Unveiling Hero Creatives: Predicting and Understanding Key Features in Video-Audio Content through Machine Learning

No Thumbnail Available

URL

Journal Title

Journal ISSN

Volume Title

School of Science | Master's thesis

Date

2024-10-14

Department

Major/Subject

Data Science

Mcode

Degree programme

Master's Programme in ICT Innovation

Language

en

Pages

64

Series

Abstract

Creative videos are an important component in the digital marketing domain, which enables marketing managers to convey complex visual and auditory information via diverse formats to acquire targeting customers. Normally, those videos who perform extremely well and acquire a large amount of valuable customers are called hero creatives. This study is initiated to help the collaborative partner company to better identify hero creatives instead of testing all candidate creative videos in advertising networks using real money. Besides, a derivative second objective is to understand key features from previous hero creatives. For this purpose, firstly, data cleaning and feature engineering on the original creative videos and audios inside are conducted to extract useful information. Secondly, a set of machine learning models is tested and evaluated, including linear models such as Logistic Regression, ensemble models such as XGBoost etc. and Convolutional Neural Networks (CNN)-based models. The experiment results show that only Logistic Regression and one specific augmented CNN-based model turn out to surpass the performance of the baseline method, which is a simple 3-line rule based on previous domain knowledge. Both methods balanced and improved recall both on positive and negative samples, leading to macro recalls of 63% and 69% respectively whereas 58% for the baseline method. Last but not the least, important features are identified and discussed for both methods. For example, audio features do not stand out when feature importance is checked, which may suggest that background music or the creative videos do not play an important role in determining a hero creative.

Description

Supervisor

Savioja, Lauri

Thesis advisor

Carrasco, Jorge

Keywords

video classification, binary classification, machine learning, audio content analysis, digital marketing, convolutional neural network

Other note

Citation