Koinder Practitioner Track 16+ ← My Learning About the track Clubhouse

Data Analysis: Where 70% of the Work Actually Is

Is machine learning just data analysis? Partly yes, importantly no — and the difference decides what you build.

Module 3 · Practitioner Track · 18+

First, the honest answer to the question

You asked whether all of this comes down to data analysis. The answer is partly yes, and importantly no — and the difference decides what you build.

Data analysisMachine learning
QuestionWhat happened, and why?What will happen with this new case?
OutputAn explanation for a humanA decision made automatically
Who actsA person, having read itThe system, thousands of times a day
Succeeds whenSomebody understands something they did notIt is right about cases nobody has seen
The same data, two different jobs

Analysis: "Sales fell 22% in the north-east last quarter, and it tracks the fuel price almost exactly." A human reads that and decides something.
Machine learning: a model that predicts next month's sales per region, running every night, with nobody reading anything.

But here is why you were right to ask.

You cannot build a model without doing the analysis first, and roughly 70% of any machine learning project is data work. Every failure in the previous module — overfitting, leakage, imbalance, drift — is a data problem, not a model problem.

There is also a commercial point worth hearing. Data analysis alone is a career, it is easier to sell to Nigerian businesses than machine learning, and the person who can clean a filthy dataset is hired ahead of the person who has only used clean teaching data.

The pipeline — nine steps, and where the time really goes

1 · The question

Written down as one sentence, before any data is touched. "Which customers are about to stop buying?" not "let us look at the sales data". A vague question produces months of work nobody uses.

2 · Collect

Where does it live, who owns it, are you allowed to have it. In Nigerian businesses this is often the hardest step, and it is a relationship problem more than a technical one.

3 · Look at it. Actually look.

Open the raw file. Read a hundred rows with your eyes. Every experienced analyst has a story about the thing they only found because they looked.

4 · Clean

Duplicates, missing values, three date formats, "N/A" and "n/a" and blank meaning different things, ₦ signs inside number columns, names with trailing spaces.

5 · Explore

Count things. Plot things. Find what is strange before you find what is true.

6 · Features

Turn raw records into something a model can use. A date becomes day-of-week, month, is-it-a-holiday. This step decides more outcomes than the choice of model does — you will see it proven with numbers in the next module.

7 · Model

The famous step. Frequently the shortest.

8 · Evaluate honestly

On data the model has never seen. Broken down by group, never as a single number. The next module shows a model at 76% that is actually perfect at three things and useless at one.

9 · Deploy and watch

A model is not a finished product. It is a claim about the world that slowly expires.

The four questions to ask of any dataset, before anything else

1. How was this collected, and who decided what to record?
The hospital records trap

A dataset of Nigerian patients contains only people who reached a hospital. Everyone treated at home, or who could not afford transport, is missing. Any model built on it is a model of people who reach hospitals, not of Nigerians.

2. Who is missing?
The smartphone assumption

App data describes people who own smartphones and can afford data. A model of "customer behaviour" built from it will be confidently wrong about most of the market.

3. What does each column actually mean?
The column called "amount"

Is it naira or kobo? Before or after discount? Including VAT? Nobody documented it and the person who built the table has left. This single question has ruined more analyses than any algorithm.

4. What is the date range, and what happened during it?
Context is a column you cannot see

Nigerian data spanning 2020 contains the lockdown. Data spanning mid-2023 contains the fuel subsidy removal. Data across an election contains an election. None of that is in a column, and all of it is in the numbers.

Practical — before the next module

Take any real dataset. Your own sales, market prices, school results, anything. Do not model it. Just answer, in writing:

1

What is the one question worth asking of this?

2

How was it collected, and who is missing from it?

3

List every dirty thing in it. Most people find between eight and twenty.

4

What was happening in the world during this date range?

That list is the job. In the next module you will build a real model — and you will watch the data work matter more than the model does, measured in actual percentage points.

🤖

One question before we start

The Practitioner track is built for adults. It assumes you are ready to deploy real systems that real strangers will use, and to be answerable for what they do.

We ask because in Nigeria you become an adult at 18, and this track is a paid commitment. Nothing here is stored for anyone under that age.