StarSpace - Embed All The Things! (Wu et al. 2017)

Overview

• Provides a general method for embedding entities composed of discrete features

Key Results

• Denote all features of entities as n dimensional matrix

• Each entity vector is expressed as summation of all its feature vectors

• Each feature can be regarded as each entity (in case of embedding words)

• Train the feature vectors by minimizing loss on the objective function that compares positive sample to negative samples

Comments

• Provides very strong baseline for any embedding methods but seems to take more time to train compared to other embedding methods that are made for specific purposes