Generalized Feature Embedding for Supervised, Unsupervised, and Online Learning Tasks

Feature embedding is an emerging research area which intends to transform features from the original space into a new space to support effective learning. Many feature embedding algorithms exist, but they often suffer from several major drawbacks, including (1) only handle single feature types, or users have to clearly separate features into different feature views and supply such information for feature embedding learning; (2) designed for either supervised or unsupervised learning tasks, but not for both; and (3) feature embedding for new out-of-training samples have to be obtained through a retraining phase, therefore unsuitable for online learning tasks. In this paper, we propose a generalized feature embedding algorithm, GEL, for both supervised, unsupervised, and online learning tasks. GEL learns feature embedding from any type of data or data with mixed feature types. For supervised learning tasks with class label information, GEL leverages a Class Partitioned Instance Representation (CPIR) process to arrange instances, based on their labels, as a dense binary representation via row and feature vectors for feature embedding learning. If class labels are unavailable, CPIR is naturally degenerated and treats all instances as one class. Based on the CPIR representation, GEL uses eigenvector decomposition to convert the proximity matrix into a low-dimensional space. For new out-of-training samples, their low-dimensional representation are derived through a direct conversion without a retraining phase. The learned numerical embedding features can be directly used to represent instances for effective learning. Experiments and comparisons on 28 datasets, including categorical, numerical, and ordinal features, demonstrate that embedding features learned from GEL can effectively represent the original instances for clustering, classification, and online learning.