Learning framework for eCPM prediction in online advertisement

Abstract

Nowadays, online advertising becomes the most popular tool for buying and selling all products. Especially after the appearance of COVID-19, online advertising has become a big challenge for all web advertisers and web publishers to increase their revenues. All web advertisers seek to buy empty ad slots at low prices. At the same time, all web publishers seek to sell their ads spaces and achieve high revenues. Selling publishers ad space happens in real-time through Real-Time Bidding (RTB) platform. So, there is no specific price for selling publisher’s ads spaces. The ad space price is specified based on the market demand at a time. The publishers wish if they can know the optimal ad space price that achieves the highest revenue. In this thesis, we are trying to help publishers to increase their revenues by trying to predict the average Effective Cost per Mille (eCPM) price for ads spaces. We are trying to forecast the average eCPM around certain date for the next 7, 14, and 30 days from today based on the eCPM history. We used different models trying to achieve this goal. In this thesis, we used K-Nearest Neighbors (KNN), Multi Layer Perceptron (MLP), and Long Short-Term Memory (LSTM) models on four different datasets trying to reach the best model accuracy.

Date	11 Feb 2022
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Pascal Giard (Supervisor)