The most basic regression model assumes linear dependency between features and target variable. The model is often fitted using least squares approach, that is, the best model minimizes the squares of the errors. In many cases, linear regression is not able to model complex relations; for example, the following diagram shows four different sets of points having the same linear regression line. The upper-left model captures the general trend and can be considered as a proper model, whereas the bottom-left model fits points much better (except for one outlier, which should be carefully checked), and the upper and lower-right side linear models completely miss the underlying structure of the data and cannot be considered proper models: