What is the logic of the extra columns in Tensorflow categorical encoding? - tensorflow

Question

I am following the official Tensorflow tutorial for preprocessing layers, and I am not sure I get why I end up getting these extra columns after the categorical encoding.
Here is a stripped-down minimal reproducible example (including the data):
import numpy as np
import pandas as pd
import tensorflow as tf
from tensorflow.keras import layers
from tensorflow.keras.layers.experimental import preprocessing
import pathlib
dataset_url = 'http://storage.googleapis.com/download.tensorflow.org/data/petfinder-mini.zip'
csv_file = 'datasets/petfinder-mini/petfinder-mini.csv'
tf.keras.utils.get_file('petfinder_mini.zip', dataset_url, extract=True, cache_dir='.')
df = pd.read_csv(csv_file)
# In the original dataset "4" indicates the pet was not adopted.
df['target'] = np.where(df['AdoptionSpeed']==4, 0, 1)
# Drop un-used columns.
df = df.drop(columns=['AdoptionSpeed', 'Description'])
# A utility method to create a tf.data dataset from a Pandas Dataframe
def df_to_dataset(dataframe, shuffle=True, batch_size=32):
dataframe = dataframe.copy()
labels = dataframe.pop('target')
ds = tf.data.Dataset.from_tensor_slices((dict(dataframe), labels))
if shuffle:
ds = ds.shuffle(buffer_size=len(dataframe))
ds = ds.batch(batch_size)
ds = ds.prefetch(batch_size)
return ds
batch_size = 5
ds = df_to_dataset(df, batch_size=batch_size)
[(train_features, label_batch)] = ds.take(1)
def get_category_encoding_layer(name, dataset, dtype, max_tokens=None):
# Create a StringLookup layer which will turn strings into integer indices
if dtype == 'string':
index = preprocessing.StringLookup(max_tokens=max_tokens)
else:
index = preprocessing.IntegerLookup(max_values=max_tokens)
# Prepare a Dataset that only yields our feature
feature_ds = dataset.map(lambda x, y: x[name])
# Learn the set of possible values and assign them a fixed integer index.
index.adapt(feature_ds)
# Create a Discretization for our integer indices.
encoder = preprocessing.CategoryEncoding(max_tokens=index.vocab_size())
#encoder = preprocessing.CategoryEncoding(max_tokens=2)
# Prepare a Dataset that only yields our feature.
feature_ds = feature_ds.map(index)
# Learn the space of possible indices.
encoder.adapt(feature_ds)
# Apply one-hot encoding to our indices. The lambda function captures the
# layer so we can use them, or include them in the functional model later.
return lambda feature: encoder(index(feature))
So, after running
type_col = train_features['Type']
layer = get_category_encoding_layer('Type', ds, 'string')
layer(type_col)
I get a result as:
<tf.Tensor: shape=(5, 4), dtype=float32, numpy=
array([[0., 0., 1., 0.],
[0., 0., 1., 0.],
[0., 0., 0., 1.],
[0., 0., 0., 1.],
[0., 0., 0., 1.]], dtype=float32)>
similar to what is shown in the tutorial indeed.
Notice that this is a binary classification problem (Cat/Dog):
np.unique(type_col)
# array([b'Cat', b'Dog'], dtype=object)
So, what is the logic of the 2 extra columns after the categorical encoding shown in the result above? What do they represent, and why they are 2 (and not, say, 1, or 3, or more)?
(I am perfectly aware that, should I wish for a simple one-hot encoding, I could simply use to_categorical(), but this is not the question here)

Answer 1

As already implied in the question, categorical encoding is somewhat richer that simple one-hot encoding. To see what these two columns represent it suffices to add a diagnostic print somewhere inside the get_category_encoding_layer() function:
print(index.get_vocabulary())
Then the result of the last commands will be:
['', '[UNK]', 'Dog', 'Cat']
<tf.Tensor: shape=(5, 4), dtype=float32, numpy=
array([[0., 0., 1., 0.],
[0., 0., 1., 0.],
[0., 0., 0., 1.],
[0., 0., 0., 1.],
[0., 0., 0., 1.]], dtype=float32)>
The hint should hopefully be clear: the extra two columns here represent the empty value '' and unknown ones '[UNK]', respectively, which could be present in future (unseen) data.
This is actually determined from the default arguments, not of CategoryEncoding, but of the preceding StringLookup; from the docs:
mask_token=''
oov_token='[UNK]'
You can end up with a somewhat more tight encoding (only 1 extra column instead of 2) by asking for oov_token='' instead of oov_token='[UNK]'; replace the call to StringLookup in the get_category_encoding_layer() function with
index = preprocessing.StringLookup(oov_token='',mask_token=None, max_tokens=max_tokens)
after which, the result will be:
['', 'Dog', 'Cat']
<tf.Tensor: shape=(5, 3), dtype=float32, numpy=
array([[0., 1., 0.],
[0., 1., 0.],
[0., 0., 1.],
[0., 0., 1.],
[0., 0., 1.]], dtype=float32)>
i.e. with only 3 columns (without a dedicated one for '[UNK]'). AFAIK, this is the lowest you can go - attempting to set both mask_token and oov_token to None will result to an error.