Self-training with Noisy Student for Click-Through Rate Prediction

CONTENIDO DE LIBRE ACCESO

Este contenido es de libre acceso. Solo haz clic en el siguiente botón.

Descripción: Abstract: Improved click-through rate (CTR) prediction is a task of increasing importance in the digital world, especially in the current e-commerce boom, where dozens of billions of dollars are spent every year in online advertising. Advertisers usually have a wealth of unlabelled data that has historically been ignored, since machine learning classifiers need labelled data. This work explores the use of semisupervised learning techniques to take advantage of the availability of unlabelled data in an attempt to improve CTR prediction performance. Gradient boosting methods and deep neural networks are evaluated with several noising/regularization techniques. Most combinations fail to surpass a fullysupervised baseline, except a deep neural network that showed a slight increase in performance, albeit not high enough to warrant deployment. The results show that models used in CTR prediction with self-training are quite sensitive to hyperparameters and particularly prone to overfitting.
Citación recomendada (normas APA)
: Nicolás Martínez Baquero, "Self-training with Noisy Student for Click-Through Rate Prediction", Colombia:-, 2022. Consultado en línea en la Biblioteca Digital de Bogotá (https://www.bibliotecadigitaldebogota.gov.co/resources/3711718/), el día 2025-05-20.

Información adicional