Online Learning

  |   Source
Online Learning
  • Online learning algorithm is used when we have streamed input, having continuous data that didn’t stop. And what we like to do is set learning algorithm to keep learning as the data streamed and making learning algorithm making better and better decision live
  • Nowadays the website use learning algorithm to make better decision receiving continous input



  • Here we have a website where we have continuous data from our user
  • We have labeled when user accepted shipping service y= 1, or not y = 0
  • The goal is we optimize our asking price so user is willing to accept our shipping service
  • We include demographic of user, ori/dest/ and notice that we also include our asking price
  • So in doing this, we scrap our fixed set training example. Everythime there’s user purchase and his option whether or not they accept, we then directly update our learning algorithm
  • So instead of fixed training examples, we have one example and have continuous update.
  • This way we can handle continuous update and learning it.
  • Learning algorithm can be done effectively when we have major website, with a lot of users, and we can keep doing better. Whereas website with small users, don’t get learning much.
  • It’s better for small websites to just gather all examples to fixed, and learn from it.
  • Learning algorithm can also have other benefit, can adapt dynamically. For example if there’s economic change, or trends that making users more price-sensitive, learning algorithm can adapt very quickly


  • Here we have store that will provide 10 best guesses to user word search
  • We can input the feature X and the labeled whether the user click the link or not.
  • Often called predicted CTR (Click To Rate)
  • Another example is when we using as stated above
  • Basically others learning algorithm can be used to online learning, also collaborative filtering at product recommendation.
  • We can also just wait few days, then taking fixed training examples to learn from the data.
  • But these are algorithms used by major website, that doesn’t need to store the data(because it’s so huge). Instead just receiving training example, learn from it, and discard it. This way the website sdoesn’t need a lot of storage.


  • So here’s the online learning which really similar to stochastic GD, where we using one example
  • This is advantageous as we can adapt user’s preference, where the direction the user preferably go