1ofN

1ofN, also known as one-of-N encoding or one-hot encoding, is a way of representing categorical variables as vectors in which exactly one of the N elements is active. In the standard form, the active element is set to 1 and all other elements are 0, producing a sparse, orthogonal code for each category. The concept is widely used in data processing, information encoding, and machine learning to convert nominal categories into numerical form suitable for algorithms.

In supervised learning, a common use is to encode the target variable when there are N classes. The true class corresponds to the basis vector with a 1 in the appropriate position. Many algorithms, including neural networks, logistic regression, support vector machines, and tree-based methods, can process one-hot encoded features. In neural networks, the final layer often outputs a probability distribution over the N classes via a softmax function, with the target represented as a one-hot vector for training.

Advantages of 1ofN encoding include its simplicity, interpretability, and ease of use with linear models and distance-based methods. Distances between distinct categories are straightforward to compute, and each category is equidistant from the others in the encoded space. Disadvantages involve increasing dimensionality as the number of categories grows, which can lead to sparsity and higher memory or computation costs. For very large category sets, alternatives such as target encoding, hashing tricks, or learned embeddings may offer more compact representations while capturing similarity between categories.

See also: one-hot encoding, categorical variable encoding, embeddings, softmax.