Improving Accuracy of Recommender System using Stack Autoencoder

Abstract

The massive amount of information available on the World Wide Web has made a requirement for business organisation to developed a Recommendation Systems. Recommendation Systems are becoming gradually important to individual users and
businesses for providing Personalized Recommendations. Collaborative Filtering aims at using the feedback of users to deliver personalized recommendations. It is surprisingly
amazing that Neural Networks models are capable of learning latent features on huge and heterogeneous datasets. These thesis proposes the use of Stack Autoencoder to develop
recommendation systems and also a comparison between User rating vector and Item rating vector as input to the model has been done, here the Autoencoder tries to predict the input on the output layer through a bottleneck i.e. fewer hidden nodes in the hidden layer, these hidden units become the learned compressed information of the input/output data. Experimental result has shown us that Recommendation System using Stack Autoencoder outperforms Matrix Factorization, and also feeding the Stacked Autoencoder with Item rating vectors performs better in speed and accuracy than feeding the model with User rating vectors on ML dataset 100k and 1M User-Item rating. This work clearly establishes the value of using Stack Autoencoder to obtain useful features for predicting unseen movies to the user. We also shown that Autencoder can learn the data representation better in lower dimension, and computationally its faster and more accurate than the current state-of-the-art methods in neural approaches to Collaborative Filtering.