PerAugy: Diversity Augmentation of Dynamic User Preference Data for Boosting Personalized Text Summarizers

This repository provides a step-by-step pipeline for preparing and processing the PENS dataset for sequential recommendation tasks using synthetic user interaction graphs (UIGs).

Dataset

• PENS Dataset: Includes train, validation, test, and news sets.

---

Processing Steps

1. Download the PENS Dataset

Start by downloading the PENS dataset, which contains the following files:

• train.json
• val.json
• test.json
• news.json

Ensure these files are placed in the appropriate directory for further processing.

---

2. Arrange Dataset into Seed UIG Format

Run the script:

python Scripts/PENS_augmentation.py

This script:

• Sorts user interactions by timestamp.
• Appends summary nodes from the test set.
• Generates a Seed UIG (user interaction graph) structure.

The following file will be created:

• synthetic-original.csv – the initial synthetic UIG dataset.

Additionally, a corresponding summary dataset will be generated with metadata about the summary nodes for each synthetic user.

---

3. Create Double Shuffled Trajectories

Open and run:

DS/DoubleShuffling.ipynb

This notebook generates Double Shuffled (DS) user trajectories. You can experiment with:

• offset
• gap
• segment length
• other relevant hyperparameters

---

4. Apply History-Influenced Perturbation

Run the following notebook to refine the DS dataset:

perturbation/perturbation_D2.ipynb

This step smooths the shuffled trajectories by applying history-aware perturbations to improve recommendation dynamics.

---

5. Convert to PENS Sequential Format

Run:

KG2PENS/KG2PENS_Trainer_New_Convertor.ipynb

This notebook:

• Replaces <d-s> pairs with corresponding s-nodes only.
• Converts the processed dataset into PENS-compatible knowledge graph format.

The final output format will be:

UserID, ClickedNewsID, PositiveNewsID, NegativeNewsID

---

Output

After completing all steps, you’ll obtain:

• A PENS-style dataset ready for sequential recommendation experiments.
• Summary statistics and user graph structures aligned with experimental design.

---

Notes

• Make sure all dependencies are installed before running the notebooks.
• Intermediate datasets are saved automatically after each step.
• You can tweak parameters in the Jupyter notebooks for different experimental settings.

Links

• Paper :- https://openreview.net/forum?id=JVx7Qi8tz3
• Datasets :- https://doi.org/10.6084/m9.figshare.30327451