03.01 · Concept
Entities, Events, and Observations
Define entities, events, observations, labels, and timestamps for a supervised ML problem.
No video curated for this lesson yet
This lesson is written, ordered and part of the path - the video slot is the only thing still open. We are working through Everything Data lesson by lesson; 55 of 85 have their video so far.
The written notes below cover this idea in full - you lose nothing by reading instead of watching.
Rows in supervised ML training data need a precise meaning: an entity at a timestamp, with past-only features and a future-defined label. Separating entities, events, observations, labels, and time boundaries prevents leakage, conflicting targets, and datasets that cannot match production prediction behaviour.
What this lesson answers
- what is an observation in supervised machine learning
- how to define labels without data leakage
- entity versus event in machine learning data
Notes
In a supervised machine learning problem, the first modeling step is to be precise about what each row of training data represents. The entity is the thing you are making predictions about, such as a user, account, device, transaction, customer, or support ticket. An event is something that happens involving an entity, such as a login, purchase, payment failure, page view, or claim submission. An observation is a training example: a specific entity at a specific point in time, with features computed only from information available up to that point.
Common questions
- What is the difference between an entity and an observation?
- An entity is the thing being predicted about, such as an account, device, transaction, or ticket. An observation is one training example for that entity at a particular timestamp. The same entity can produce many observations over time, but each observation needs one clear prediction target and a fixed time boundary.
- Why does the observation timestamp matter so much?
- The observation timestamp defines what the model was allowed to know when making the prediction. Features must be computed from information available before or at that time. The label is then derived from what happens afterwards. Without this boundary, future information can slip into features and make offline performance misleading.
- How should I phrase a supervised ML prediction problem?
- Use a sentence that names the entity, timestamp, label, and future window. For example: for each entity, at this time, predict whether a specific outcome will happen later. That phrasing forces decisions about eligible rows, feature history, label construction, and how the training data matches production use.