.a.TheMultiplicityofGoodModels)

Age

Numbe

rofP

urchases

Lostprofits.

Wastedmarketing.

“Foraoneunitincreaseinage,thenumberofpurchasesincreasesby0.8 onaverage.”

𝑔 𝑥 = 0.8𝑥

LinearModels

MachineLearning

Exact explanationsforapproximatemodels.

Approximate explanationsforexactmodels.

Age

“Slopebeginstodecreasehere.Acttooptimizesavings.”

“Slopebeginstoincreaseheresharply.Acttooptimizeprofits.”

Numbe

rofP

urchase

𝑔 𝑥 ≈ 𝑓(𝑥)

AFewofOurFavoriteThings

Partialdependenceplots

Source:http://statweb.stanford.edu/~tibs/ElemStatLearn/printings/ESLII_print10.pdf

HomeValue ~ MedInc + AveOccup + HouseAge + AveRooms

Surrogatemodels

Localinterpretable model-agnosticexplanations

Source:https://www.oreilly.com/learning/introduction-to-local-interpretable-model-agnostic-explanations-lime

VariableimportancemeasuresGlobalvariableimportanceindicatestheimpactofavariableonthemodelfortheentiretrainingdataset.

Localvariableimportancecanindicatetheimpactofavariableforeachdecisionamodelmakes– similartoreasoncodes.

Resources

MachineLearningInterpretabilitywithH2ODriverlessAIhttps://www.h2o.ai/wp-content/uploads/2017/09/MLI.pdf(ORcomebythebooth!!)

IdeasonInterpretingMachineLearninghttps://www.oreilly.com/ideas/ideas-on-interpreting-machine-learning

FAT/MLhttp://www.fatml.org/

Questions?