Getting started

This tutorial highlights the basics of the high-level API for ensemble classes, the model selection suite and features visualization.

============================ ================================================= Tutorials Content ============================ ================================================= ensemble-guide How to build, fit and predict with an ensemble model-selection-guide How to compare several estimators in one go visualization-guide Plotting functionality ============================ =================================================

The advanced high-level API tutorials <ensemble-tutorial> shows how to leverage advanced features such as probabilistic layers, feature propagation etc. For tutorials on low-level mechanics, see the mechanics guides <learner_tutorial>.

Preliminaries

We use the following setup throughout:

Ensemble guide

Building an ensemble ^^^^^^^^^^^^^^^^^^^^ Instantiating a fully specified ensemble is straightforward and requires three steps: first create the instance, second add the intermediate layers, and finally the meta estimator.

To check the performance of estimator in the layers, call the data attribute. The attribute can be wrapped in a :class:pandas.DataFrame, but prints in a tabular format as is.

To round off, let's see how the ensemble as a whole fared.

Multi-layer ensembles ^^^^^^^^^^^^^^^^^^^^^

With each call to the add method, another layer is added to the ensemble. Note that all ensembles are sequential in the order layers are added. For instance, in the above example, we could add a second layer as follows.

We now fit this ensemble in the same manner as before:

.. py:currentmodule:: mlens.model_selection

Model selection guide

The work horse class is the :class:Evaluator, which allows you to grid search several models in one go across several preprocessing pipelines. The evaluator class pre-fits transformers, thus avoiding fitting the same preprocessing pipelines on the same data repeatedly. Also, by fitting all models over all parameter draws in one operation, parallelization is maximized. The following example evaluates a Naive Bayes estimator and a K-Nearest-Neighbor estimator under three different preprocessing scenarios: no preprocessing, standard scaling, and subset selection. In the latter case, preprocessing is constituted by selecting a subset of features.

The scoring function ^^^^^^^^^^^^^^^^^^^^

.. currentmodule:: mlens.metrics

An important note is that the scoring function must be wrapped by :func:make_scorer, to ensure all scoring functions behave similarly regardless of whether they measure accuracy or errors. To wrap a function, simple do:

.. currentmodule:: mlens.model_selection

The make_scorer wrapper is a copy of the Scikit-learn's :func:sklearn.metrics.make_scorer, and you can import the Scikit-learn version as well. Note however that to pickle the :class:Evaluator, you must import make_scorer from mlens.

A simple evaluation ^^^^^^^^^^^^^^^^^^^

Before throwing preprocessing into the mix, let's see how to evaluate a set of estimator. First, we need a list of estimator and a dictionary of parameter distributions that maps to each estimator. The estimators should be put in a list, either as is or as a named tuple ((name, est)). If you don't name the estimator, the :class:Evaluator will automatically name the model as the class name in lower case. This name must be the key in the parameter dictionary.

We can now run an evaluation over these estimators and parameter distributions by calling the fit method.

The full history of the evaluation can be found in cv_results. To compare models with their best parameters, we can pass the results attribute to a :obj:pandas.DataFrame or print it as a table. We use m to denote mean values and s to denote standard deviation across folds for brevity. Note that the timed prediction is for the training set, for comparability with training time.

Preprocessing ^^^^^^^^^^^^^

Next, suppose we want to compare the models across a set of preprocessing pipelines. To do this, we first need to specify a dictionary of preprocessing pipelines to run through. Each entry in the dictionary should be a list of transformers to apply sequentially.

The fit methods determines automatically whether there is any preprocessing or any estimator jobs to run, so all we need to do is specify the arguments we want to be processed. If a previous preprocessing job was fitted, those pipelines are stored and will be used for subsequent estimator fits.

.. currentmodule:: mlens.preprocessing

This can be helpful if the preprocessing is time-consuming, for instance if the preprocessing pipeline is an ensemble itself. All ensembles implement a transform method that, in contrast to the predict method, regenerates the predictions made during the fitcall. More precisely, the transform method uses the estimators fitted with cross-validation to construct predictions, whereas the predict method uses the final estimators fitted on all data. This allows us use ensembles as preprocessing steps that mimicks how that ensemble would produce predictions for a subsequent meta learner or layer. Since fitting large ensembles is highly time-consuming, fixing the lower layers as preprocessing input is highly valuable for tuning the higher layers and / or the final meta learner. See the model-selection-tutorial tutorial for an example. To fit only the preprocessing pipelines, we omit any estimators in the fit call.

Model Selection across preprocessing pipelines ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

To evaluate the same set of estimators across all pipelines with the same parameter distributions, there is no need to take any heed of the preprocessing pipeline, just carry on as in the simple case:

You can also map different estimators to different preprocessing folds, and map different parameter distribution to each case. We will map two different parameter distributions

.. py:currentmodule:: mlens.visualization

Visualization guide

Explained variance plot

The :class:exp_var_plot function plots the explained variance from mapping a matrix X onto a smaller dimension using a user-supplied transformer, such as the Scikit-learn :class:sklearn.decomposition.PCA transformer for Principal Components Analysis.

Principal Components Analysis plot

The :class:pca_plot function plots a PCA analysis or similar if n_components is one of [1, 2, 3]. By passing a class labels, the plot shows how well separated different classes are.

Principal Components Comparison plot

The :class:pca_comp_plot function plots a matrix of PCA analyses, one for each combination of n_components=2, 3 and kernel='linear', 'rbf'.

Correlation matrix plot

The :class:corrmat function plots the lower triangle of a correlation matrix and is adapted the Seaborn_ correlation matrix.

Clustered correlation heatmap plot

The :class:clustered_corrmap function is similar to :class:corrmat, but differs in two respects. First, and most importantly, it uses a user supplied clustering estimator to cluster the correlation matrix on similar features, which can often help visualize whether there are blocks of highly correlated features. Secondly, it plots the full matrix (as opposed to the lower triangle).

We duplicate all features, note that the heatmap orders features as duplicate pairs, and thus fully pick up on this duplication.

Input-Output correlations

The :class:corr_X_y function gives a dashboard of pairwise correlations between the input data (X) and the labels to be predicted (y). If the number of features is large, it is advised to set the no_ticks parameter to True, to avoid rendering an illegible x-axis. Note that X must be a :class:pandas.DataFrame.