The code has been tested running under Python 3.8.10. The required packages are follows:

• numpy==1.18.5
• tqdm==4.61.2
• pandas==1.3.1
• scikit_learn==1.0.2
• tensorflow==2.4.1

The instruction of commands has been clearly stated in the shell scripts: We uploaded some shell scripts as a reference to run the code, however, the pathes should be modified accordingly.

• run_pretrain.sh:

pretrain ckpts using the first 30 days information. we also provide the model checkpoints to reproduce the results on the public Criteo30d.

• run_base(_1d).sh

obtain baseline results under streaming setting.

• run_defuse.sh

obtain our DEFUSE results.

The criteo dataset is available at https://drive.google.com/file/d/1x4KktfZtls9QjNdFYKCjTpfjM4tG2PcK/view?usp=sharing

A preprint version of this paper is available at https://arxiv.org/abs/2202.06472.pdf