In this chapter, we briefly presented the concept of recommendation systems and how collaborative filtering works. We then presented how neural networks can be utilized in order to avoid explicitly defining rules that dictate how unrated items would be rated by a user, using embedding layers and dot products. Following that, we showed how the performance of these models can be improved if we allow the networks to learn how to combine the embedding layers themselves. This gives the models considerably higher degrees of freedom without drastically increasing the number of parameters, leading to considerable increases in performance. Finally, we showed how the same architecture—with variable numbers of embedding features—can be utilized in order to create base learners for a stacking ensemble. In order to combine the base learners, we utilized a Bayesian ridge regression, which resulted in better results than any individual base learner.

This chapter serves as an introduction to the concept of using ensemble learning techniques for deep recommendation systems, rather than a fully detailed guide. There are many more options that can lead to considerable improvements in the system. For example, the usage of user descriptions (rather than indices), additional information about each movie (such as genre), and different architectures, can all greatly contribute to performance improvements. Still, all these concepts can greatly benefit from the usage of ensemble learning techniques, which this chapter adequately demonstrates.

In the next and final chapter, we will use ensemble learning techniques in order to cluster data from the World Happiness Report as we try to uncover patterns in the data.