Machine learning and deep learning models, like those in Keras, require all input and output variables to be numeric.

This means that if your data contains categorical data, you must encode it to numbers before you can fit and evaluate a model.

The two most popular techniques are an integer encoding and a one hot encoding, although a newer technique called learned embedding may provide a useful middle ground between these two methods.

In this tutorial, you will discover how to encode categorical data when developing neural network models in Keras.

After completing this tutorial, you will know:Let’s get started.

How to Encode Categorical Data for Deep Learning in KerasPhoto by Ken Dixon, some rights reserved.

This tutorial is divided into five parts; they are:A categorical variable is a variable whose values take on the value of labels.

For example, the variable may be “color” and may take on the values “red,” “green,” and “blue.

”Sometimes, the categorical data may have an ordered relationship between the categories, such as “first,” “second,” and “third.

” This type of categorical data is referred to as ordinal and the additional ordering information can be useful.

Machine learning algorithms and deep learning neural networks require that input and output variables are numbers.

This means that categorical data must be encoded to numbers before we can use it to fit and evaluate a model.

There are many ways to encode categorical variables for modeling, although the three most common are as follows:We will take a closer look at how to encode categorical data for training a deep learning neural network in Keras using each one of these methods.

As the basis of this tutorial, we will use the so-called “Breast cancer” dataset that has been widely studied in machine learning since the 1980s.

The dataset classifies breast cancer patient data as either a recurrence or no recurrence of cancer.

There are 286 examples and nine input variables.

It is a binary classification problem.

A reasonable classification accuracy score on this dataset is between 68% and 73%.

We will aim for this region, but note that the models in this tutorial are not optimized: they are designed to demonstrate encoding schemes.

You can download the dataset and save the file as “breast-cancer.

csv” in your current working directory.

Looking at the data, we can see that all nine input variables are categorical.

Specifically, all variables are quoted strings; some are ordinal and some are not.

We can load this dataset into memory using the Pandas library.

Once loaded, we can split the columns into input (X) and output (y) for modeling.

Finally, we can force all fields in the input data to be string, just in case Pandas tried to map some automatically to numbers (it does try).

We can also reshape the output variable to be one column (e.

g.

a 2D shape).

We can tie all of this together into a helpful function that we can reuse later.

Once loaded, we can split the data into training and test sets so that we can fit and evaluate a deep learning model.

We will use the train_test_split() function from scikit-learn and use 67% of the data for training and 33% for testing.

Tying all of these elements together, the complete example of loading, splitting, and summarizing the raw categorical dataset is listed below.

Running the example reports the size of the input and output elements of the train and test sets.

We can see that we have 191 examples for training and 95 for testing.

Now that we are familiar with the dataset, let’s look at how we can encode it for modeling.

An ordinal encoding involves mapping each unique label to an integer value.

As such, it is sometimes referred to simply as an integer encoding.

This type of encoding is really only appropriate if there is a known relationship between the categories.

This relationship does exist for some of the variables in the dataset, and ideally, this should be harnessed when preparing the data.

In this case, we will ignore any possible existing ordinal relationship and assume all variables are categorical.

It can still be helpful to use an ordinal encoding, at least as a point of reference with other encoding schemes.

We can use the OrdinalEncoder() from scikit-learn to encode each variable to integers.

This is a flexible class and does allow the order of the categories to be specified as arguments if any such order is known.

Note: I will leave it as an exercise for you to update the example below to try specifying the order for those variables that have a natural ordering and see if it has an impact on model performance.

The best practice when encoding variables is to fit the encoding on the training dataset, then apply it to the train and test datasets.

The function below, named prepare_inputs(), takes the input data for the train and test sets and encodes it using an ordinal encoding.

We also need to prepare the target variable.

It is a binary classification problem, so we need to map the two class labels to 0 and 1.

This is a type of ordinal encoding, and scikit-learn provides the LabelEncoder class specifically designed for this purpose.

