Parallel Ranking of Ads and Creatives in Real-Time Advertising Systems
Authors: Zhiguang Yang, Liufang Sang, Haoran Wang, Wenlong Chen, Lu Wang, Jie He, Changping Peng, Zhangang Lin, Chun Gan, Jingping Shao
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments to compare ours with two state-of-the-art approaches. The results demonstrate the effectiveness of our approach in both offline evaluations and real-world advertising platforms online in terms of response time, CTR, and CPM. |
| Researcher Affiliation | Industry | Zhiguang Yang, Liufang Sang*, Haoran Wang, Wenlong Chen, Lu Wang, Jie He, Changping Peng, Zhangang Lin, Chun Gan, Jingping Shao JD.com zgyang1996@gmail.com, sangliufang@jd.com, wanghaoran35@jd.com, chenwenlong17@jd.com, wanglu319@jd.com, hejie67@jd.com, pengchangping@jd.com, linzhangang@jd.com, ganchun1@jd.com, shaojingping@jd.com |
| Pseudocode | Yes | Algorithm 1: Evaluation Metrics NSCTR Input: impression data I with Ad, Creative, and y(click or not), creative ranker fcr Output: NSCTR 1: Let impressions 0; 2: Let clicks 0; 3: Let {Imp A1, ..., Imp AM} 0; 4: Let {Imps A1, ..., Imps AM} 0; 5: Let {Clks A1, ..., Clks AM} 0; 6: for all impression {(Am, Cn, y)i}I i=1 do 7: Imp Am Imp Am + 1; 8: for Ck in Creatives {C1, ..., CN} Given in Am do 9: Get predicted scores yk = fcr(Am, Ck); 10: end for 11: Choose the creative Ck = argmax(y1,...,y K); 12: if Ck = Cn then 13: Imps Am Imps Am + 1; 14: Clks Am Clks Am + y; 15: end if 16: end for 17: impressions = P{Imp A1, ..., Imp AM}; 18: clicks = P{ Clks A1 Imp A1 Imps A1 , ..., Clks AM Imp AM Imps AM }; 19: NSCTR = clicks/impressions; 20: return NSCTR; |
| Open Source Code | No | The paper does not provide any explicit statement about making its source code open, nor does it include a link to a code repository. |
| Open Datasets | No | Experiments are conducted on a log dataset gathered from a real-world ad system from May 1st to June 30th. We use the data of the first 60 consecutive days as the training set and that of June 30th as the test set. In total, There were about 18 billion training samples and 300 million test samples. |
| Dataset Splits | No | The paper states, "We use the data of the first 60 consecutive days as the training set and that of June 30th as the test set," but does not explicitly describe a separate validation dataset split with percentages or sample counts. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running experiments, such as GPU models, CPU models, or memory specifications. |
| Software Dependencies | No | The paper mentions software components like 'Re LU activation', 'sigmoid output layer', and 'Adagrad optimizer', but it does not specify any version numbers for these or any other software dependencies. |
| Experiment Setup | Yes | The DCN has 4 hidden layers with sizes of 512 512 256 128. In contrast, the CR model uses simpler features and network structures, including 11 user features, 5 ad features, creative ID features, and content features. The embedding dimension is 4. The MLP for CR has 3 hidden layers with sizes of 128 64 32. The pctrad embedding size K is set as 8192 and the dimension D is 128 by default. Both models use Re LU activation and sigmoid output layer to bound predictions within (0, 1). The batch size is 512 with Adagrad optimizer at a 0.05 learning rate. As usual in ranking models, the training epoch is set to 1, and we did not use the dropout strategy. |