Collect feedback of your ML model
How to evaluate and get collaborative feedback on your Machine Learning model
To review your model, you can either
- Inspect it by yourself
- Invite peer ML engineer colleagues
- Invite business users who might have business knowledge of your model
- Clicking on the Models tab
- Selecting the model version that you want to inspect by clicking on the Inspect button
- Selecting the dataset that you want to use for the inspection
- Selecting the actual target variable from your dataset. This is the true value of your target variable. If there is no actual target variable in your dataset, just select the blank cell
Choose the dataset for the inspection session
You can play with the model by changing examples, feature values, or generating local explanations.
Use the Next and Previous buttons of the Dataset Explorer to explore your dataset from one example to another.
The Data Explorer
You can also select the random button of the Data Explorer to explore randomly your dataset.
You can also artificially change the feature values for a given example.
Changing the credit history feature
By checking the change of the result and its explanation, you'll gain some insights into the local behavior of your model.
Every time you change examples or feature values, a local explanation is generated and provides the contribution of each feature to the final result.
Local Shap values for a credit scoring model
The explanation is given as a bar plot giving the absolute value of the Shapley value for each feature: the longer the bar of a given feature, the greater its contribution to the final result of the model.
You can also generate a text explanation to have a look at the contribution of each word on a text field. To do that, click on the tab TEXT in the explanation section,
- select a text feature you want to explain
- select the classification label for which you want to compute the word contribution
Explanation for a textual feature
It then returns the LIME values of each word in your textual feature: green highlighting means a positive contribution to the selected classification label, while red highlighting means a negative contribution.
Once you played with the model, you might have things to say about the behavior of the model. The Feedback button can help you.
Here are some feedback examples:
- Switching from 2012 to 2013 makes the prediction change a lot. This is weird!
- When I increase the credit amount, the default probability is decreasing. I would expect the reverse though.
- This example is a very common case. It should have been treated differently!
- This feature has a very big importance to the final decision. I don't understand why...
- This feature value has no business sense. It should be modified.
- When I replace the word "big" with "tall", the prediction is changing a Iot. I would expect the prediction to be invariant though…
To send your feedback, you have two possibilities.
To provide feedback by feature, click on the green feedback button on the right of the feature field
Provide feedback by feature
If your feedback is related to the whole example, you may send general feedback by clicking on the Feedback bottom at the bottom right
Provide general feedback
To analyze feedback, select the Feedback tab. It lists all the feedback that were collected in your project.
List of collected feedback
To better target some feedback, you can filter them depending on:
- Model version
- Dataset
- Type of feedback: value perturbation, by feature or general
You can also aggregate the feedback using the "Group by feature" checkbox.
Grouping feedback by feature enables you to check which feature collected the most feedback. Analyzing your feedback allows you to schedule your feature engineering tasks for your next model version.
You can click on particular feedback to re-inspect the example and discuss it with the feedback provider.
Discussion can be important to clarify the feedback context and find the best action to correct the model.
Last modified 2mo ago