Recommending What Video to Watch Next: A Multitask Ranking System

data in paper:

• video: video meta-data and video content signals as its representation
• context: user demographics, device, time, and location

Steps:

1. Generate candidtae models

2. ranking:

   • should be very efficient
   • model:
     • input: given a query, candidate, and context
     • output: predict the probability of user taking actions asuch as clicks, watches, likes and dismissals
   • measure: engagement objects: binary classifcation: user click; regress: time spent satisfaction: binary classficatioin: like or not; regression: rating
   • loss:
     • binary classification: cross entropy loss
     • regression task: squared loss

3. implicit bias: selection bias(ranking order decided by current system) => shallow tower => a scalar serving as a bias term to the final prediction of the main model

data used in this repo from here

** Note **: only for testing model structure, training data in ranking.py is manipulated and not correct.