• Annoying message interfere us to view the comment of youtube
• Make a SPAM classifier to filter the annoying comment
• Eventually, we will retrieve a pure social media

• Preprocessing
  • Use encoding UTF-8 to read the data
  • Split the comment by space to feed the D2V
• Doc2Vec
  • Input: many different length comment
  • Output: fixed length vectors e.g. $V_{100\times1}$
  • The vector size is dominated by max of comment length
• DNN
  • Input: vector of comments
  • Output: spam or ham
  • Loss function: cross entropy of sigmoid
  • Optimizer: Adaptive Moment Estimation
  • Tuning the parameters e.g. batch size, learning rate

• Demo

  $ python spamClassifier.py [--data=Youtube03-LMFAO.csv] [--text="I am spam"]
  

  • --data: file you wanna train
  • --text: comment you wanna classify

• Performance

  

• Furthermore

  

  • Compare the ACC on identical model
  • LMFAO is close to Eminem and Shakira

Github