In recent research on recommendation systems, there has been a focus on exploring the use of linear regression (autoencoder) models to learn item similarity. This study aims to establish a connection between linear autoencoder models and ZCA whitening for recommendation data. The paper demonstrates that the dual form solution of a linear autoencoder model has ZCA whitening effects on item feature vectors, even though items are regarded as input features in the primal problem of the autoencoder/regression model. Additionally, the study validates the application of a linear autoencoder to low-dimensional item vectors obtained through embedding methods like Item2vec to estimate item-item similarities. Preliminary experimental results suggest that whitening low-dimensional item embeddings can be effective.
The Unseen Whitening Effects of Linear Autoencoders for Recommendation
