Categorical Target Drift
TL;DR: The report explores the changes in the categorical target function (prediction).
- Performs a suitable statistical test to compare target (prediction) distribution.
- Plots the relations between each individual feature and the target (prediction)
The Target Drift report helps detect and explore changes in the target function and/or model predictions.
The Categorical Target Drift report is suitable for problem statements with the categorical target function: binary classification, multi-class classification, etc.
To run this report, you need to have input features, and target and/or prediction columns available.
You will need two datasets. The reference dataset serves as a benchmark. We analyze the change by comparing the current production data to the reference data.
You can potentially choose any two datasets for comparison. But keep in mind that only the reference dataset will be used as a basis for comparison.
We estimate the drift for the target (actual values) and predictions in the same manner. If both columns are passed to the dashboard, we build two sets of plots.
If only one of them (either target or predictions) is provided, we build one set of plots. If neither target nor predictions column is available, you will get an error.
To estimate the categorical target (prediction) drift, we compare the distribution of the target (prediction) in the two datasets. There is a default logic to choosing the appropriate statistical test, based on:
- the number of observations in the reference dataset
- the number of unique values in the target (n_unique)
For small data with <= 1000 observations in the reference dataset:
- For binary categorical target (n_unique <= 2), we use the proportion difference test for independent samples based on Z-score.
All tests use a 0.95 confidence level by default.
The report includes 2 components. All plots are interactive.
The report first shows the comparison of target (prediction) distributions in the current and reference dataset. The result of the statistical test and P-value are displayed in the title.
For a classification problem with three classes, it can look like this (an example of the extreme target drift with the appearance of a new class):
The report generates an interactive table with the visualizations of dependencies between the target and each feature.
If you click on any feature, you get a plot that shows the feature distribution for the different target labels.
These plots help analyze how feature values relate to the target labels and identify the differences between the datasets.
We recommend paying attention to the behavior of the most important features since significant changes might confuse the model and cause higher errors.
1. Before model retraining. Before feeding fresh data into the model, you might want to verify whether it even makes sense.
2. When you are debugging the model decay. If you observe a drop in performance, this report can help see what has changed.
3. When you are flying blind, and no ground truth is available. If you do not have immediate feedback, you can use this report to explore the changes in the model output and the relationship between the features and prediction. This can help anticipate data and concept drift.
If you choose to generate a JSON profile, it will contain the following information: