Exploring random indexing for profile learning

Abstract

Random Indexing is a recent technique for dimensionality reduction that allows to obtain a word space model from a set of contexts. This technique is less computationally expensive in comparison with others like LSI, Word2Vec or LDA. These characteristics turn it an attractive prospect to be used in an online learning environment. In this work, we compare several variants reported in the Random Indexing literature with the aim of using on the profile learning task. Experiments conducted in a subcollection of the dataset Reuter-21578 show that Random Indexing produces promising results, identifying some versions without actual advantage for the task at hand. Results obtained, by comparing Random Indexing with LDA, Word2Vec or LSI, also show that this technique is a viable alternative for representing documents.