Understanding Consumer Preferences Through Latent Spaces / Entendendo as preferências do consumidor através dos espaços latentes

Abstract

The Latent Spaces technique has applications in areas such as natural language processing, image recognition and multi-language translation. It permits embedding vectors use. These are vectors in a k-dimensional vector space representing increasingly advanced study objects learning models, forming entirely new basis areas. Those vectors can capture semantic study object features and once trained can be reused in other models, decreasing training time and increasing knowledge transfer. This happens in Google Word2Vec and Facebook FastText pre-trained word vectors set. This work explores Latent Spaces techniques to understand preferences through recommendation mechanism implementation on top of MovieLens dataset from University of Minnesota. From it were extracted a sequence of triples (userId, movieId, rating) , representing ratings given by users to particular films. Two k-dimensional Latent Vector Spaces representing film characteristics and the corresponding user preferences were created using Google Tensorflow, Machine Learning techniques like SGD and Matrix Factorization. The performance benchmark was the mean-square error on the dimensionality k of the Latent Spaces. The capacity of the Latent Spaces was than evaluated to abstract non-trivial information about films and compared with small cosine distance technique. The system was able to find subjective related similarities that would be tough to code in a straightforward manner. Finally, it was explored an alternative to generate user vectors through neural networks. The techniques exposed here support the case that Machine Learning techniques like Latent Spaces can be used in business decision making.