How a computer learns from examples

Type "dog" into the search on your phone's photos. It finds your dog pictures, even ones you never named.

Nobody wrote a rule for what a dog looks like. The phone's software learnt from millions of pictures that people had already labelled. Today you build a tiny version that learns busy days from quiet ones.

In machine learning, you do not write the rule. You give the computer examples with the right answer attached, its label.

Here each day of the year is an example. Its label is busy if it had 300 hires or more, and quiet if not. The computer learns from the training data, then guesses the label of days it was not shown. Its accuracy is the share it gets right.

docker exec jupyter python3 steps/knn.py

The model in knn.py is nearest neighbours. To label a new day, it finds the past days most like it. Then it lets them vote.

How alike two days are is a gap: here, the difference in temperature plus the difference in rain. The number of neighbours that vote is called k. With k = 3, the three smallest gaps vote.

# Past days: temperature, rain in mm, and the label.
past = [[17, 0, "quiet"], [15, 2, "busy"], [9, 0, "quiet"],
        [16, 6, "quiet"], [14, 0, "busy"]]

# A new day to label.
new_temp = 16
new_rain = 0

for day in past:
    # The gap: the difference in temperature plus the difference in rain.
    gap = abs(day[0] - new_temp) + abs(day[1] - new_rain)
    print(day[2], "gap", gap)

# Which three days have the smallest gaps? What do they vote?
# Change new_rain to 5 and press Run again.

docker exec jupyter python3 steps/knnk.py

With k = 1, the model scores 0.963 on the days it learnt from. On days kept back it scores only 0.728.

It has learnt each training day by heart, oddities and all, so it copes badly with new ones. With k = 5 the two scores are 0.883 and 0.815, much closer. A model has learnt when it does well on examples it has never seen.

A model only knows its examples. If they are unfair, the model is unfair, and it hides this behind a number.

One large company trained a tool to sort job applications on ten years of its past hires, who were mostly men. The tool marked down applications that mentioned a women's club. The company stopped using it. This unfairness is called bias.

import pandas as pd

hires = pd.read_csv("hires.csv")
# True on a day with 300 hires or more, the label busy.
hires["busy"] = hires["hires"] >= 300
# The mean of True and False is the share of days that were busy.
hires.groupby("weather")["busy"].mean().round(2)

Type the cell above into a notebook and run it. Which weather had the largest share of busy days?

The model only knew temperature and rain. Name one thing it never saw that also makes a day busy, and say how you would add it.