Analysis of YouTube thumbnails : a deep neural decision forest implementation

• Main Author: Shankkar Magandran
• Other Authors: Ponnuthurai Nagaratnam Suganthan
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/140188
• Institution: Nanyang Technological University

The advent of social media has transformed the way in which content is consumed by the masses. Driven by people’s interest in creating and sharing information, this phenomenon has propelled social media platforms such as YouTube as a dominant force impacting everyday life. The explosion of content on YouTube in recent decades has spurred growing competition amongst content creators, jostling for the attention from users. This project looks to create a social media analytics tool capable of helping content creators on YouTube better captivate their audience. The overall approach is to target the thumbnail image of a YouTube video since it is the first point of contact between the content and the user. By training several machine learning algorithms to approximate the number of views, likes, dislikes and comments a video would garner based on its thumbnail image, we aim to construct a tool capable of estimating the level of interaction a video would receive based on the design of its thumbnail. We anticipate that such a tool would be immensely useful to YouTube content creators as it offers them a way to evaluate the attractiveness of their thumbnail design. In the course of this project, the first phase illustrates in detail how the entire YouTube dataset used in this project was created from scratch. We then proceed on to construct and train popular Convolutional Neural Network architectures. This is followed by the employment of traditional machine learning algorithms like Random Forests. We also explore a fully differentiable combination of the Convolutional Neural Network and Random Forest known as Deep Neural Decision Forests which is a core requirement in this Final Year Project. The project concludes with the construction of a Graphical User Interface with the above mentioned models serving as its predictive engines. The challenges encountered in the execution of this project are discussed. This report also presents significant results and evaluates them against that of related publications while concurrently examining the limitations of some of the algorithms in performing the stated approximations.