Abstract
The growing diffusion of web-based services allows an increasingly large number of users to access
and consume online content. People access web pages, purchase items using e-commerce sites, watch
online content through streaming services, and interact within social networks and media. Understanding the factors that characterize user preferences and shape their future behavior is crucial to provide
users with a better experience through the recommendation of new content and data that users are likely to appreciate. Generally, providing enjoyable and quality recommendations is followed by user satisfaction, and, eventually more users and revenue for the streaming platform. Hence, these reasons make
recommender systems an application with a very high commercial impact.
Unlike a recommender system that models the global user preferences, which might be applicable
perhaps sometime later in the user life-span, in this thesis, we tackle a much different and a more
practically relevant problem – sequential recommendation. In sequential recommendation, we want to
capture the user’s short-term preferences more accurately so that we can predict what the user might
like right now rather than sometime later.
There have been numerous existing works for our concerned problem-statement, ranging from the
relatively simple nearest-neighbor collaborative filtering based methods to highly sophisticated recurrent
neural network-based approaches. In this thesis, we propose three methods, all for the task of sequential recommendation. The first two methods are suited only for next song recommendation scenarios,
whereas the third method applies to any item recommendation problem.
This thesis iterates in increasing order of complexity of proposed algorithms. We start by proposing
a novel, subsession-based method for next song recommendation. The proposed method uses a custom
similarity measure among subsessions to find closest subsessions to the active subsession and then takes
a majority vote to recommender new songs. Despite the method being highly competent, in this thesis,
we treat it as a great starting point for the subsequently proposed neural network-based methods.
Inspired from the previous method, we now propose a much more sophisticated deep neural architecture for better performance in the next song recommendation problem. Specifically, we use two recurrent
neural networks to model the short-term user preferences. At each time-step in the user consumption
sequence, we model the probability of each song occurring next. We learn all the model parameters by
maximizing the likelihood of the actually observed song.
Finally, not limiting ourselves to the domain of music recommendation, we propose an item recommendation algorithm, independent of the kind of items involved. Specifically, we propose a model that
extends variational autoencoders (VAEs) by exploiting the rich information present in the past preference history. We introduce a recurrent version of the VAE, where we pass the consumption sequence
through a recurrent neural network. At each time-step of the RNN, the sequence is fed through a series
of non-linear transformations. The output of these transformations models the probability distribution of
the most likely future preferences. We show that handling temporal information is crucial for improving
the accuracy of the VAE for our task of next item recommendation.
All three proposed methodologies are highly competent and out-perform all existing methods. In
both music recommendation and item recommendation scenarios, our methods are up to 51% and 52%
better, respectively, compared to the best existing method on benchmark music-recommendation and
item-recommendation datasets.
