Add counterfactual analysis for regression models (What-If Tool) #2647

grovina · 2019-09-12T09:16:12Z

Motivation for features / changes

Add the ability to show nearest counterfactual example in regression models.

Technical description of changes

The feature and its use are analogous to the existing counterfactual analysis available in non-regression models.

There is one fundamental difference in the way we compare inference classes. In binary classification, the counterfactual examples are the ones belonging to the alternative class. In regression, we will define them as the ones whose inferred values are farther from the original example than a given delta, which will be defined by the user.

For example, if delta is set to 2 and the inferred value of the selected example is 7, then we'll be looking for most similar example with an inferred value outside of the (5, 9) interval.

Screenshots of UI changes

For regression only:

Detailed steps to verify changes work correctly (as executed by you)

Open binary classification problem and check that nothing's changed. (multi_demo)
Open regression classification problem (age_demo) and wait for data to load, then:
- select an example
- enable "Show nearest counterfactual datapoint"
- play around with Delta and distance
- repeat for some other datapoints

Plotting Distance to datapoint vs Inference value also helps to confirm that the counterfactual found is reasonable.

Alternate designs / implementations considered

Some other ways of defining the nearest counterfactual example:

Define separate upper and lower deltas for the factual interval.
Within a given distance in the feature space, find the datapoint with highest difference in inference value.
Finding the datapoint with the highest ratio between difference in inference value and distance in features space.

These alternative ways could be added in the future.

cc @jameswex @tolga-b

When computing closest counterfactuals, we need to evaluate if two datapoints are in the same class or not. For regression, this is more than ordinary comparison. Let's move this evaluation to a separate method to be able to keep the logics clear.

For regression, we add a way to specify the distance in inferred value above which the points can be considered counterfactual. Also adding a small explanation in the information dialog.

Set the maximum slider value for the datapoint when calculating nearest counterfactuals. Since the counterfactual threshold is based on the distance between inferred values, we calculate the maximum value distance between the selected point and the others.

Ignore the example itself when looking for counterfactuals, since a 0 threshold in regression would mean that the closest counterfactual is always the example itself, which gives no information. By skipping the example, the 0 threshold means looking for the closest [different] example, regardless of the difference in inferred values.

googlebot · 2019-09-12T10:40:28Z

We found a Contributor License Agreement for you (the sender of this pull request), but were unable to find agreements for all the commit author(s) or Co-authors. If you authored these, maybe you used a different email address in the git commits than was used to sign the CLA (login here to double check)? If these were authored by someone else, then they will need to sign a CLA as well, and confirm that they're okay with these being contributed to Google.
In order to pass this check, please resolve this problem and then comment @googlebot I fixed it.. If the bot doesn't comment, it means it doesn't think anything has changed.

ℹ️ Googlers: Go here for more info.

googlebot · 2019-09-12T10:42:03Z

CLAs look good, thanks!

ℹ️ Googlers: Go here for more info.

jameswex

Thanks for getting this out so quickly! Some small comments.

...ractive_inference/tf_interactive_inference_dashboard/tf-interactive-inference-dashboard.html

Instead of 0, use the standard deviation of the inferred values as the initial value for the counterfactual delta.

Fixed vertical alignment.

jameswex · 2019-09-13T16:17:02Z

@stephanwlee please review.

stephanwlee · 2019-09-13T16:30:08Z

...ractive_inference/tf_interactive_inference_dashboard/tf-interactive-inference-dashboard.html

+        adjustMaxCounterfactualValueDist_: function(selected, valueStr) {
+          this.maxCounterfactualValueDist = Math.max(
+            this.distanceStats_[valueStr].max -
+              this.visdata[selected][valueStr],


Do not know about the nature of distanceStats and visData: can all of these values be negative? Can visData be lower than both max and min of the distanceStats? Can this be doing Math.max on, for example, Math.max(-2, -3)?

distanceStats_ is based on the visData data, so that:
distanceStats_[valueStr].min <= visdata[selected][valueStr] <= distanceStats_[valueStr].max

This Math.max op then receives only non-negative values.

...ractive_inference/tf_interactive_inference_dashboard/tf-interactive-inference-dashboard.html

Addressing PR conversations.

...ractive_inference/tf_interactive_inference_dashboard/tf-interactive-inference-dashboard.html

One-way binding is enough.

`yarn fix-lint`

grovina added 4 commits September 11, 2019 17:51

Add slider to control counterfactual threshold

d54dd47

For regression, we add a way to specify the distance in inferred value above which the points can be considered counterfactual. Also adding a small explanation in the information dialog.

googlebot added the cla: yes label Sep 12, 2019

googlebot added cla: no and removed cla: yes labels Sep 12, 2019

Fix lint warnings

e5ed88e

grovina force-pushed the grovina-wit-reg-counterfactual branch from e54b2df to e5ed88e Compare September 12, 2019 10:41

googlebot added cla: yes and removed cla: no labels Sep 12, 2019

jameswex reviewed Sep 12, 2019

View reviewed changes

grovina added 2 commits September 13, 2019 14:58

Initialize counterfactual delta to std. deviation

7de7f02

Instead of 0, use the standard deviation of the inferred values as the initial value for the counterfactual delta.

Adjust display of counterfactual parameters

6d1b1cb

Fixed vertical alignment.

jameswex requested a review from stephanwlee September 13, 2019 16:16

jameswex approved these changes Sep 13, 2019

View reviewed changes

stephanwlee requested changes Sep 13, 2019

View reviewed changes

Adjust naming, style and comments.

009fd5a

Addressing PR conversations.

grovina requested a review from stephanwlee September 17, 2019 10:49

stephanwlee approved these changes Sep 17, 2019

View reviewed changes

...ractive_inference/tf_interactive_inference_dashboard/tf-interactive-inference-dashboard.html Outdated Show resolved Hide resolved

...ractive_inference/tf_interactive_inference_dashboard/tf-interactive-inference-dashboard.html Show resolved Hide resolved

grovina added 3 commits September 17, 2019 16:48

Change max value binding in counterfactual delta

2a3c3cf

One-way binding is enough.

Merge branch 'master' into grovina-wit-reg-counterfactual

a9dcb6c

Fix lint

8ad428c

`yarn fix-lint`

jameswex merged commit b7b68b1 into tensorflow:master Sep 20, 2019

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add counterfactual analysis for regression models (What-If Tool) #2647

Add counterfactual analysis for regression models (What-If Tool) #2647

grovina commented Sep 12, 2019 •

edited

Loading

googlebot commented Sep 12, 2019

googlebot commented Sep 12, 2019

jameswex left a comment

jameswex commented Sep 13, 2019

stephanwlee Sep 13, 2019

grovina Sep 16, 2019

Add counterfactual analysis for regression models (What-If Tool) #2647

Add counterfactual analysis for regression models (What-If Tool) #2647

Conversation

grovina commented Sep 12, 2019 • edited Loading

googlebot commented Sep 12, 2019

googlebot commented Sep 12, 2019

jameswex left a comment

Choose a reason for hiding this comment

jameswex commented Sep 13, 2019

stephanwlee Sep 13, 2019

Choose a reason for hiding this comment

grovina Sep 16, 2019

Choose a reason for hiding this comment

grovina commented Sep 12, 2019 •

edited

Loading