TNTOOD & TextTopoOOD

This repository contains the implementation for the EMNLP2025 paper "Text Meets Topology: Rethinking Out-of-distribution Detection in Text-Rich Networks".

Intro

Out-of-distribution (OOD) detection remains challenging in text-rich networks, where textual features intertwine with topological structures. Existing methods primarily address label shifts or rudimentary domain-based splits, overlooking the intricate textual-structural diversity. For example, in social networks, where users represent nodes with textual features (name, bio) while edges indicate friendship status, OOD may stem from the distinct language patterns between bot and normal users. To address this gap, we introduce the TextTopoOOD framework for evaluating detection across diverse OOD scenarios: (1) attribute-level shifts via text augmentations and embedding perturbations; (2) structural shifts through edge rewiring and semantic connections; (3) thematically-guided label shifts; and (4) domain-based divisions. Furthermore, we propose TNT-OOD to model the complex interplay between Text aNd Topology using: 1) a novel cross-attention module to fuse local structure into node-level text representations, and 2) a HyperNetwork to generate node-specific transformation parameters. This aligns topological and semantic features of ID nodes, enhancing ID/OOD distinction across structural and textual shifts. Experiments on 11 datasets across four OOD scenarios demonstrate the nuanced challenge of TextTopoOOD for evaluating OOD detection in text-rich networks.

Dependencies

• Python 3.9.2
• Cuda 12.4
• Torch 2.5.1
• ogb 1.3.6
• torch_geometric 2.6.1
• torch_sparse 0.6.18
• nltk 3.9.1

Usage

1. Installation

Clone the repository and install dependencies:

git clone https://github.com/DannyW618/TNT.git
cd TNT
pip install -r requirements.txt

2. Download datasets

Download the datasets via:

pip install gdown
gdown 1cZDOK12AVhT2MnrE6ZeLWJZLlAp81bey -O datasets.zip
unzip datasets.zip -x "__MACOSX/*"

Optional (Recommended): Download the pre-processed OOD splits via:

# File content is relatively large ~52GB, ensure enough disk space and bandwidth
wget -c https://huggingface.co/datasets/DannyW618/TextTopoOOD/resolve/main/embeddings.tar.gz -O embeddings.tar.gz
tar -xvzf embeddings.tar.gz # Ensure this is placed at top-level directory according to the Repo structure given below

3. Running Experiments

Run an experiment with:

e.g., Cora [email protected]:

python main.py --datasets cora --shift_type feature --noise_level 0.9 --num_layers 2 --learning_rate 0.0001 --dropout 0.3 --use_tntood

Or multiple shifts and configs together:

python main.py --datasets cora --shift_type feature structure text text_swap semantic_connection --semantic_selection_mode threshold threshold threshold --target_density_factor 1 1 1 --threshold_percentile 0.75 0.85 0.95 --swap_scope both intra inter --swap_ratio 1 1 1 --text_augmentation_type synonym antonym --text_noise_level 1 1 --noise_level 0.5 0.7 0.9 --p_ii_factor 0.7 0.6 0.4 --p_ij_factor 0.5 0.3 0.7 --structure_noise_level 0.2 0.5 1.0 --num_layers 2 --learning_rate 0.0001 --dropout 0.3 --use_tntood

• More examples can be found in runtnt.sh.

Key arguments:

• --datasets : Choose dataset (e.g., cora, citeseer, pubmed, arxiv, etc.).
• --shift_type : Select OOD shift type (feature, structure, text, text_swap, semantic_connection).
• Default settings are defined in config.py and more arguments are in argparser.py.

We have provided the OOD split embeddings/data, you may choose to generate your own with different perturbation configs using ood_generation.py.

Supported OOD Shifts

• Feature Shift - Noisy or perturbed node features.
• Structure Shift - Graph topology altered using SBM-based rewiring.
• Text Shift - Corruption or perturbation of node text attributes.
• Text Swap - Random swapping of text features across nodes.
• Semantic Connection Shift - Alteration of graph connectivity based on semantic similarity metrics.
• Label Leave Out - Leave out a set of classes as OOD data.

Repository Structure

tntood/                  # Primary folder containing the code for TNTOOD
├── main.py              # Config loading, initialisation, training, dataset preparation, evaluation
├── runtnt.sh            # Example scripts to run experiments
├── ood_generation.py    # OOD shifts generation code
│
datasets/                # Contains the dataset used
│
embeddings/              # Cache folder for created OOD splits after running experiments

Acknowledgement

We thank the authors of TSGFM and GLBench for the datasets.

Citation

The bib ref for our paper is as follow:

@inproceedings{tntood,
  title = {Text Meets Topology: Rethinking Out-of-distribution Detection in Text-Rich Networks},
  author = {Danny Wang and Ruihong Qiu and Guangdong Bai and Zi Huang},
  booktitle = {Empirical Methods in Natural Language Processing (EMNLP)},
  year = {2025}
  }