Addressing Cold-start Problem in Click-Through Rate Prediction via Supervised Diffusion Modeling

CMD

The following command reproduce our results on the TaobaoAD dataset as reported in the main paper:

base=deepfm
warm=diff
epochs=2
seed=42
T=100
ts=10
iter=2
run=10
diff_coef=1
python3 main.py --dataset_name taobaoAD --model_name ${base} \
    --warmup_model ${warm} \
    --diff_epochs ${epochs} \
    --diff_iters ${iter} \
    --T ${T} \
    --timesteps ${ts} \
    --noise_min 0.0001 \
    --noise_max 0.0001 \
    --noise_scale 0.1 \
    --diff_coef ${diff_coef} \
    --run ${run} --seed ${seed}

Note

In our experiments, we observed that employing a larger position embedding in the U-Net enhanced the performance of DIFF on the ML data. However, increasing the size of the position embedding had a negligible effect on performance for the other datasets. We configured the position embedding to be twice the size of the ID embedding for a balance.

Acknowledgement

This code is based on CVR.

Citation

If you find our method useful, please kindly cite our paper.

@misc{zhu2025csdm,
      title={Addressing Cold-start Problem in Click-Through Rate Prediction via Supervised Diffusion Modeling}, 
      author={Wenqiao Zhu and Lulu Wang and Jun Wu},
      year={2025}
}