.. _multilogreg: Multinomial logistic regression ############################### The multinomial logistic regression model is a statistical model that models the probability of a record being associated with one category (out of N), based on one or more features. Here, all of the categories are independent of each other, and cannot be ordered in any meaningful way. For instance, one could try to predict the blood type of a person based on medical diagnostic values. For more information about multinomial logistic regression, see `Wikipedia `_. This guide explains all of the steps to fit a multinomial logistic regression model on a shared dataset, along with their options and considerations. This is done through an example with the Iris petal dataset (more about this dataset later). The logistic regression API in crandas follows the `scikit-learn API `_ wherever possible. It is often possible to run existing scikit-learn code with minimal modifications in crandas. That being said, there are still some differences, so it is a good idea to go through this guide once. Note that crandas specifically implements the ``multinomial`` model, not the ``ovr`` model that is also used by scikit-learn. Setup ===== Before moving on with the guide, it is necessary to import a few modules and functions from crandas first: .. code:: python import crandas as cd from crandas.crlearn.logistic_regression import LogisticRegression from crandas.crlearn.metrics import classification_accuracy from crandas.crlearn.utils import min_max_normalize from crandas.crlearn.metrics import confusion_matrix Reading the data ================ In this example, the dataset is read from a local CSV: .. code:: python tab = cd.read_csv("../../test/logreg_test_data/iris_scaled.csv") This imports the *Iris petal* dataset, which contains records of three different types of flower: ``iris-setosa``, ``iris-versicolor`` and ``iris-virginica``, along with measured features of these flowers. This guide demonstrates how multinomial logistic regression can be used to predict the class of flower, based on these features. The dataset looks as following: >>> print(tab.open().head()) class sepal length sepal width petal length petal width 0 0 0.222222 0.625000 0.067797 0.041667 1 0 0.166667 0.416667 0.067797 0.041667 2 0 0.111111 0.500000 0.050847 0.041667 3 0 0.083333 0.458333 0.084745 0.041667 4 0 0.194445 0.666667 0.067797 0.041667 Note that the features are all numerical (the ``sepal length``, ``sepal width``, ``petal length``, ``petal width`` columns). .. note:: See `Kaggle `_ for more information about this dataset. Preparing the data ================== Getting rid of null values --------------------------- The logistic regression can only be executed on a :class:`.CDataFrame` without null values (specifically, without nullable columns). If the dataset contains any missing values, one can get rid of all rows with null values using :meth:`.CDataFrame.dropna`. .. code:: python tab = tab.dropna() An alternative to deleting the rows with null values is performing `data imputation `_ using :meth:`.CSeries.fillna`. However, this might introduce bias and is not recommended in the general case. Normalizing ----------- If the dataset contains any numerical values (e.g. ``sepal length`` in this example), these first need to be normalized to values between 0 and 1. The way in which this is commonly done is through `Min-Max-Normalization `_. .. code:: python tab_normalized = min_max_normalize(tab, columns=['sepal length', 'sepal width', 'petal length', 'petal width']) Here, ``columns`` can be used to specify which columns need to be normalized. The remaining columns will remain untouched. .. attention:: It is **essential** to normalize your numerical features to within [0, 1] **before** you fit the model. Otherwise, fitting will **not** work correctly, and will return erroneous results. Negative values are also **not** allowed. Splitting into predictors, response ----------------------------------- First, split the predictor variables from the response variable: .. code:: python X = tab_normalized[['sepal length', 'sepal width', 'petal length', 'petal width']] y = tab_normalized[['class']] Creating the model ================== The logistic regression functionality in crandas is made accessible through the :class:`.LogisticRegression` class, which can be used to fit the model and make predictions. The model can be created using: .. code:: python model = LogisticRegression(solver='lbfgs', multi_class="multinomial", n_classes=3) Here, the ``multi_class`` argument specifies the type of regression to be performed, in this case ``multinomial``. The ``n_classes`` argument specifies the number of classes in the dataset. .. note:: The ``solver`` argument indicates which numerical solver the model should use to fit the model. Currently, the available options are: - ``lbfgs`` (which stands for `Limited-memory BFGS `_) - ``gd`` (which stands for `Gradient Descent `_) The ``lbfgs`` solver gives better results and fits the model faster. As such, there is normally no reason to deviate from it. .. attention:: It is **required** to specify the number of classes in the dataset. Unlike in scikit-learn, the crandas model does not detect this automatically due to the dataset being secret-shared. Fitting the model ================= Now that the data has been prepared and the model has been created, the model can be fitted to the training set: .. code:: python model.fit(X, y, max_iter=40) Here, the ``max_iter`` argument specifies how many iterations the numerical solver should perform to fit the model. The default of 10 is sufficient in some cases but sometimes it is necessary to increase this number in order for the model to fully converge. In this case, we need 40 iterations for the Iris dataset. .. note:: Fitting a logistic regression model in crandas can take quite some time, depending on the number of records in the dataset, the number of features, and the number of iterations that you specify. The fitted model parameters can now be accessed as following: .. code:: python beta = model.get_beta() This returns a matrix of size ``num_features + 1`` by ``n_classes``. The first column contains all of the intercept terms, while the remaining columns correspond to each of the features. Each row contains all of the terms associated with that specific class. Predicting ========== Now that the model has been fitted, it can be used to make predictions. We distinguish two different types in crandas: - probabilities: the model can predict the probability of each class being associated with the record - classes: the model can predict the class with the highest likelihood Probabilities ------------- First, to predict the probabilities corresponding to each record of the test dataset: .. code:: python y_pred_probabilities = model.predict_proba(X) This returns a table with three columns (one for each class), containing the probability for each class. These sum up to one. Classes ------- Alternatively, if you are interested in making actual class predictions rather than the probabilities, you can directly predict the classes through: .. code:: python y_pred_classes = model.predict(X) .. tip:: Predicting *classes* is a quick operation, that takes *significantly* less time than predicting probabilities. Assessing prediction quality ============================ After fitting the model, it is important to assess the quality of the model and its predictions. crandas provides a couple of methods for doing this, namely: - Classification Accuracy - `Confusion Matrix `_. Accuracy -------- To compute the accuracy of the (class) predictions, you can use: .. code:: python accuracy = classification_accuracy(y, y_pred_classes, n_classes=3) print("Classification Accuracy:", accuracy.open()) .. attention:: It is **required** to specify the number of classes in the dataset. The function does not detect this automatically, due to the dataset being secret-shared. Confusion Matrix ---------------- The confusion matrix visualizes the relation between the predicted classes and the actual classes. The Y-axis represents the true class, while the X-axis represents the class predicted by the model. To compute the confusion matrix obliviously, you can use: .. code:: python matrix = confusion_matrix(y, y_pred_classes, n_classes=3) matrix.open() [[50, 0, 0], [0, 44, 6], [0, 4, 46]]