This repository is an official implementation of PerPEFT, a personalized parameter-efficient fine-tuning method for multimodal recommendation.

• Title: PerPEFT: Personalized Parameter-Efficient Fine-Tuning of Foundation Models for Multimodal Recommendations
• Authors: Sunwoo Kim, Hyunjin Hwang, Kijung Shin
• Venues: The Web Conference 2026 (WWW 2026)
• Affiliation: KAIST AI

In this work, we use the four datasets from https://amazon-reviews-2023.github.io/.

The entire dataset, including (1) user-item interactions, (2) item images, and (3) item titles, is presented in the link below:

• Link: https://www.dropbox.com/scl/fo/olpz13hyfcdn5jg6a4tzy/AN1Na5w_ySyO4nnYFyYnRpE?rlkey=5i6iyloeq9fpa48tz29ztmaqj&st=ibnulph8&dl=0
• Description: Refer to the README.txt file within the link.

Once you download the datasets in the above link, all files will be downloaded in mmrec_datasets.zip file. Then, locate the zip file in your directory, and run the file below to make the correct dataset directories:

python3 unzip_data.py

Our implementation is conducted upon the following key packages:

python == 3.8.20
pytorch == 2.2.0+cu121
numpy == 1.26.3
peft == 0.18.1
transformers == 4.49.0

Note that several other versions are also compatible. Feel free to leave any issues caused by the package error.

In this work, we use three PEFT modules: (1) LoRA, (2) (IA)3, and (3) IISAN.

• For LoRA (standard PEFT module), one can use the code below:

python3 perpeft_lora_ia3.py --dataset sports_outdoors --device cuda:0 --peft_type lora --wdecay 1e-4 --lr 1e-4

• For (IA)3 (the strongest PEFT module in our setting), one can use the code below:

python3 perpeft_lora_ia3.py --dataset sports_outdoors --device cuda:0 --peft_type ia3 --wdecay 1e-4 --lr 1e-4

• For IISAN (the fastest PEFT module in our setting), one can use the code below:

python3 perpeft_iisan.py --dataset sports_outdoors --device cuda:0 --wdecay 1e-4 --lr 1e-4

Description for each hyperparameter is as follows:

• --dataset indicates the target dataset one aims to use. One can choose: (1) sports_outdoors, (2) toys_games, (3) beauty_care, and (4) arts.
• --device indicates the target GPU device one aims to use.
• --wdecay indicates the coefficient for the weight normalization of the (1) PEFT modules and (2) backbone SASRec.
• --lrindicates the learning rate of the model for the training.

One can refer to perpeft_hyperparameter.pickle for the optimal hyperparameter configuration of each model on each dataset.

• If you have any questions regarding our paper and/or code, feel free to reach me out!
• You can either (1) contact me via kswoo97@kaist.ac.kr or (2) leave questions in the GitHub issue.