The estimate of the importance of a feature is a piece of information that can help you operate on the best choices to determine the features to be used. Sometimes, you may need to understand better why a variable is important in predicting a certain outcome. Gradient Boosting Trees, by controlling the effect of all the other variables involved in the analysis, provide you with a clear point of view of the relationship of a variable with respect to the predicted results. Such information can provide you with more insights into causation dynamics than what you may have obtained by using a very effective EDA:

In: from sklearn.ensemble.partial_dependence import 
    plot_partial_dependence
    from sklearn.ensemble import GradientBoostingRegressor
    GBM = GradientBoostingRegressor(n_estimators=100, 
                                    random_state=101).fit(X, y)
    features = [5,12,(5,12)]
    fig, axis = plot_partial_dependence(GBM, X, features, 
                                        feature_names=feature_names)   

As an output, you get three plots, which constitute the partial plots of RM and LSTAT features:

The plot_partial_dependence class will automatically provide you with the visualization after you provide an analysis plan on your part. You need to present a list of indexes of the features to be plotted singularly, and the tuples of the indexes of those that you would like to plot on a heat map (the features are the axis, and the heat value corresponds to the outcome).

In the preceding example, both the average number of rooms and the percentage of the lower status population have been represented, thus displaying an expected behavior. Interestingly, the heat map, which explains how they together contribute to the value of the outcome, reveals that they do not interact in any particular way (it is single hill-climbing). However, it is also revealed that LSTAT is a strong delimiter of the resulting housing values when it is above 5.