Overview

This is an auto-bidding training framework to help participants implement and evaluate their bidding strategies. This framework includes three modules: data processing, strategy training, and offline evaluation. Several Generative Models baseline strategies, such as Decision-transformer and Decision-diffuser, are included in the framework. Participants can utilize this framework to develop a well-trained auto-bidding strategy based on the training dataset. Since the auction system cannot be used for evaluation during offline training, participants can rely on the provided framework for a basic offline assessment to ensure the code implementation meets the competition requirements.

Dependencies

conda create -n nips-bidding-env python=3.9.12 pip=23.0.1
conda activate nips-bidding-env
pip install -r requirements.txt

Usage

Dataset Link

Due to the large size of the data file, it has been split into multiple parts for download.

https://alimama-bidding-competition.oss-cn-beijing.aliyuncs.com/share/aigb_track_data_period_7-8.zip

https://alimama-bidding-competition.oss-cn-beijing.aliyuncs.com/share/aigb_track_data_period_9-10.zip

https://alimama-bidding-competition.oss-cn-beijing.aliyuncs.com/share/aigb_track_data_period_11-12.zip

https://alimama-bidding-competition.oss-cn-beijing.aliyuncs.com/share/aigb_track_data_period_13.zip

https://alimama-bidding-competition.oss-cn-beijing.aliyuncs.com/share/aigb_track_data_trajectory_data.zip

Data Processing

We provide traffic granularity data and additional trajectory data used for model training. Download the traffic granularity data & trajectory data and place it in the biddingTrainENv/data/ folder. The directory structure under data should be:

biddingTrainEnv
|── data
    |── traffic
        |── period-7.csv
        |── period-8.csv
        |── period-9.csv
        |── period-10.csv
        |── period-11.csv
        |── period-12.csv
        |── period-13.csv
    |── trajectory
        |── trajectory_data.csv

Train Model

Decision-Transformer

Load the training data and train the DT bidding strategy.

python main/main_decision_transformer.py

Use the DtBiddingStrategy as the PlayerBiddingStrategy for evaluation.

bidding_train_env/strategy/__init__.py
from .dt_bidding_strategy import DtBiddingStrategy as PlayerBiddingStrategy

Decision-Diffusion

Load the training data and train the DD bidding strategy.

python main/main_decision_diffuser.py

Use the DdBiddingStrategy as the PlayerBiddingStrategy for evaluation.

bidding_train_env/strategy/__init__.py
from .dd_bidding_strategy import DdBiddingStrategy as PlayerBiddingStrategy

offline evaluation

Load the training data to construct an offline evaluation environment for assessing the bidding strategy offline.

python main/main_test.py

Appendix

Traffic granularity data format

The training dataset is derived from advertising delivery data generated via the auction system where multiple advertisers compete against each other. Participants can use this dataset to recreate the historical delivery process of all advertisers across all impression opportunities. The training dataset includes 7 delivery periods. Each delivery period contains approximately 500,000 impression opportunities and is divided into 48 steps. There are 48 advertisers competing for these opportunities. The dataset consists of approximately 170 million records, with a total size of 25G. The specific data format is as follows:

(c1) deliveryPeriodIndex: Represents the index of the current delivery period.
(c2) advertiserNumber: Represents the unique identifier of the advertiser.
(c3) advertiserCategoryIndex: Represents the index of the advertiser's industry category.
(c4) budget: Represents the advertiser's budget for a delivery period.
(c5) CPAConstraint: Represents the CPA constraint of the advertiser.
(c6) timeStepIndex: Represents the index of the current decision time step.
(c7) remainingBudget: Represents the advertiser's remaining budget before the current step.
(c8) pvIndex: Represents the index of the impression opportunity.
(c9) pValue: Represents the conversion action probability when the advertisement is exposed to the customer.
(c10) pValueSigma: Invalid variable, Constantly zero, please ignore.
(c11) bid: Represents the advertiser's bid for the impression opportunity.
(c12) xi: Represents the winning status of the advertiser for the impression opportunity, where 1 implies winning the opportunity and 0 suggests not winning the opportunity.
(c13) adSlot: Represents the won ad slot. The value ranges from 1 to 3, with 0 indicating not winning the opportunity .
(c14) cost: Represents the cost that the advertiser needs to pay if the ad is exposed to the customer.
(c15) isExposed: Represents whether the ad in the slot was displayed to the customer, where 1 implies the ad is exposed and 0 suggests not exposed.
(c16) conversionAction: Represents whether the conversion action has occurred, where 1 implies the occurrence of the conversion action and 0 suggests that it has not occurred.
(c17) leastWinningCost: Represents the minimum cost to win the impression opportunity,i.e., the 4-th highest bid of the impression opportunity.
(c18) isEnd: Represents the completion status of the advertising period, where 1 implies either the final decision step of the delivery period or the advertiser's remaining budget falling below the system-set minimum remaining budget.

