CTR prediction Result - Githubissues

hzn666 / RLBid_EA

This is a repository of the experiment code supporting the paper Real-time Bidding Strategy in Display Advertising: An Empirical Analysis.

Apache License 2.0

24 stars 9 forks source link

CTR prediction Result #1

Open franktseng0718 opened 1 year ago

franktseng0718 commented 1 year ago

I've tried ctr prediction with LR in this repo. I also refered to some paper seeing their AUROC results on 1458, 3358, 3427, 3476 can easily surpass 90 even 95, but I only reached 8X. I also haved tuned many parameters but can't get better results.

Is anyone know how to get better results?

hzn666 commented 1 year ago

This repo intends to study bidding strategies so the performance requirements of the click-through rate prediction model are not high. Regarding your question, I will answer you from the following aspects：

The LR model in RLBid_EA/ctr/model.py is implemented using PyTorch. Strictly speaking, it is different from other LRs.
The features used to construct the dataset may differ from other papers. For details, you can check RLBid_EA/ctr/ctr_data.py.

franktseng0718 commented 1 year ago

Thanks for yor answer. I'm now trying run drlb's code, but I found that cannot converge after even 1000 episodes. The results of each episode fluctuate greatly, and most of them are even worse than LIN. The result I got is like this: But the LIN's method's result is like: