Building Recommendation Systems in Python and JAX

Chapter 11. Personalized Recommendation Metrics

Having explored the powerful methodologies of MF and neural networks in the context of personalization, we are now equipped with potent tools to craft sophisticated recommendation systems. However, the order of recommendations in a list may have a profound impact on user engagement and satisfaction.

Our journey so far has primarily been focused on predicting what a user may like, using latent factors or deep learning architectures. However, the manner in which we present these predictions, or more formally, how we rank these recommendations, holds paramount significance. Therefore, this chapter will shift our gaze from the prediction problem and will unravel the complex landscape of ranking in recommendation systems.

This chapter is dedicated to understanding key ranking metrics including mean average precision (mAP), mean reciprocal rank (MRR), and normalized discounted cumulative gain (NDCG). Each of these metrics takes a unique approach toward quantifying the quality of our rankings, catering to different aspects of the user interaction.

We’ll dive into the intricacies of these metrics, unveiling their computational details and discussing their interpretation, covering their strengths and weaknesses, and pointing out their specific relevance to various personalization scenarios.

This exploration forms an integral part of the evaluation process in recommendation systems. It not only gives us a robust framework to measure the performance ...