Categorical features

by: Chris Cave and Emil Larsen
(Feyn version 3.0 or newer)

Categorical features

A feature is categorical if there is no clear ordering in the values the feature can take. The values a categorical feature can take are called categories. Below is an example dataset containing categorical features and their categories:

Country	Favourite colour	Gender	Smoker/non-smoker
Denmark	Red	Male	0
Spain	Yellow	Female	1
UK	Blue	Male	1
Brazil	Green	Male	0
USA	Yellow	Female	1
Italy	Red	Female	0

In the example above each feature does not have an obvious ordering.

How the QLattice treats categorical features

Before we pass the dataset above through the QLattice we need to specify the semantic types of the categoricals.

import feyn
import numpy as np
import pandas as pd

data = pd.DataFrame({
    "Country": ["Denmark", "Spain", "UK", "Brazil", "USA", "Italy"],
    "Favourite colour": ["Red", "Yellow", "Blue", "Green", "Yellow", "Red"],
    "Gender": ["Male", "Female", "Male", "Male", "Female", "Female"],
    "Smoker": [0,1,1,0,1,0]
})

stypes = {
    'Country': 'c',
    'Favourite colour': 'c',
    'Gender': 'c',
    }

After which we can pass into the QLattice.

ql = feyn.QLattice(random_seed=42)
models = ql.auto_run(
    data=data,
    output_name="Smoker",
    stypes=stypes,
    n_epochs=1,
    max_complexity=2
)
model = models[0]

Here is one graph from the output of auto run.

How has the QLattice interpreted the categories of this feature as a number so it can be an input to a mathematical function?

Each category is associated to a numerical value which we call its weight. The weight gets learnt while the model is being fitted to the data.

We can see the weights associated to each category by calling params:

input_node = model[2]
print(f"category weights: {input_node.params['categories']}")
print(f"bias: {input_node.params['bias']}")

category weights: [
    ('Yellow', 0.26123558974851063),
    ('Blue', 0.26123554327444204),
    ('Red', -0.25637270012528196),
    ('Green', -0.25637278717475576)
]
bias: 0.12080737359066837

When we predict on a sample that contains a category in the list above then we first convert that category to category_weight + bias.

This value gets passed into the functions inside the model and behaves like a normal numerical value.

If there are any NaN values present for some observations in a categorical feature, then the NaN values will be interpretated as a category.