COMPARATIVE ANALYSIS OF AUTOENCODER AND LSTM IN MUSIC SESSION DATA

ABSTRACT
Deep learning is now being used in the very complicated area of item
recommendation, where each e-commerce site has its own quirks, thanks
to the exponential increase in computing power and memory. The
difficulty offered by data sparsity in this domain is immense, and it has
had a significant impact on the performance of existing algorithms in this
domain. Our research aimed to uncover a new method of eliciting latent
factors implicit in the sequence of items consumed by a user over time by
listening to a song or watching a movie. This is possible because a song
playlist created by a user is indicative of the user’s preference for the songs
in the sequence in a given session. We examine the effectiveness of
sequential LSTM and Autoencoder LSTM in learning the latent
characteristics present in the session sequence of consumable online
items, exhibiting the usefulness derived for enhanced recommender
system.

Keywords: Autoencoder-LSTM, Sequential-LSTM, Session, Item
sequence, Latent-factors.