Very basic but how to understand data statistically for machine learning?

Question

I’m trying to solidify my statistics so I really know how to use them in my analysis/models. However my concept of statistical testing gets completely messed up by context. I’m unsure defining exactly what is a population vs a sample to even begin with in real cases. This is rather very embarrassing to write but I need help as a rookie.

For instance, let’s say I want to predict a value for some business case. If I use my customer base, is that technically the population? Or is it a sample of all potential customers? I'm unsure on defining exactly what is an is not a population, and if it is data of an entire population, how does that affect the training process/metrics?

And when I am training a model, if I'm using a subset (train set), then that would then be considered a sample?

I’m not sure if I’m explaining it well, but hopefully someone can see what I’m not understanding when it comes to relating statistics to ML datasets, training or use cases. :’)

Answer 1

That's actually a good question - this is nothing to be embarrassed about IMO :)

I'll attempt to answer your question as a practitioner (but hopefully people who are more knowledgeable can provide better structured answers)

Population

So, what is a population ? Given a measurement setting, a population is the entirety of possible values of a variable.

Let's take an example :

Problem : What is the average age of my customer base ?

Answer : measure = average, variable = age and, population = my customer base.

Key take-away: Defining a population depends on the question and a measure (or a modeling goal) about said population.

Let's take another example :

Problem : The company I work for wants to launch a new product and the marketing team is asking What the average age of our customer base will look like in a year ?

Answer : measure = average, variable = age and, population = future customer base.

Again, think of a population as entirety of possible values of a variable you want to retrieve some sort of information/measure about.

So now the question becomes : what is the future customer base?

This is where your modeling starts ! (Start here to get an overview)

Possible Approaches:

• for example framing a hypothesis stating that my future customer base age variable will look like my current customer base age variable. And testing it to see if it holds.
• or predicting the age of the future customer base based on previous churn rates, acquisition rates, and other customer attributes. (enter machine learning - if applicable or viable)
• you can also add other feature depending on the level of complexification you require to answer your problem. For example: features based on customer consumption trends, demographics of the region the marketing team is targeting, income of targeted customers - features that would determine if those customers would be acquired (and should be acquired) and hence if those customers fall within the definition of the modeled future customer base.

Key take-away: Defining a population is a modeling exercise.

Sample

A sample is a portion of a population. To take a sample, other modeling questions can be asked and answered depending on the problem at hand.

Questions such as:

• What is the hypothesis I am working on / testing ?
• What is a representative sample of my population ?
• How much bias your sampling can afford ?

You can start here for an overview

Hope this helps !