Machine learning glossary of important terms
The following list is a compilation of important machine learning terms that are useful as you build your custom models in ML.NET.
In classification, accuracy is the number of correctly classified items divided by the total number of items in the test set. Ranges from 0 (least accurate) to 1 (most accurate). Accuracy is one of evaluation metrics of the model performance. Consider it in conjunction with precision, recall, and F-score.
Area under the curve (AUC)
In binary classification, an evaluation metric that is the value of the area under the curve that plots the true positives rate (on the y-axis) against the false positives rate (on the x-axis). Ranges from 0.5 (worst) to 1 (best). Also known as the area under the ROC curve, i.e., receiver operating characteristic curve. For more information, see the Receiver operating characteristic article on Wikipedia.
Calibration is the process of mapping a raw score onto a class membership, for binary and multiclass classification. Some ML.NET trainers have a
NonCalibrated suffix. These algorithms produce a raw score that then must be mapped to a class probability.
In ML.NET, a catalog is a collection of extension functions, grouped by a common purpose.
For example, each machine learning task (binary classification, regression, ranking etc) has a catalog of available machine learning algorithms (trainers). The catalog for the binary classification trainers is: BinaryClassificationCatalog.BinaryClassificationTrainers.
When the data is used to predict a category, supervised machine learning task is called classification. Binary classification refers to predicting only two categories (for example, classifying an image as a picture of either a 'cat' or a 'dog'). Multiclass classification refers to predicting multiple categories (for example, when classifying an image as a picture of a specific breed of dog).
Coefficient of determination
In regression, an evaluation metric that indicates how well data fits a model. Ranges from 0 to 1. A value of 0 means that the data is random or otherwise cannot be fit to the model. A value of 1 means that the model exactly matches the data. This is often referred to as r2, R2, or r-squared.
Data is central to any machine learning application. In ML.NET data is represented by IDataView objects. Data view objects:
- are made up of columns and rows
- are lazily evaluated, that is they only load data when an operation calls for it
- contain a schema that defines the type, format and length of each column
A class in ML.NET that implements the IEstimator<TTransformer> interface.
An estimator is a specification of a transformation (both data preparation transformation and machine learning model training transformation). Estimators can be chained together into a pipeline of transformations. The parameters of an estimator or pipeline of estimators are learned when Fit is called. The result of Fit is a Transformer.
A .NET method that is part of a class but is defined outside of the class. The first parameter of an extension method is a static
this reference to the class to which the extension method belongs.
Extension methods are used extensively in ML.NET to construct instances of estimators.
A measurable property of the phenomenon being measured, typically a numeric (double) value. Multiple features are referred to as a Feature vector and typically stored as
double. Features define the important characteristics of the phenomenon being measured. For more information, see the Feature article on Wikipedia.
Feature engineering is the process that involves defining a set of features and developing software that produces feature vectors from available phenomenon data, i.e., feature extraction. For more information, see the Feature engineering article on Wikipedia.
A parameter of a machine learning algorithm. Examples include the number of trees to learn in a decision forest or the step size in a gradient descent algorithm. Values of Hyperparameters are set before training the model and govern the process of finding the parameters of the prediction function, for example, the comparison points in a decision tree or the weights in a linear regression model. For more information, see the Hyperparameter article on Wikipedia.
The element to be predicted with the machine learning model. For example, the breed of dog or a future stock price.
In classification, an evaluation metric that characterizes the accuracy of a classifier. The smaller log loss is, the more accurate a classifier is.
A loss function is the difference between the training label values and the prediction made by the model. The parameters of the model are estimated by minimizing the loss function.
Different trainers can be configured with different loss functions.
Mean absolute error (MAE)
Traditionally, the parameters for the prediction function. For example, the weights in a linear regression model or the split points in a decision tree. In ML.NET, a model contains all the information necessary to predict the label of a domain object (for example, image or text). This means that ML.NET models include the featurization steps necessary as well as the parameters for the prediction function.
A feature extraction scheme for text data: any sequence of N words turns into a feature value.
Normalization is the process of scaling floating point data to values between 0 and 1. Many of the training algorithms used in ML.NET require input feature data to be normalized. ML.NET provides a series of transforms for normalization
Numerical feature vector
A feature vector consisting only of numerical values. This is similar to
All of the operations needed to fit a model to a data set. A pipeline consists of data import, transformation, featurization, and learning steps. Once a pipeline is trained, it turns into a model.
In classification, the precision for a class is the number of items correctly predicted as belonging to that class divided by the total number of items predicted as belonging to the class.
In classification, the recall for a class is the number of items correctly predicted as belonging to that class divided by the total number of items that actually belong to the class.
Regularization penalizes a linear model for being too complicated. There are two types of regularization:
- $L_1$ regularization zeros weights for insignificant features. The size of the saved model may become smaller after this type of regularization.
- $L_2$ regularization minimizes weight range for insignificant features, This is a more general process and is less sensitive to outliers.
A supervised machine learning task where the output is a real value, for example, double. Examples include predicting stock prices. For more information, see the Regression section of the Machine learning tasks topic.
Relative absolute error
Relative squared error
Root of mean squared error (RMSE)
In regression, an evaluation metric that is the square root of the average of the squares of the errors.
Scoring is the process of applying new data to a trained machine learning model, and generating predictions. Scoring is also known as inferencing. Depending on the type of model, the score may be a raw value, a probability, or a category.
Supervised machine learning
A subclass of machine learning in which a desired model predicts the label for yet-unseen data. Examples include classification, regression, and structured prediction. For more information, see the Supervised learning article on Wikipedia.
The process of identifying a model for a given training data set. For a linear model, this means finding the weights. For a tree, it involves identifying the split points.
An ML.NET class that implements the ITransformer interface.
Unsupervised machine learning
A subclass of machine learning in which a desired model finds hidden (or latent) structure in data. Examples include clustering, topic modeling, and dimensionality reduction. For more information, see the Unsupervised learning article on Wikipedia.