Investigating the nature of interaction between two features and its effect on a model’s prediction might provide insightful knowledge about both the data and the model. As explained in Features-Interaction, interactions captured the residual effect on a model's prediction of two feature (or more) feature values, after considering the individual effect. The interaction strength can be conveniently displayed on a graph where the axis corresponds to the features values. The int

Working point is an easy knob to use in order to optimize the model performance to fit our needs. Working point (WP) is a value between 0 and 1 that represents the threshold we use in a binary classification model. Anything above this threshold will be classified as having label 1 (positive) and anything below this threshold will be classified as having label 0 (negative). A commonly used working point is 0.5, however the working point is problem-dependent and can be changed

The categorical features table summarizes multiple statistics about every categorical feature in the data, allowing it to be investigated all at once. The statistics presented in the table include: Uniques - Number of unique categories in each categorical feature in the data-set. N/A - Number of missing values in each categorical feature in the data-set. Least frequent- the least frequent category and it's appearance frequency in percentage for each categorical feature in the

Features often interact with each other under model predictions, and therefore interaction analysis can provide meaningful insights for this model. The predictions of a model can be decomposed into the sum of feature effects and the feature interaction effects. Here, Interactions are defined as the effect on the models' predictions that occurs by varying the features values after considering their individual contributions. For example, a house pricing model can use individual

The numerical features table summarizes multiple statistics about every numerical feature in the data, allowing it to be investigated all at once. The statistics presented in the table include: Mean - the mean value of each numerical feature in the data-set Stdev - the standard deviation of each numerical feature in the data-set Min - the minimal value of each numerical feature in the data-set Max - the maximal value of each numerical feature in the data-set N/A - Number of m

Shared distribution heatmap is a very useful and visual way to examine the spread and legitimacy of the data and understand how different attributes appear together in the data. For each pair of features we count the amount of samples that have every combination of their values. For categorical features this is done per each category. For continuous features, the values are first binned and the count is done per each bin. The x-axis and y-axis of the heatmap represent the cho

Feature histogram enables a single glance to view each features general distribution split between the labels categories. For continuous features, the x-axis of the histogram shows the feature values while the y-axis shows the density (local amount normalized by the total amount). The separate histograms depict the different labels of the samples (in the example bellow, the samples with the positive label vs. the negative label) For categorical features, the x-axis of the his

The feature leakage ROC plot is a graphic way to evaluate the potential leakage there might be between each feature and the label. Data leakage - "...The use of information in the model training process which would not be expected to be available at prediction time..." (source). One way a leakage can occur is by using features that are proxy of the label or partially give away the label (or the label itself!). To identify whether there is a leakage between one of the features

Counterfactuals provide an easy to understand local explanation to models prediction using a counter example providing a what-if intuition to the model's predictions. Counterfactuals provide an alternative way to look at the models decision making process. Instead of explaining what feature contributed most to this specific prediction, counterfactuals show the minimal changes needed in order to reverse the prediction. If all the changes written in each counterfactual will be

Anchors provide an intuitive set of rules that emphasis the features that are locally sufficient for the model to make a decision. These are set of rules in the form of if-and-then that are locally sufficient in order to ensure a certain prediction with a high degree of confidence. Meaning, that for instances in which the anchors hold, the prediction is almost always the same. Each anchor consists of the set of rules, the prediction, precision and coverage. For example: IF (f

The partial dependence plot (PDP) shows the marginal effect one or two features have on the predicted outcome of a machine learning model. The algorithm averages the model prediction results over all the samples while the feature of interest, Xs, is kept fixed at some value a. The result is the PDP value for feature Xs at point a. Doing this for all possible values of the feature Xs yields the PDP plot. This gives us a good visualization of how the model changes its predictio

A prediction histogram is an intuitive graphic way for visualizing model predictions and the separation it achieves for the actual labels. The models predictions (confidences) are binned and colored according to the true class label. For example, on the figure on the right, red bars are model confidences for negative labeled samples, and green bars are the model confidences for positive labeled samples. This can help users understand the trade-offs of choosing different thres

Force plots are a very intuitive and visual way to understand the influence of each feature on the models prediction for a specific instance of the data. The Force plot, much like Feature importance, is based on a game theory method - Shapley values. Unlike many other methods, the Shapley method has a strong theoretical basis. In short, the features are treated as players that can form a coalition and play games. The outcome of the game is the prediction of the model. The imp

Calibration plots give us a snapshot to understand how well a model's prediction correlate with a measure of probability that it is correct in its prediction. Many binary classification models output a value that is named confidence or probability. This allows for flexibility in the classification by using a working point that decides the final prediction. While confidence is often interpreted as the probability of the sample to belong to the predicted class or as model’s unc

Receiver Operating Characteristic Curve is a graphic way to evaluate the model performance and the trade-offs under different working points. A receiver operating characteristic curve, or ROC curve, is a plot that illustrates the diagnostic ability of a binary classification system as its discrimination threshold (or working point) is varied, where the predictions of the model are on a scale of zero to one. The x-axis is the False Positive Rate (FPR) or false alarm rate, and

Feature Importance plots are a great way to get a sense quickly how much a model is relying on each feature to make it predictions. Feature importance plot shows the global effect each feature has on the model. There are a few different techniques to calculate feature importance. Our implementation is based on a game theory method called Shapley values. Unlike many other methods, the Shapley method has a strong theoretical basis. In short, the features are treated as players