Aggregation

Aggregation means: many individual values are combined into one or a few overall values. An example from school is the report card grade. Twenty individual grades become a single average. The average says something about the whole, but the individual pieces of work disappear within it. That is precisely the core of every aggregation: you gain an overview and lose detail. Typical computational steps for this are sum, average, count, minimum, and maximum.

Why no one could work with data without summarizing

Raw data is mostly worthless to humans. An online shop stores every single click, every shopping cart, every order. That can amount to billions of rows per day. No one reads through something like that. Only the summarized figure — revenue per day, orders per country — yields a meaningful statement.

Aggregation is also an important tool in data protection. As long as individual values are stored, they are tied to a specific person. Once combined, often only an anonymous statistic remains. A statement like “340 cars charge per day in this neighborhood” reveals no one. The list of individual charging events with timestamps, on the other hand, does.

The price for this is a loss of information, and that can be misleading. An average salary of 6,000 euros sounds good, but it might result from nine salaries of 2,000 euros and one of 42,000 euros. That’s why, alongside the average, the median is often given as well — the value exactly in the middle. Anyone who sees only a single aggregated figure should always ask what it is hiding.

From individual value to key metric

Technically, aggregation almost always happens in two steps. First, the data is grouped, i.e., sorted by a characteristic: by day, by product, by country. Then a computation rule is applied to each group, for example “compute the sum”. At the end, there is exactly one row per group. In database languages, this command has simply been called GROUP BY for decades.

In neural networks, the learning models behind today’s AI, the same principle appears in several places. During training, adjustments are not made after every single example. Instead, the model processes a batch of roughly 32 to 1,000 examples and averages their results. This averaged feedback is more stable than that of a single example, which might just happen to be an outlier.

Another place is the attention mechanism in language models. There, for each word, a weighted summary of all other words in the text is computed. Important words contribute strongly, unimportant ones barely at all. This too is aggregation, just with varying weights instead of a simple average. The term should be distinguished from filtering: filtering discards rows, aggregation combines them through computation.

Aggregation in products, news, and statistics

You encounter this principle constantly in everyday life. Your fitness app doesn’t show every single step, but the daily total. A weather app combines measurements from many stations into a single temperature for your location. A news aggregator like Google News collects articles from many outlets and bundles them into one topic.

In economic news, almost every figure mentioned is an aggregate. The inflation rate arises from hundreds of thousands of individual prices. The DAX combines the share prices of 40 companies into a single number. Such figures are useful, but they are always a selection and a computational rule, not a pure observation.

A current example from the world of AI is federated learning. Here, many smartphones train a model locally using their users' data. What goes to the central server is not the chat history, but only an aggregated change to the model’s values. The server averages the contributions of thousands of devices into a new model. This is how, for example, keyboard prediction improves without your messages ever leaving your phone.