alan-j-lin/yt_monetization_improvement

Metis Project 5: Improving Youtube's Ad-Revenue Program for Sex Educators

yt_monetization_improvement's Issues

MVP Feedback

This is a great project idea. I'm sure you're already looking into the librosa library to generate audio features to use in machine learning.

I recommend using both NLP features as well as audio features and using an ensemble to tune the weights. You can also look at Mattie's work from last cohort that leveraged some of the same concepts as you (NLP + audio). https://github.com/mterzolo/first-impressions. I think there will be a lot of signal in using the NLP and audio features only, but if we hit any roadblocks we can look at which videos it's failing and brainstorm how to move forward.

I think treating each 30 second sample as an individual point makes most sense. I do not know what YouTube does algorithm-wise to classify videos as explicit material, so this casts a wider net.

I'm assuming you're using an AWS instance with a lot of memory to store your video. How many videos have you download for your dataset?