Thumbnail Evaluator

• 

  The Thumbnail Evaluator Webbapp.

YouTube is a platform where click-through rate is a crucial factor in the success of a channel. This project aims at helping people create thumbnails for their videos so that they can be effective in reaching their audience.

This thumbnail evaluator looks at an image and predicts for you the success of it as a video thumbnail on YouTube.

Inspiration

It is common knowledge that YouTube's infamous "algorithm" relies on some combination of machine learning solutions to decide what videos to show to users on their feeds. I wanted to see if I can train a model to replicate a small but important part of this "algorithm": the thumbnail evaluator.

How I built it

I experimented with increasingly deep and complex convolutional neural network models up until I reached the Inception models and finally settled on a custom reduced version of the Xception model for its exceptional learning capabilities.

The main challenges I faced during this project were over-fitting and hitting plateaus in validation accuracy. Neural networks in general tend to over-fit very easily if one doesn't make sure to properly apply regularization techniques, but even once I had applied all the regularization that made sense to my model, I still was hitting constant plateaus of the model never reaching satisfactory validation accuracies.

Results

I am not very satisfied with the final accuracies I got with my model, but the constant struggle to try to get it to go up has pushed me to do much more research on CNNs, learn much more about regularization and gain a much better intuition on the whole process of training CNNs than I would have if the project had gone more smoothly. I am the most satisfied with the learning experience that I got through this project.

Next Steps

What's next for this project is to experiment with using pre-trained image feature extraction models to guarantee a certain level of understanding of the image in my model. I would also try to find a larger dataset to train with that would contain only more recent thumbnail images as the YouTube "thumbnail game" has evolved a lot since 2008 and settled on more consistent strategies around 2016. Ideally, I would also score thumbnails using click-through rate statistics instead of view count but this information isn't available through the YouTube APIs.

Built With

• flask
• keras
• pipenv
• python