Predicting clicks in online display advertising with latent features and side-information

TL;DR

This paper reviews a click-through rate prediction method combining collaborative filtering, matrix factorization, and side-information, demonstrating that latent features can modestly improve prediction performance in real-world online advertising scenarios.

Contribution

It provides detailed modeling and experimental insights into a CTR prediction approach that fuses collaborative filtering with side-information, validated on real-world datasets.

Findings

Latent features improve CTR prediction performance.

Small but statistically significant performance gains.

Validated on multiple real-world datasets.

Abstract

We review a method for click-through rate prediction based on the work of Menon et al. [11], which combines collaborative filtering and matrix factorization with a side-information model and fuses the outputs to proper probabilities in [0,1]. In addition we provide details, both for the modeling as well as the experimental part, that are not found elsewhere. We rigorously test the performance on several test data sets from consecutive days in a click-through rate prediction setup, in a manner which reflects a real-world pipeline. Our results confirm that performance can be increased using latent features, albeit the differences in the measures are small but significant.