Machine learning

What machine learning means

Machine learning is an approach where a system learns patterns from data instead of being given rules. The difference from classical programming lies in the direction: there a person writes the rule and the program delivers the result. Here the system receives examples and derives the rule itself.

The consequence is a different kind of reliability. A learned system is usually right and never guaranteed. It can state a probability but not a proof. Anyone confusing the two builds applications that work well on average and surprise in the individual case.

The approaches found in retail

Three basic kinds are common. In supervised learning the system receives examples with a known answer and should sort unknown cases afterwards. That is how fraud detection, churn prediction or assigning customers to groups is built. In unsupervised learning there is no known answer; groups and patterns are sought, for instance when bundling similar items. In reinforcement learning the system learns through trial and feedback, which in practice appears mainly in steering advertising budgets.

Language models are a special case. They do not learn decisions but the probability of text sequences, which places them closer to patterns than to forecasts.

What training data has to deliver

• Volume: too few examples produce a system that knows the examples by heart and sorts new cases badly.
• Balance: if fraud cases only come from one country and one payment method, the system learns that pattern and marks everything else as unremarkable.
• Currency: when the market changes, the model ages. A forecast model from before a crisis calculates with conditions that no longer apply.
• Traceability: without documentation of which data was used when, a decision cannot be explained later.

Where the limits lie

The first limit is causality. A model finds connections, not causes. Two things occurring together does not mean one causes the other, and a measure built on a correlation can remain without effect.

The second limit is drift. When customer behaviour changes, the hit rate falls without anything changing in the model. That is why every use needs ongoing control, not only a sign off at the start. The third limit is responsibility. A model makes no decision, it delivers a number. Whoever decides on that number carries the decision.

Frequent questions

• Is machine learning the same as AI? It is a part of it. AI is the broader term.
• Do I need large data volumes? Not for simple tasks, but usually yes for forecasts.
• How do I recognise a good use? By a task that repeats often, gives clear feedback and whose mistakes are bearable.