/human-like-explainations

Language models for recommendation

Primary LanguagePython

PERLM: Faithful Path-Based Explainable Recommendation via Language

This repository contains the source code of the submitted paper "PERLM: Faithful Path-Based Explainable Recommendation via Language Modeling over Knowledge Graphs".

Requirements

This repo requires the creation of a virtual environment. A python version below 3.12 is required.

This repo was adapted for Python 3.11 - on 19 March 2024.

Install the required packages: pip install -r requirements.txt

Some packages were adapted compared to original requirements.

Download the datasets and the embeddings(to run the plm-rec implementation) from the data.zip and embedding-weights.zip archive at the drive repository: https://drive.google.com/drive/folders/1e0uFWb6iJ6MXHtslZsqV8qRYC0Pl_AR7?usp=sharing Then extract both data.zip and embedding-weights.zip inside the top level of the repository (i.e. the level in which setup.py is located).

Usage

Design philosophy:

The experiments which are reported in the paper can be run with ease by means of the provided bash scripts. This holds both for dataset generation and model training. To access the lower level details, one can directly use the python scripts which are called by the same bash scripts.

Note: all experiments have been run with fixed seed in order to ease reproducibility of the results.

0. Install the repository

From the top-level (i.e. the folder which contains setup.py and the pathlm folder) Run:

pip install .

1. Path Dataset generation

To create the preprocessed/mapping folder needed by the random walk algorithm, run from the top level:

python pathlm/data_mappers/map_dataset.py --data <dataset_name> --model pearlm

To generate all datasets, run from the top level:

source build_datasets.sh

Each dataset is generated by the pipeline described in 'create_dataset.sh' which is in charge of:

  1. Generation of a dataset of at most X unique paths per user
  2. Concatenation of the results into a single .txt file
  3. (Optional) Pruning of the concatenated .txt file (This is only useful if the start entity is chosen instead of the standard 'USER')
  4. Move of the concatenated and pruned .txt file into the 'data' folder which is used to tokenize and train the models

2. Bulk Training

From the top-level (i.e. the folder which contains setup.py and the pathlm folder). Install the repository with pip install .

Then, proceed according to the chosen experiment to run as described below. Each bash script can be customised as desired in order to run alternative experiments

PERLM

To bulk train PERLM, run from the top level:

CUDA_DEVICE_NUM=0
source run_perlm_experiments.sh $CUDA_DEVICE_NUM
PLM-Rec

To train PLM-Rec, run from the top level:

CUDA_DEVICE_NUM=0
source run_plm-rec_experiments.sh $CUDA_DEVICE_NUM

3. Training

Before training a specific model, tokenize the dataset, running from the top level:

python pathlm/models/lm/tokenize_dataset.py --data <dataset_name> --sample_size <sample_size>
PERLM

To train a specific PEARLM, run from the top level:

python pathlm/models/lm/pearlm_main.py --data <dataset_name> --model <base-clm-model> --sample_size <sample_size>
PLM-Rec

To train a specific PLM, run from the top level:

python pathlm/models/lm/plm_main.py --data <dataset_name> --model <base-clm-model> --sample_size <sample_size>