Data drift impact in a machine learning model
Techniques for determining a drift impact score in a machine learning model are disclosed. The techniques can include: obtaining a reference distribution of a machine learning model; obtaining a current distribution of the machine learning model; determining a statistical distance based on the reference distribution and the current distribution; determining a local feature importance parameter for each feature associated with a prediction made by the machine learning model; determining a cohort feature importance parameter for a cohort of multiple features based on the local feature importance parameter of each feature in the cohort; and determining a drift impact score for the cohort based on the statistical distance and the cohort feature importance parameter.
1. A computer-implemented method for detecting a bias in a machine learning model, the method comprising:
obtaining a reference distribution of a machine learning model;
obtaining a current distribution of the machine learning model;
binning the reference distribution and the current distribution into a plurality of corresponding bins;
determining a statistical distance based on the binned reference distribution and the binned current distribution;
determining a local feature importance parameter for each feature associated with a prediction made by the machine learning model, wherein the local feature importance parameter of a feature indicates a difference between the model's prediction and an expected prediction;
determining a cohort feature importance parameter for a cohort of multiple features based on averaging values of the local feature importance parameter of each feature in the cohort;
determining a drift impact score for the cohort based on a multiplication of the statistical distance and the cohort feature importance parameter; and
detecting a bias in the machine learning model based on the drift impact score.
2. The method of claim 1 , wherein the determining of the statistical distance is based on a population stability index metric.
3. The method of claim 1 , wherein the determining of the statistical distance is based on a Kullback-Leibler (KL) divergence metric.
4. The method of claim 1 , wherein the determining of the statistical distance is based on a Jensen-Shannon (JS) divergence metric.
5. The method of claim 1 , wherein the determining of the statistical distance is based on an Earth Mover's distance (EMD) metric.
6. The method of claim 1 , wherein the reference distribution is a distribution across a fixed time window or a moving time window.
7. The method of claim 1 , wherein the reference distribution is from a training environment or a production environment.
8. A system for detecting a bias in a machine learning model, the system comprising a processor and an associated memory, the processor being configured for:
obtaining a reference distribution of a machine learning model;
obtaining a current distribution of the machine learning model;
binning the reference distribution and the current distribution into a plurality of corresponding bins;
determining a statistical distance based on the binned reference distribution and the binned current distribution;
determining a local feature importance parameter for each feature associated with a prediction made by the machine learning model, wherein the local feature importance parameter of a feature indicates a difference between the model's prediction and an expected prediction;
determining a cohort feature importance parameter for a cohort of multiple features based on averaging values of the local feature importance parameter of each feature in the cohort;
determining a drift impact score for the cohort based on a multiplication of the statistical distance and the cohort feature importance parameter; and
detecting a bias in the machine learning model based on the drift impact score.
9. The system of claim 8 , wherein the determining of the statistical distance is based on a population stability index metric.
10. The system of claim 8 , wherein the determining of the statistical distance is based on a Kullback-Leibler (KL) divergence metric.
11. The system of claim 8 , wherein the determining of the statistical distance is based on a Jensen-Shannon (JS) divergence metric.
12. The system of claim 8 , wherein the determining of the statistical distance is based on an Earth Mover's distance (EMD) metric.
13. The system of claim 8 , wherein the reference distribution is a distribution across a fixed time window or a moving time window.
14. The system of claim 8 , wherein the reference distribution is from a training environment or a production environment.