We could just as easily use the OrdinalEncoder and achieve the same result, although the LabelEncoder is designed for encoding a single variable.

The prepare_targets() integer encodes the output data for the train and test sets.

We can call these functions to prepare our data.

We can now define a neural network model.

We will use the same general model in all of these examples.

Specifically, a MultiLayer Perceptron (MLP) neural network with one hidden layer with 10 nodes, and one node in the output layer for making binary classifications.

Without going into too much detail, the code below defines the model, fits it on the training dataset, and then evaluates it on the test dataset.

If you are new to developing neural networks in Keras, I recommend this tutorial:Tying all of this together, the complete example of preparing the data with an ordinal encoding and fitting and evaluating a neural network on the data is listed below.

Running the example will fit the model in just a few seconds on any modern hardware (no GPU required).

The loss and the accuracy of the model are reported at the end of each training epoch, and finally, the accuracy of the model on the test dataset is reported.

Your specific results will vary given the stochastic nature of the learning algorithm.

Try running the example a few times.

In this case, we can see that the model achieved an accuracy of about 70% on the test dataset.

Not bad, given that an ordinal relationship only exists for some of the input variables, and for those where it does, it was not honored in the encoding.

This provides a good starting point when working with categorical data.

A better and more general approach is to use a one hot encoding.

A one hot encoding is appropriate for categorical data where no relationship exists between categories.

It involves representing each categorical variable with a binary vector that has one element for each unique label and marking the class label with a 1 and all other elements 0.

For example, if our variable was “color” and the labels were “red,” “green,” and “blue,” we would encode each of these labels as a three-element binary vector as follows:Then each label in the dataset would be replaced with a vector (one column becomes three).

This is done for all categorical variables so that our nine input variables or columns become 43 in the case of the breast cancer dataset.

The scikit-learn library provides the OneHotEncoder to automatically one hot encode one or more variables.

The prepare_inputs() function below provides a drop-in replacement function for the example in the previous section.

Instead of using an OrdinalEncoder, it uses a OneHotEncoder.

Tying this together, the complete example of one hot encoding the breast cancer categorical dataset and modeling it with a neural network is listed below.

The example one hot encodes the input categorical data, and also label encodes the target variable as we did in the previous section.

The same neural network model is then fit on the prepared dataset.

Your specific results will vary given the stochastic nature of the learning algorithm.

Try running the example a few times.

In this case, the model performs reasonably well, achieving an accuracy of about 72%, close to what was seen in the previous section.

A more fair comparison would be to run each configuration 10 or 30 times and compare performance using the mean accuracy.

Recall, that we are more focused on how to encode categorical data in this tutorial rather than getting the best score on this specific dataset.

Ordinal and one hot encoding are perhaps the two most popular methods.

A newer technique is similar to one hot encoding and was designed for use with neural networks, called a learned embedding.

A learned embedding, or simply an “embedding,” is a distributed representation for categorical data.

Each category is mapped to a distinct vector, and the properties of the vector are adapted or learned while training a neural network.

The vector space provides a projection of the categories, allowing those categories that are close or related to cluster together naturally.

This provides both the benefits of an ordinal relationship by allowing any such relationships to be learned from data, and a one hot encoding in providing a vector representation for each category.

Unlike one hot encoding, the input vectors are not sparse (do not have lots of zeros).

The downside is that it requires learning as part of the model and the creation of many more input variables (columns).

The technique was originally developed to provide a distributed representation for words, e.

g.

allowing similar words to have similar vector representations.

As such, the technique is often referred to as a word embedding, and in the case of text data, algorithms have been developed to learn a representation independent of a neural network.

For more on this topic, see the post:An additional benefit of using an embedding is that the learned vectors that each category is mapped to can be fit in a model that has modest skill, but the vectors can be extracted and used generally as input for the category on a range of different models and applications.

That is, they can be learned and reused.

Embeddings can be used in Keras via the Embedding layer.

