Data Science and Goodhart's Law

Preview:

Citation preview

Data Science and Goodhart’s Law

Kyle PolichData Science, Inc.

2

Goodhart’s Law

When a measure becomes a target, it ceases to be a good measure

3

Sales Rep Compensation Example

• Base pay + variable commission• For monthly <50k, commission = 3%• For monthly 50-99k, commission = 5%• For monthly 100k+, commission = 7%

4

Some Examples

Spam filtering arms race Search engine ranking Clearing cookies to get better airline prices Keep account open to manipulate FICO score Retail discounting/couponing strategies Bidding in AdTech marketplaces

5

Measuring with Cross ValidationCross Validation• You should be doing this anyway!• Set production performance expectation• Measure post deployment• Total deviation =

deviation due to overfit + deviation due to incomplete training+ deviation due to Goodhart’s Law

6

Measuring via Homogeneity Assumption

Can you train a model to accurately predict the date at which the observation was created?

7

Measuring Drift

8

Measuring DriftTypical failure from a web application release

9

Measuring DriftPossible failure from a web application release

10

Dealing with it

• Detection is key• Experimentation is required• Agile methods for model

deployment

11

Causal Impact• An approach to

estimating the causal effect of a designed intervention on a time series.

• Predicts counterfactual (how response likely would have evolved absent the intervention)

12

Self Fulfilling Prophecies

• Beware!• Case study: lead qualification

– Try to predict leads that will close– Relearn the bias of your training

13

Fast Iterations

• Outside normal SWLC release cycle– State updates– Parameter tuning

• Run experiments

14

Explanatory power

• Goodhart’s law will often manifest on only a subset of (possibly significant) instances.

• Model interpretability for effected instances is key

15

Interpretable Models

16

Interpretable Models

17

Why Should I Trust You?Explaining the Predictions of Any Classifier

Ribeiro, Singh, Guestrin

Model Interpretability

18

Summary• Goodhart’s law: When a measure becomes a target, it

ceases to be a good measure• As a data scientist, if your work is meaningful, you will

encounter it• Try to measure it in the data• Work on explanatory models to mitigate• Don’t let the average case blind you

19

DataScience

facebook.com/datascience

@DataSkeptic@datascienceinc

linkedin.com/company/datascience-inc

(310) 579 - 6200

Recommended