Training data example

example-1

c1	c2	c3	c4	c5	c6	c7	c8	c9	c11	c12	c13	c14	c15	c16	c17
1	31	2	6500.00	27.00	5	5962.49	101000	0.0103542	0.2845	1	1	0.2702	1	0	0.1832
1	22	6	7000.00	38.00	5	5988.25	101000	0.0070297	0.2702	1	2	0.2154	1	1	0.1832
1	15	7	7000.00	42.00	5	6132.52	101000	0.0051392	0.2154	1	3	0.1832	0	0	0.1832
1	39	3	6000.00	30.00	5	5443.27	101000	0.0062134	0.1832	0	0	0	0	0	0.1832
1	43	9	7500.00	25.00	5	6421.81	101000	0.0045392	0.1099	0	0	0	0	0	0.1832

This example presents an impression opportunity involving the top five advertisers. The top three advertisers, numbered 31, 22, and 15, won the impression opportunity with the highest bids and were allocated to ad slots 1, 2, and 3, respectively. During this impression, slots 1 and 2 were exposed to the customer, while slot 3 remained unexposed. Consequently, ads in slots 1 and 2 need to pay 0.2702 and 0.2154, respectively. Additionally, the customer engaged in a conversion action with the ad in slot 2.

example-2

c1	c2	c3	c4	c5	c6	c7	c8	c9	c10	c11	c12	c13	c14	c15	c16	c17	c18
3	48	6	7500.00	40.00	1	7500.00	1	0.0032157	0	0.1345	0	0	0	0	0	0.1628	0
3	48	6	7500.00	40.00	1	7500.00	2	0.0146256	0	0.5852	0	0	0	0	0	0.6421	0
3	48	6	7500.00	40.00	1	7500.00	3	0.0054324	0	0.1924	1	1	0.1673	1	1	0.1454	0
3	48	6	7500.00	40.00	1	7500.00	4	0.0073145	0	0.2786	0	0	0	0	0	0.2862	0
…
3	48	6	7500.00	40.00	2	7341.25	20901	0.0076453	0	0.2856	0	0	0	0	0	0.3245	0
3	48	6	7500.00	40.00	2	7341.25	20902	0.0139234	0	0.5629	1	2	0	0	0	0.6782	0
3	48	6	7500.00	40.00	2	7341.25	20903	0.0077212	0	0.3045	0	0	0	0	0	0.3122	0
3	48	6	7500.00	40.00	2	7341.25	20904	0.0021341	0	0.0926	0	0	0	0	0	0.1151	0
…
3	48	6	7500.00	40.00	43	0.00	895201	0.0065274	0	0.0000	0	0	0	0	0	0.1243	1
3	48	6	7500.00	40.00	43	0.00	895202	0.0032125	0	0.0000	0	0	0	0	0	0.2986	1
3	48	6	7500.00	40.00	43	0.00	895203	0.0112986	0	0.0000	0	0	0	0	0	0.0932	1
3	48	6	7500.00	40.00	43	0.00	895204	0.0051678	0	0.0000	0	0	0	0	0	0.1687	1

This example presents a data sample illustrating an advertiser's bidding process across time steps within a delivery period. The advertiser has a budget of 7500, a CPA constraint of 40, and belongs to industry category 6. Throughout different time steps, the advertiser engages in bidding for every available impression and obtains the corresponding results. During this period, the advertiser's remaining budget decreases correspondingly. Additionally, the advertiser adjusts their bidding strategy based on prior performance, although this adjustment will not be directly evident in the data.

Trajectory data format

Trajectory data is converted from traffic granularity data. It records information for multiple advertisers over different time steps across multiple periods as (s, a, r, s'). you can refer to the code provided below.

python  bidding_train_env/dataloader/rl_data_generator.py

(c1) deliveryPeriodIndex: Represents the index of the current delivery period.
(c2) advertiserNumber: Represents the unique identifier of the advertiser.
(c3) advertiserCategoryIndex: Represents the index of the advertiser's industry category.
(c4) budget: Represents the advertiser's budget for a delivery period.
(c5) CPAConstraint: Represents the CPA constraint of the advertiser.
(c6) timeStepIndex: Represents the index of the current decision time step.
(c7) state: Represents the advertiser's state in this timeStep.
(c8) action: Represents the advertiser's action in this timeStep.
(c9) reward: Represents the advertiser's sparse reward(total conversion) in this timeStep.
(c10) reward_continuous: Represents the advertiser's continuous reward(The sum of the pValues of all exposed traffic) in this timeStep.
(c11) done: Represents the completion status of the advertising period, where 1 implies either the final decision step of the delivery period or the advertiser's remaining budget falling below the system-set minimum remaining budget.
(c12) next_state: Represents the advertiser's next state in this timeStep.

Reference

Reference for Decision Transformer Implementation： https://github.com/kzl/decision-transformer

Reference for Decision Diffusion Implementation： https://github.com/anuragajay/decision-diffuser/tree/main/code

get-david / aigb_for_bidding Goto Github PK

aigb_for_bidding's Introduction