For an example of learning word embeddings for text data in Keras, see the post:One embedding layer is required for each categorical variable, and the embedding expects the categories to be ordinal encoded, although no relationship between the categories is assumed.

Each embedding also requires the number of dimensions to use for the distributed representation (vector space).

It is common in natural language applications to use 50, 100, or 300 dimensions.

For our small example, we will fix the number of dimensions at 10, but this is arbitrary; you should experimenter with other values.

First, we can prepare the input data using an ordinal encoding.

The model we will develop will have one separate embedding for each input variable.

Therefore, the model will take nine different input datasets.

As such, we will split the input variables and ordinal encode (integer encoding) each separately using the LabelEncoder and return a list of separate prepared train and test input datasets.

The prepare_inputs() function below implements this, enumerating over each input variable, integer encoding each correctly using best practices, and returning lists of encoded train and test variables (or one-variable datasets) that can be used as input for our model later.

Now we can construct the model.

We must construct the model differently in this case because we will have nine input layers, with nine embeddings the outputs of which (the nine different 10-element vectors) need to be concatenated into one long vector before being passed as input to the dense layers.

We can achieve this using the functional Keras API.

If you are new to the Keras functional API, see the post:First, we can enumerate each variable and construct an input layer and connect it to an embedding layer, and store both layers in lists.

We need a reference to all of the input layers when defining the model, and we need a reference to each embedding layer to concentrate them with a merge layer.

We can then merge all of the embedding layers, define the hidden layer and output layer, then define the model.

When using a model with multiple inputs, we will need to specify a list that has one dataset for each input, e.

g.

a list of nine arrays each with one column in the case of our dataset.

Thankfully, this is the format we returned from our prepare_inputs() function.

Therefore, fitting and evaluating the model looks like it does in the previous section.

Additionally, we will plot the model by calling the plot_model() function and save it to file.

This requires that pygraphviz and pydot are installed, which can be a pain on some systems.

If you have trouble, just comment out the import statement and call to plot_model().

Tying this all together, the complete example of using a separate embedding for each categorical input variable in a multi-input layer model is listed below.

Running the example prepares the data as described above, fits the model, and reports the performance.

Your specific results will vary given the stochastic nature of the learning algorithm.

Try running the example a few times.

In this case, the model performs reasonably well, matching what we saw for the one hot encoding in the previous section.

As the learned vectors were trained in a skilled model, it is possible to save them and use them as a general representation for these variables in other models that operate on the same data.

A useful and compelling reason to explore this encoding.

To confirm our understanding of the model, a plot is created and saved to the file embeddings.

png in the current working directory.

The plot shows the nine inputs each mapped to a 10 element vector, meaning that the actual input to the model is a 90 element vector.

Note: Click to the image to see the large version.

Plot of the Model Architecture With Separate Inputs and Embeddings for each Categorical VariableClick to Enlarge.

This section lists some common questions and answers when encoding categorical data.

Or, what if I have a mixture of categorical and ordinal data?You will need to prepare or encode each variable (column) in your dataset separately, then concatenate all of the prepared variables back together into a single array for fitting or evaluating the model.

Or, what if I concatenate many one hot encoded vectors to create a many thousand element input vector?You can use a one hot encoding up to thousands and tens of thousands of categories.

Also, having large vectors as input sounds intimidating, but the models can generally handle it.

Try an embedding; it offers the benefit of a smaller vector space (a projection) and the representation can have more meaning.

This is unknowable.

Test each technique (and more) on your dataset with your chosen model and discover what works best for your case.

This section provides more resources on the topic if you are looking to go deeper.

In this tutorial, you discovered how to encode categorical data when developing neural network models in Keras.

Specifically, you learned:Do you have any questions?.Ask your questions in the comments below and I will do my best to answer.

with just a few lines of PythonDiscover how in my new Ebook: Deep Learning With PythonIt covers end-to-end projects on topics like: Multilayer Perceptrons, Convolutional Nets and Recurrent Neural Nets, and more.

Skip the Academics.

Just Results.

.. More details