Skip to content

Latest commit

 

History

History
53 lines (40 loc) · 2.12 KB

README.md

File metadata and controls

53 lines (40 loc) · 2.12 KB
Raw

Knowledge Plugins: Enhancing Large Language Models for Domain-Specific Recommendations

This repository contains the source code of the paper: Knowledge Plugins: Enhancing Large Language Models for Domain-Specific Recommendations.

Usage

1. Preprocess

1.1 Download and process raw data

Follow the documents and notebooks in preprocess/step1-Raw_Dataset_Parsing to download and process the Amazon/Beauty, ML1M, and Online Retail dataset.

This step will generate three files for each dataset: sequential_data.txt, negative_samples.txt, metadata.json.

1.2 Train base recommender model

Follow the instructions outlined in preprocess/step2-Base_models to train recommendation models and generate collaborative filtering (CF) information from two perspectives: MF_embedding for Item-2-Item and SASRec_embedding for User-2-Item, respectively.

This step will generate two files for each dataset: MF_embedding.pkl and SASRec_embedding.pkl.

2. Knowledge Extraction

We consider providing CF information from two main perspectives: Item-2-Item and User-2-Item.

For each type of CF information, we extract it from the embedding and organize it into structured data for future use.

For example:

cd Knowledge_Extraction
python extract_I2I.py \
    --dataset ml1m \
    --negative_type pop
python extract_U2I.py \
    --dataset ml1m \
    --negative_type pop

This step will generate a series of json files for each dataset, like global_CF.json, MF_CF_pop.json, etc.

3. DOKE(Prompt Construction and Response Generation via API)

In the DOKE/config directory, we have stored all the prompt templates used in our paper. You can also refer to them to compose your own prompt templates.

With the following command, you can evaluate the recommendation performance under different settings.

cd DOKE
python generate_prompt.py \
    --config config/ml1m/popneg_his_I2I.json \
    --dataset ml1m
python call_openai.py \
    --prompt out/prompts/ml1m/popneg_his_I2I.json \
    --model ChatGPT \
    --dataset ml1m
bash metric.bash out/result/ml1m/ChatGPT_popneg_his_I2I ml1m