embedding-disruptiveness

Measure how disruptive a paper or patent is — using graph embeddings on citation networks.

embedding-disruptiveness learns node2vec-style embeddings from citation graphs and computes an Embedding Disruptiveness Measure (EDM) that captures whether a work disrupts or consolidates its field. It also provides the classic Disruption Index (DI) as a built-in utility.

Paper | Blog Post | PyPI

---

Features

• End-to-end pipeline — load a citation network, train embeddings, compute disruptiveness in a few lines of code
• Node2Vec with directional skip-gram — biased random walks with (p, q) parameters for flexible neighborhood exploration
• Model parallelism — splits in-vectors and out-vectors across two GPUs for large-scale networks
• Multiple negative samplers — Configuration Model, Stochastic Block Model, Erdos-Renyi, and conditional context samplers
• Modularity-aware training — optional modularity regularization for community-sensitive embeddings
• Mixed-precision training — AMP support for faster training with lower memory usage
• Built-in disruption index — classic DI₁ / DI₅ / CDI computation alongside embedding-based measures

Installation

Using pip

pip install embedding-disruptiveness

Using uv

uv pip install embedding-disruptiveness

Install from source

git clone https://github.com/MunjungKim/embedding-disruptiveness.git
cd embedding-disruptiveness
pip install -e .

Or with uv:

git clone https://github.com/MunjungKim/embedding-disruptiveness.git
cd embedding-disruptiveness
uv pip install -e .

Requirements

• Python >= 3.8
• PyTorch (with CUDA for GPU training)
• NumPy, SciPy, scikit-learn, numba, tqdm

Quick Start

import embedding_disruptiveness as edm

# Train embeddings on a citation network (scipy sparse matrix in .npz format)
trainer = edm.EmbeddingTrainer(
    net_input="citation_network.npz",
    dim=128,
    window_size=5,
    device_in="0",       # GPU for in-vectors
    device_out="1",      # GPU for out-vectors
    q_value=1,
    epochs=5,
    batch_size=1024,
    save_dir="./output",
)

trainer.train()
# Embeddings and cosine distances are saved to save_dir

Usage Guide

Input Format

Your citation network should be a scipy sparse matrix saved as .npz. Rows/columns represent nodes (papers/patents), and non-zero entries represent citation edges.

import scipy.sparse as sp

# Example: build a sparse adjacency matrix and save it
net = sp.csr_matrix(adjacency_data)
sp.save_npz("citation_network.npz", net)

You can also convert an edge list (numpy array) to a sparse adjacency matrix using to_adjacency_matrix:

import numpy as np
from embedding_disruptiveness.utils import to_adjacency_matrix

# (N, 2) edge list: [src, dst]
edges = np.array([[0, 1], [1, 2], [2, 3]])
net = to_adjacency_matrix(edges, edgelist=True)

# (N, 3) weighted edge list: [src, dst, weight]
weighted_edges = np.array([[0, 1, 0.5], [1, 2, 1.0], [2, 3, 0.8]])
net = to_adjacency_matrix(weighted_edges, edgelist=True)

Training Embeddings

EmbeddingTrainer handles the full pipeline:

1. Loads the sparse network
2. Generates biased random walks (node2vec)
3. Trains a Word2Vec-style model with triplet loss
4. Saves in-vectors, out-vectors, and cosine distance matrices

trainer = edm.EmbeddingTrainer(
    net_input="network.npz",
    dim=128,               # Embedding dimension
    window_size=5,         # Context window for skip-gram
    device_in="0",         # CUDA device for in-vectors
    device_out="1",        # CUDA device for out-vectors
    q_value=1,             # Node2Vec return parameter (q)
    epochs=5,              # Training epochs
    batch_size=1024,       # Batch size
    save_dir="./results",  # Output directory
    num_walks=10,          # Random walks per node (default: 10)
    walk_length=80,        # Walk length (default: 80)
)

trainer.train()

Computing the Disruption Index

import embedding_disruptiveness as edm

# net: sparse adjacency matrix (citing → cited)
# 1-step disruption index
di = edm.calc_disruption_index(net)

# 2-step (multistep) disruption index
di_2step = edm.calc_multistep_disruption_index(net)

Two computation methods are available via the method parameter:

• "matrix" — sparse matrix multiplication. Fast for small networks (< 1M nodes), but uses O(N²) memory.
• "iterative" — Numba-JIT row-wise loop. Memory-efficient, scales to 100M+ nodes.
• "auto" (default) — automatically picks "matrix" if N < 1M, otherwise "iterative".

# Force iterative method for a large network
di = edm.calc_disruption_index(large_net, method="iterative")

# Force matrix method with batching for medium networks
di = edm.calc_disruption_index(net, method="matrix", batch_size=2**15)

Choosing a Negative Sampler

Different null models yield different notions of "expected" connections:

from embedding_disruptiveness.utils import (
    ConfigModelNodeSampler,
    SBMNodeSampler,
    ErdosRenyiNodeSampler,
)

# Configuration Model — preserves degree distribution
sampler = ConfigModelNodeSampler(adj_matrix)

# Stochastic Block Model — preserves community structure
sampler = SBMNodeSampler(adj_matrix, group_membership)

# Erdos-Renyi — uniform random baseline
sampler = ErdosRenyiNodeSampler(adj_matrix)

Custom Training Loop

For fine-grained control, use the components directly:

from embedding_disruptiveness.models import Word2Vec
from embedding_disruptiveness.loss import Node2VecTripletLoss
from embedding_disruptiveness.datasets import TripletDataset
from embedding_disruptiveness.torch import train

model = Word2Vec(vocab_size=num_nodes, embedding_size=128, padding_idx=num_nodes, device_in="cuda:0", device_out="cuda:1")
dataset = TripletDataset(center, context, negative_sampler, epochs=5)
loss_fn = Node2VecTripletLoss()

train(model=model, dataset=dataset, loss_func=loss_fn, batch_size=1024)

API Reference

Module	Key Exports	Description
embedding	EmbeddingTrainer	High-level training orchestrator
models	Word2Vec	Dual-device embedding model
loss	Node2VecTripletLoss, ModularityTripletLoss	Loss functions
datasets	TripletDataset, ModularityDataset	PyTorch datasets for triplet sampling
torch	train()	Training loop with AMP and SparseAdam
utils	RandomWalkSampler, *NodeSampler, calc_disruption_index, calc_multistep_disruption_index	Samplers and metrics

How It Works

1. Random Walks: Node2Vec-style biased walks explore the citation graph, capturing both local and global structure via (p, q) parameters.

2. Directional Skip-Gram: A Word2Vec model learns separate in-vectors (as a target) and out-vectors (as a context) for each node, preserving the directionality of citations.

3. Embedding Disruptiveness: Cosine distances between a focal paper's in-vector and its references'/citations' out-vectors quantify how much the paper departs from — or reinforces — existing knowledge.

Citation

If you use this package in your research, please cite:

@article{kim2024embedding,
  title={Embedding Disruptiveness Measure},
  author={Kim, Munjung and others},
  year={2024}
}

License

MIT License. See LICENSE for details.

Contributing

Contributions are welcome! Please open an issue or submit a pull request on GitHub.