Code and Data for paper Efficient Optimal Selection for Composited Advertising Creatives with Tree Structure

Data

We provide the generated synthetic data for testing in 'data' directory.

• 'tree_constraint.txt' details the constraints between ingredients with adjacency matrix.
  • '1' represents feasible connections
  • '0' represents illegal connections
• 'tree_struct2.txt' details the Ingredient Tree
• 'ctr_new.txt'/'ctr_new2.txt' details the simulated data, where each row is formulated as 'ID + element list + Generated CTR'

Online Raw Data

'data/ctr_online.txt' has same format with ctr_new.txt/ctr_new2.txt

The element list contains

• ID for the template : {0}
• ID for the background : {0: dark, 1:light}
• ID for picture sizes : {0: 88%, 1:91%, 2:94%, 3:97%, 4:100%}
• ID for text color : {0,1,2,3,4,5,6,7}, 0-3 for the dark background, 4-7 for the light background
• ID for text font : {0,1,2,3}

Code

Run an example with python run.py -b 1000 -p 1000000 -j AES -r 20

• 'run.py' is the main function and different configurations can be found here for tuning
• 'policy' directory contains different methods

Parameters

• -f file path of creatives
• -p total pv
• -r multiple running repetitions
• -b batch size(update intervals)
• -j method : EGreedy,thompson,ucb,IndEgreedy,Edge_TS,LinUCB,TEgreedy,Full_TS,Vertex_TS,MVT,AES(proposed method)
• -e parameter for egreedy
• -t EE type : 0,1,2
  • 0: regular exploration
  • 1: exploration after DP
  • 2: exploration during DP
• -a parameter for thompson methods

Generate Simulation Data

Run python gen_ctr.py

• the generated file is saved in "data/ctr_new2.txt"