efficiently-scaling-up-rtb-turn.md

From 0.5 Million to 2.5 Million: Efficiently Scaling up Real-Time Bidding by Jianqian Shen et al. ICDM 2015.

• Shen et al. proposes a hierarchical allocation to distribute bid requests and a logistic sigmoid to further filter bid requests
• each request can be hierarchically allocated based on features (e.g. device, publisher)
  • example hierarchy:
    1. publisher 1 2. site 1 3. device 1 3. device 2 2. site 2 3. device 1 3. device 2
    2. publisher 2 2. site 1 3. device 1 3. device 2 2. site 2 3. device 1 3. device 2
• each leaf node wants R_t requests, so a hierarchical filtering system is created, where each node has selection rate theta_i
• theta_i is a threshold for each request's utilization score, which is an estimation of how likely this request is useful
• Shen et al. proposes a logistic sigmoid to compute the utilization score, for 4 reasons:
  1. higher utilization gets higher score
  2. the "S" shape of sigmoid ensures score is in a limited range
  3. normalized
  4. relatively easy to balance exploration vs exploitation