A technical report on a two-stage architecture for revenue-optimal ad impression allocation in operator-run mobile advertising: offline predictive modeling and linear-programming allocation planning, online table-driven decisioning with pacing, and a closed feedback loop.
- Read the paper (HTML): https://apartsinprojects.github.io/optimal-impression-allocation/
- PDF: https://apartsinprojects.github.io/optimal-impression-allocation/paper.pdf
| File | Description |
|---|---|
index.html |
The paper (GitHub Pages index) |
paper.pdf |
PDF rendering of the paper |
alloc_sim.py |
Monte-Carlo simulation comparing greedy serving with LP-planned allocation (Section 7 of the paper) |
pip install numpy scipy
python alloc_sim.pyTwenty replications of a synthetic market (60 audience segments, 25 campaigns, ~200,000 impressions per run). The planned policy achieves a mean 6.5% revenue lift and 33% expected-click lift over greedy serving on identical event streams, capturing 94% of the hindsight offline optimum. The harness verifies both policies against the hindsight bound in every replication and includes a uniform-probability control whose measured lift is zero.
h2_replay.py reconstructs an allocation market (with quota-stress, re-solved-plan, and tuned/warm dual-price policies) from the public Taobao display-ad dataset (Tianchi dataset 56 / Kaggle mirror) and replays greedy, LP-planned, and dual-price pacing policies under walk-forward dual-model scoring; h2_errorbars.py and h2_wf_analyze.py compute the closed-form error bars and fold-combined statistics; h2_cluster_boot.py runs the user-clustered bootstrap and h2_matched_action.py the matched-action outcome bracket reported in the paper.