Sample data generation

The Sample-data generation is the process of generating fake, but as much as realistic as possible, measurements associated to the mock Basket.

The “mock Basket” is a special Basket used for tests that is characterized by the ID 0x23232323 in hex, or 589505315 in decimal.

The problem consist in generating three type of measurements:

• Accelerometer data: the movement of the Basket or of the Basket board above a certain threshold.
• People detection data: the detection of people standing in front of the basket.
• Basket data: anything that occluded the IR detector (most commonly, the ball entering the basket).

The task of the generator, given a range of dates, is to generate for each day within the range fake measurements.

In order to do this, the generator distinguish three types of days in a year:

• Busy day: a day for example of school or of work.
• Unplayable day: a day that is considered to be unplayable for example because of bad weather conditions (rain or cold).
• Playable day: any other day.

For every day type and for every measurement type, it’s built a probability distribution from which the generator will sample pseudo-realistic timestamps for the measurement.

Even the number of timestamps drawn within a day per measurement are dependent on the day type. Trivially it’s expected to have more baskets, people detected and accelerometer data on a Playable day rather than on an Unplayable day.

How a day is classified?

A date is classified using the following procedure:

• If a day is unplayable then it’s Unplayable
• Otherwise the day is playable, but if it’s busy, then it’s Busy
• Otherwise it’s Playable

Unplayable day classification

The condition for a day to be unplayable should be derived from realistic weather data. However we don’t have access to a weather service that could give us the exact weather conditions back in the past and its integration would have been an over-engineering for the generation of sample fake-data.

For this purpose, it’s been implemented a predicate that given a date returns whether the day was playable or unplayable. This predicate has to be deterministic, thus the day condition can’t involve a random generator. The randomness has been replaced with a noise function that given a numerical representation of the date (e.g. Unix timestamp) returns a [0, 1] value. This value is then converted to a boolean value giving more probability to autumn/winter month’ days to be unplayable and less to spring/summer days.

def noise(n: float):
    return abs(modf(sin(n) * 43758.5453123)[0])

def is_unplayable_day(date: datetime):
    value = noise(mod(date.year, 5) * 365 + date.month * 31 + date.day)
    return value <= [0.9, 0.8, 0.6, 0.38, 0.3, 0.1, 0.08, 0.07, 0.1, 0.3, 0.79, 0.9][date.month - 1]

Busy day classification

A day is considered busy trivially if respects these hand-crafted conditions:

• The month is within school period (January to May or September to December)
• It’s not a Sunday
• It’s not Christmas holiday