Walk2Pers

Code and notebooks for the paper:

Beyond Markovian Drifts: Action-Biased Geometric Walks with Memory for Personalized Summarization

This repository contains a small, notebook-driven pipeline to (i) sanity-check / preprocess personalized summarization data, (ii) build text embeddings (e.g., T5 encodings), (iii) train Walk2Pers model variants, and (iv) run evaluation / testing.

Note: This repo is organized primarily around Jupyter notebooks (see Notebooks below). If you prefer scripts, you can convert notebooks to .py via jupyter nbconvert --to script.

---

What is Walk2Pers?

Walk2Pers operationalizes the Structured Walk Hypothesis (SWH) for personalized summarization: user preference evolves over time and should be modeled beyond short-memory / purely Markov drift assumptions. The method updates a user preference state using action-conditioned steps (direction + magnitude) and dual memory lanes (reinforce vs. suppress), designed specifically for summarization.

---

Repository Contents (high level)

• README.md — you’re reading it
• check_data.ipynb — dataset inspection / sanity checks
• test_data.ipynb — quick testing utilities on prepared data
• embedding_T5_encoding.ipynb — embedding generation (T5-based encoding)
• w2p_model_test.ipynb — model loading + evaluation / inference tests
• w2p_model_training_lmhead+nlayers.ipynb — training notebook (variant)
• w2p_model_training_lmhead_only (1).ipynb — training notebook (variant)

---

Setup

1) Environment

This codebase assumes a standard Python ML stack.

Recommended:

• Python 3.9+ (3.10/3.11 typically fine)
• PyTorch (CUDA optional)
• HuggingFace transformers, datasets
• numpy, pandas, scikit-learn, tqdm, matplotlib

If you already have your own environment management:

• Use pip / conda as usual and install missing packages based on notebook import errors.

Example (minimal, adjust as needed):

pip install torch transformers datasets numpy pandas scikit-learn tqdm matplotlib jupyter