When a model finishes training, click on the model to see the results.
For all models trained in Python (e.g., scikit-learn, keras, custom models), Dataiku can compute and display partial dependence plots. These plots can help you to understand the relationship between a feature and the target.
Partial dependence plots are a post-hoc analysis that can be computed after a model is built. Since they can be expensive to produce when there are many features, you need to explicitly request these plots on the Partial dependence page of the output.
The plot shows the dependence of the target on a single selected feature. The x axis represents the value of the feature, while the y axis displays the partial dependence. A negative partial dependence represents a negative relationship between that feature value and the target, while a positive partial dependence represents a positive relationship.
For example, in the figure below, there is a negative relationship between age_first_order and the target for people under 40. The relationship appears to be roughly parabolic with a minimum somewhere between ages 25 and 30. After age 40, the relationship is slowly increasing, until a precipitous dropoff in late age.
The plot also displays the distribution of the feature, so that you can determine whether there is sufficient data to interpret the relationship between the feature and target.
For regression and binary classification models trained in Python (e.g., scikit-learn, keras, custom models), Dataiku can compute and display subpopulation analyses. These can help you to assess if your model behaves identically across subpopulations; for example, for a bias or fairness study.
Subpopulation analyses are a post-hoc analysis that can be computed after a model is built. Since they can be expensive to produce when there are many features, you need to explicitly request these plots on the Subpopulation analysis page of the output.
The primary analysis is a table of various statistics that you can compare across subpopulations, as defined by the values of the selected column. You need to establish, for your use case, what constitutes “fair”.
For example, the table below shows a subpopulation analysis for the column gender. The model-predicted probabilities for male and female are close, but not quite identical. Depending upon the use case, we may decide that this difference is not significant enough to warrant further investigation.
By clicking on a row in the table, we can display more detailed statistics related to the subpopulation represented by that row.
For example, the figure below shows the expanded display for rows whose value of gender is missing. The number of rows that are Actually True is lower for this subpopulation than for males or females. By comparing the % of actual classes view for this subpopulation versus the overall population, it looks like the model does a significantly better job of predicting actually true rows with missing gender than otherwise.
The density chart suggests this may be because the class True density curve for missing gender has a single mode around 0.75 probability. By contrast, the density curve for the overall population has a second mode just below 0.5 probability.
What’s not clear, and would require further investigation, is whether this is a “real” difference, or an artifact of the relatively small subpopulation of missing gender.