SSkilvy
RULog inStart free

Three main ML concepts

To talk about ML you need three basic concepts: data, features and a model. They form the basis of any ML system. Let us cover them with a clear example — predicting an apartments price.

DataFeaturesModel Training Predictions
Data consists of features; a model learns to link features to an answer.

Data

Data is the examples a model learns on. For predicting an apartments price, the data is a table of sold apartments: each has known characteristics and the real sale price. The more and better the data, the better the model learns.

Features

Features are the characteristics of each example by which the model draws a conclusion. For an apartment the features are: area, number of rooms, floor, district, condition. The model learns to link features to the answer (price). Choosing good features is critical: "area" is useful for price, "wallpaper color" — hardly.

The model

A model is what learns from data to find the link between features and the answer. After training, the model can predict: give it the features of a new apartment (not in the data) — it predicts the price based on learned patterns.

Explain data, features and a model on the example of predicting whether a customer will come again.
Example: predict whether a customer returns to a store. 1) Data: a table of past customers where for each it is known whether they returned or not (this is what we learn on). 2) Features: customer characteristics by which we predict — how much they spent, how long ago they visited, how many purchases, whether they left a review. The model looks for a link between these features and whether the customer returned. 3) Model: after training on past customers, it takes a new customers features and predicts the return probability. Key: data is examples for training; features describe each example; the model is what learns to link features to the answer. Quality depends on both the data and choosing the right features.

Training and prediction — two stages

Working with an ML model splits into two stages: training (the model learns on data with known answers) and prediction (the trained model is applied to new data). First we learn on the past, then apply to the new. Understanding this split is the key to everything that follows.

Rule: data is examples for training, features describe an example, the model is what learns to link features to the answer. Good data and features decide everything.

🧠 What are features in ML?

← BackNext →