This repository contains the implementation code for the paper:

IGD: Token Decisiveness Modeling via Information Gain in LLMs for Personalized Recommendation

# Take the book dataset as an example 
# Download the dataset
wget https://datarepo.eng.ucsd.edu/mcauley_group/data/amazon_v2/categoryFiles/Books.json.gz
wget https://datarepo.eng.ucsd.edu/mcauley_group/data/amazon_v2/metaFiles2/meta_Books.json.gz

# Unzip
gunzip Books.json.gz
gunzip meta_Books.json.gz

conda create -n IGD python=3.10
conda activate IGD
pip install -r requirements.txt

# Preprocess and extract Item-frequency information
bash compute_item_freq.sh

bash ig_monitor.sh

• beta adjusts the weight of zero-IG tokens.
• To implement the baseline, set beta=1.0.
• For our method, beta=0.1 works well in general. You can grid search over: [0.08, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6]

bash evaluate.sh

• Adjust the alpha parameter in the evaluate.sh script:
  • alpha=0.0 is the baseline.
  • In the inference script, you can set: (0.0 0.1 0.2 0.3 0.4)
  • alpha=0.2 generally yields the good results.
• For D3 method, set length penalty to 0.0
• For BIGRec method, set the length penalty to 1.0 in the script.

• Uses cft_monitor.py. According to the original paper, search over:
  • beta = 0.09, 0.16, 0.29, 0.38, 0.5, 0.66, 0.9, 0.96
  • alpha = 0.01, 0.02, 0.025, 0.05, 0.1, 0.2, 0.3

• Part of the CFT method. Set alpha=0, and only tune beta.

In our experiments, we trained our methods on an H100 96G GPU and tested on an A5000 GPU. Different hardware configurations may cause minor differences in results.