Experimenting with NLP and machine learning. Based on thumbnail information on your YouTube Home page (signed out), i.e. your recommendations, predict what video the user will click on next.

First baseline: Scrape info, store in DB, use titles, train FastText, compute how similar the scraped titles are to everthing seen AND clicked so far.

Rright now only English and German, only Firefox, only FastText model.

pip3 install pandas numpy matplotlib seaborn sklearn nltk gensim selenium

• for german:

pip3 install HanTa

Download file https://github.com/wartaal/HanTa/blob/master/HanTa/morphmodel_ger.pgz

In a python terminal run:

import nltk
ntlk.download()

Select the corpora and packages options. Also enter the download path into CONFIG.

Download selenium driver for FF for you version: https://github.com/mozilla/geckodriver/releases

Enter the other data in CONFIG.

Call

./main

Click YTScrape, choose title from list and double click (or click button).

• Make notebook to visualize your YouTube "word bubble"
• Incorporate views, clicks, upload time
• Try soft cosine similarity
• Don't use average of clicked vids. - Make robust against "outlier clicks"
• don't use selenium. Use JS mokey script in "normal" browser.
• Fix issues:
  • speed everything up (scraping, browser and GUI startup)
  • Frequent very bad language guesses by nltk
  • Bad lemmatization
  • Covid section sometimes has different css selector
  • Upload time (e.g. Streamed x y ago) in different CSS element sometimes.