This is a repository of the experiment code supporting the paper Real-Time Bidding by Reinforcement Learning in Display Advertising in WSDM 2017.
For the demo of the small-scale evaluation, please run under the folder of scripts:
$ bash small-scale.shor run under the folder of python:
$ python3 bid_ss.pyAfter running, you could get the performance table printed in the console like:
setting objective auction impression click cost win-rate CPM eCPC
ipinyou, camp=1458, algo=ss_mdp, N=1000, c0=0.03125 11 350000 52757 11 748502 15.07% 14.19 68.05
ipinyou, camp=1458, algo=mcpc, N=1000, c0=0.03125 121 350000 29312 121 752015 8.37% 25.66 6.21
ipinyou, camp=1458, algo=lin_bid, N=1000, c0=0.03125 243 350000 30152 243 564282 8.61% 18.71 2.32
ipinyou, camp=1458, algo=rlb, N=1000, c0=0.03125 271 350000 26292 271 687794 7.51% 26.16 2.54
ipinyou, camp=2259, algo=ss_mdp, N=1000, c0=0.03125 7 350000 63217 7 1008842 18.06% 15.96 144.12
ipinyou, camp=2259, algo=mcpc, N=1000, c0=0.03125 7 350000 19571 7 1017600 5.59% 52.00 145.37
ipinyou, camp=2259, algo=lin_bid, N=1000, c0=0.03125 8 350000 42781 8 1002204 12.22% 23.43 125.28
ipinyou, camp=2259, algo=rlb, N=1000, c0=0.03125 11 350000 46621 11 1001535 13.32% 21.48 91.05
ipinyou, camp=2261, algo=ss_mdp, N=1000, c0=0.03125 9 343862 81995 9 960409 23.85% 11.71 106.71
ipinyou, camp=2261, algo=mcpc, N=1000, c0=0.03125 7 343862 25217 7 963167 7.33% 38.20 137.60
ipinyou, camp=2261, algo=lin_bid, N=1000, c0=0.03125 8 343862 53249 8 912583 15.49% 17.14 114.07
ipinyou, camp=2261, algo=rlb, N=1000, c0=0.03125 10 343862 59163 10 961293 17.21% 16.25 96.13
ipinyou, camp=2821, algo=ss_mdp, N=1000, c0=0.03125 19 350000 65310 19 968224 18.66% 14.83 50.96
ipinyou, camp=2821, algo=mcpc, N=1000, c0=0.03125 10 350000 21217 10 976485 6.06% 46.02 97.65
ipinyou, camp=2821, algo=lin_bid, N=1000, c0=0.03125 19 350000 39136 19 753469 11.18% 19.25 39.66
ipinyou, camp=2821, algo=rlb, N=1000, c0=0.03125 23 350000 49261 23 964767 14.07% 19.58 41.95
ipinyou, camp=2997, algo=ss_mdp, N=1000, c0=0.03125 80 156063 40395 80 306637 25.88% 7.59 3.83
ipinyou, camp=2997, algo=mcpc, N=1000, c0=0.03125 48 156063 14752 48 307751 9.45% 20.86 6.41
ipinyou, camp=2997, algo=lin_bid, N=1000, c0=0.03125 71 156063 32208 71 203610 20.64% 6.32 2.87
ipinyou, camp=2997, algo=rlb, N=1000, c0=0.03125 78 156063 39680 78 304375 25.43% 7.67 3.90
ipinyou, camp=3358, algo=ss_mdp, N=1000, c0=0.03125 9 300928 34289 9 857285 11.39% 25.00 95.25
ipinyou, camp=3358, algo=mcpc, N=1000, c0=0.03125 81 300928 12424 81 867133 4.13% 69.79 10.71
ipinyou, camp=3358, algo=lin_bid, N=1000, c0=0.03125 202 300928 8465 202 517007 2.81% 61.08 2.56
ipinyou, camp=3358, algo=rlb, N=1000, c0=0.03125 212 300928 12970 212 836097 4.31% 64.46 3.94
ipinyou, camp=3386, algo=ss_mdp, N=1000, c0=0.03125 6 350000 51384 6 834115 14.68% 16.23 139.02
ipinyou, camp=3386, algo=mcpc, N=1000, c0=0.03125 15 350000 17839 15 840473 5.10% 47.11 56.03
ipinyou, camp=3386, algo=lin_bid, N=1000, c0=0.03125 37 350000 9279 37 213219 2.65% 22.98 5.76
ipinyou, camp=3386, algo=rlb, N=1000, c0=0.03125 39 350000 29142 39 830911 8.33% 28.51 21.31
ipinyou, camp=3427, algo=ss_mdp, N=1000, c0=0.03125 4 350000 52338 4 879547 14.95% 16.81 219.89
ipinyou, camp=3427, algo=mcpc, N=1000, c0=0.03125 48 350000 19622 48 885400 5.61% 45.12 18.45
ipinyou, camp=3427, algo=lin_bid, N=1000, c0=0.03125 160 350000 17664 160 679864 5.05% 38.49 4.25
ipinyou, camp=3427, algo=rlb, N=1000, c0=0.03125 171 350000 21678 171 851729 6.19% 39.29 4.98
ipinyou, camp=3476, algo=ss_mdp, N=1000, c0=0.03125 17 350000 51293 17 858971 14.66% 16.75 50.53
ipinyou, camp=3476, algo=mcpc, N=1000, c0=0.03125 36 350000 16254 36 865811 4.64% 53.27 24.05
ipinyou, camp=3476, algo=lin_bid, N=1000, c0=0.03125 113 350000 7915 113 352671 2.26% 44.56 3.12
ipinyou, camp=3476, algo=rlb, N=1000, c0=0.03125 113 350000 18282 113 855048 5.22% 46.77 7.57
Note these results are produced from a subset (the first 350,000 lines of each campaign in iPinYou) under T = 1000 and c0 = 1/32. For the full small-scale evaluation and large-scale evaluation, please first check the GitHub project make-ipinyou-data for pre-processing the iPinYou data. After downloading the dataset, by "make all" you can generate the standardised data.
