K-fold cross-validation
This method was invented and gained popularity in those days when the big date was not yet a problem, everyone had little data, but still needed to build reliable models. First thing we do is shuffle our dataset well, and then divide it randomly into several equal parts, say 10 (this is the k in k-fold). We hold out the first part as a test set, and on the remaining nine parts we train the model. The trained model is then assessed on the test set that did not participate in the training as usual. Next, we hold out the second of 10 parts, and train the model on the remaining nine (including those previously served as a test set). We validate the new model again on the part that did not participate in the training. We continue this process until each of the 10 parts is in the role of the test set. The final quality metrics are determined by the averaging metrics from each of the 10 tests:
In []: from sklearn.model_selection import cross_val_score scores = cross_val_score(tree_model, features, df.label, cv=10) np.mean(scores) Out[]: 0.88300000000000001 In []: plot = plt.bar(range(1,11), scores) Out[]:
From the preceding graph, you can see that the model's accuracy depends on how you split the data, but not much. By taking the average and variance of the cross-validation results, you can make a sense of how well your model can generalize on different data, and how stable it is.