SKAPP

This is a reference implementation of the SKAPP model described in "Improving Multimodal Social Media Popularity Prediction via Selective Retrieval Knowledge Augmentation", published at AAAI 2025.

Install

Use Python 3.12 and install PyTorch for your CUDA runtime, then:

python -m pip install -r requirements.txt

Download data

python src/download.py --dataset icip

Replace icip with smpd or instagram to use another dataset.

Downloads are verified and extracted to datasets/<dataset>/.

Alternatively, download the dataset ZIP from Google Drive and extract it into the datasets/ directory at the project root. For example, extracting skapp-icip.zip should give:

datasets/icip/
├── dataset.json
├── train.npz
├── valid.npz
└── test.npz

Original datasets:

• ICIP: Official | Our mirror
• SMPD: Official
• Instagram: Official

Train and evaluate

python src/train.py --dataset icip
python src/evaluate.py --dataset icip

Training runs once with seed 12 by default and selects checkpoints by validation MSE. Use --seeds followed by one or more distinct integers to customize the seeds. Models are saved to runs/icip/models/; evaluation outputs go to runs/icip/evaluation/. Use a fresh output directory for a new run, e.g., --output-dir runs/icip/experiment2.

Evaluation verifies saved validation predictions before evaluating the test split. result.json contains each seed's MSE, MAE and Spearman correlation, plus their mean and sample standard deviation. Predictions include sample IDs and labels.

The default device is cuda:0. Use --device cpu for small CPU runs, --data-root /path/to/datasets for another data location, or --output-dir for a custom run directory. Dataset features are loaded onto the selected device. See --help for options.

Recompute RRCP (optional)

The supplied RRCP was generated with auxiliary teachers trained on the training split and selected using validation data. To train your own teachers and rebuild RRCP from the packaged neighbors:

python src/train_teachers.py --dataset icip --kind all
python src/train_teachers.py --dataset icip --kind single
python src/build_rrcp.py --dataset icip
python src/train.py --dataset-path runs/icip/data-rebuilt --output-dir runs/icip/rebuilt-models
python src/evaluate.py --dataset-path runs/icip/data-rebuilt --training-dir runs/icip/rebuilt-models --output-dir runs/icip/rebuilt-evaluation

Rebuilding RRCP creates a separate dataset and preserves the downloaded package.

Citation

@inproceedings{xu2025improving,
  author = {Xu, Xovee and Zhang, Yifan and Zhou, Fan and Song, Jingkuan},
  title = {Improving Multimodal Social Media Popularity Prediction via Selective Retrieval Knowledge Augmentation},
  booktitle = {AAAI Conference on Artificial Intelligence (AAAI)},
  year = {2025},
  pages = {932--940},
  doi = {10.1609/aaai.v39i1.32078}
}

MIT License.