Revenue-Incentive Tradeoffs in Dynamic Reserve Pricing
Authors: Yuan Deng, Sebastien Lahaie, Vahab Mirrokni, Song Zuo
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We further empirically evaluate the tradeoffs on synthetic data as well as real ad auction data from a major ad exchange to verify and support our theoretical findings. ... We empirically evaluate the revenue-incentive tradeoffs of our class of nonlinear policies on both synthetic data and real ad auction data from a major ad exchange. The experiments verify and support our theoretical findings, and uncover nonlinear policies with significantly better revenue-incentive tradeoffs than both linear policies and the exponential mechanism for realistic bid distributions. |
| Researcher Affiliation | Industry | 1Google Research, New York City, NY, USA. |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any concrete access information (e.g., specific repository link, explicit code release statement) for source code. |
| Open Datasets | No | We first present synthetic-data experiments... We consider a standard log-normal distribution F1... and a shifted log-normal distribution F2... We use a log-normal distribution because it is commonly used to model bid and value distributions in practice (Lahaie & Pennock, 2007; Ostrovsky & Schwarz, 2011; Thompson & Leyton-Brown, 2013; Golrezaei et al., 2019). ... Our dataset consists of auction records of bids placed on queries from a single publisher over two consecutive days. We focus on a single bidder and only retain the auctions that involve this bidder, yielding over 100K auction records in total. The first day is used as the training set to compute the bid quantiles and their associated revenue performance as reserves under a second-price auction. The paper uses standard statistical distributions for synthetic data but does not provide a public dataset link or citation for them. For the real ad auction data, it describes using proprietary data ('from a major ad exchange') without providing public access or specific citation. |
| Dataset Splits | No | The paper mentions using a 'training set' and 'testing (next-day) data' for evaluation, but it does not specify explicit percentages or counts for these splits, nor does it mention a distinct validation set. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific ancillary software details, such as library names with version numbers, needed to replicate the experiment. |
| Experiment Setup | Yes | When simulating a policy, we use REV (i.e., the buyer s expected payment) as the weight function. By varying the parameter (in Definition 4.1), we obtain different combinations of (REVreserve/W, DIC) which yields a tradeoff curve that we plot. ... We compute the 100-quantiles, excluding the 100-th quantile (i.e., the maximum), and their revenues as the weights. ... We focus on a single bidder and only retain the auctions that involve this bidder, yielding over 100K auction records in total. The first day is used as the training set to compute the bid quantiles and their associated revenue performance as reserves under a second-price auction. Specifically, we compute the (5, 10, 25, 50, 75, 90, 95)-th bid quantiles and their counterfactual revenue. |