Addressing Cold-Start Problem in Click-Through Rate Prediction via Supervised Diffusion Modeling
Authors: Wenqiao Zhu, Lulu Wang, Jun Wu
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform extensive experiments on three CTR prediction benchmark datasets to validate the effectiveness of our proposed method. In a nutshell, the contributions of this work include: ... Extensive experiments are conducted on three public benchmark datasets, and the results show that CSDM outperforms existing cold-start methods in CTR predictions. |
| Researcher Affiliation | Industry | Hi Think Research EMAIL |
| Pseudocode | No | The paper describes the model and processes using mathematical equations and textual explanations, but it does not include a distinct 'Pseudocode' or 'Algorithm' block. |
| Open Source Code | Yes | Code https://github.com/WNQzhu/CSDM |
| Open Datasets | Yes | We evaluate our method on three publicly available datasets: Movie Lens-1M 1, Taobao Display Ad Click 2, and CIKM 2019 EComm AI 3. 1http://www.grouplens.org/datasets/movielens/ 2https://tianchi.aliyun.com/dataset/dataDetail?dataId=56 3https://tianchi.aliyun.com/competition/entrance/231721 |
| Dataset Splits | Yes | We divided the datasets into several groups following (Zhu et al. 2021) to assess the performance of our proposed method in both the cold-start and warm-up phases. First, we divide the items into two groups based on their frequency using a threshold N. Items with a frequency greater than N are classified as old items, while those with a lower frequency are considered new items. The threshold N is set to 200 for Movie Lens-1M, 2000 for Taobao AD, and 200 for CIKM 2019, ensuring the ratio of new items to old items is approximately 8:2, which mirrors a long-tail distribution as described in (Chen et al. 2020). We further divide the new item instances, sorted by timestamps, into four groups: warm-a, warm-b, warm-c, and a test set. The first 3 K instances are distributed evenly among warm-a, warm-b, and warm-c, with the remainder allocated to the test set. The value of K is set to 20 for Movie Lens-1M, 500 for Taobao AD, and 50 for CIKM 2019, respectively. |
| Hardware Specification | Yes | The tests were conducted on a single A800 GPU with a batch size set to 2048. |
| Software Dependencies | No | The paper mentions specific optimizers and network architectures (e.g., Adam optimizer, U-Net) but does not provide specific version numbers for any software libraries or programming languages used. |
| Experiment Setup | Yes | The embedding size for all features is set to 16 for compared methods. Additionally, the MLPs in the backbone models utilize two dense layers, each with 16 units. We set the learning rate to 0.001 for all methods, and the minibatch size is set to 2048 for Movie Lens-1M and Taobao AD, and 4096 for CIKM 2019. All methods are optimized using the Adam optimizer (Kingma and Ba 2015) on shuffled samples. We set T = 100 for the total number of forward steps. The parameters ρ and s are searched from the sets {0.001, 0.01, 0.1, 1} and {5, 10}, respectively. We define the sequences {αt} and {ct} using a hyper-parameter β = 10 5 for all experiments. ... we adopt dropout with p = 0.5 in the U-Net. |