KDSR (Knowledge Distillation for Sequential Recommendation)

This is the PyTorch implementation for KDSR proposed in the paper Lightweight Modality Adaptation to Sequential Recommendation via Correlation Supervision (https://arxiv.org/abs/2401.07257v1), ECIR, 2024. The model name in the code is VLGraph (Visio-linguistic Graph).

Preprocess

• First Step. Download the raw data from http://jmcauley.ucsd.edu/data/amazon/links.html
• Second Step. Create an "image" folder under the downloaded datafolder, and scrawl the images, and save them into the "image" folder.
• Third Step. Check config/preprocess.yaml for starting dataset preparation.
• Fourth Step. Enter the preprocess/dataset_name/ folder, run python image_feature_extractor.py、python text_feature_extractor_t5.py、python process_dataset.py step by step

Note that the preprocess provide several optiosn: to generate datasets for collaborative filtering task (dataset_cf) sequential recommendation task (dataset_sr), multimodal sequential recommendation task (dataset_mmsr)

How to Run

Before running the program, you need to check the model configuration file in config/model.yaml, and make sure you are using the correctly preprocessed dataset folder.

Process "python main.py dataset_name", such as,

python main.py beauty

The results will be saved into log/.

Others

You can also find an example of our execution log as shown in running_records.log, which reports the running records and results of our model on the beauty dataset.