GalaXC: Graph Neural Networks with Labelwise Attention for Extreme Classification

Related tags

Deep Learning python machine-learning deep-learning multi-label-classification extreme-classification graph-neural-networks galaxc

Overview

GalaXC

GalaXC: Graph Neural Networks with Labelwise Attention for Extreme Classification

@InProceedings{Saini21,
	author       = {Saini, D. and Jain, A.K. and Dave, K. and Jiao, J. and Singh, A. and Zhang, R. and Varma, M.},
	title        = {GalaXC: Graph Neural Networks with Labelwise Attention for Extreme Classification},
	booktitle    = {Proceedings of The Web Conference},
	month = "April",
	year = "2021",
	}

Setup GalaXC

git clone https://github.com/Extreme-classification/GalaXC.git
conda env create -f GalaXC/environment.yml
conda activate galaxc
pip install hnswlib
git clone https://github.com/kunaldahiya/pyxclib.git
cd pyxclib
python setup.py install
cd ../GalaXC

Dataset Structure

Your dataset should have the following structure:

DatasetName (e.g. LF-AmazonTitles-131K)
│   trn_X.txt   (text for trn documents, one text in each line)
|   tst_X.tst   (text for tst documents, one text in each line)
|   Y.txt       (text for labels, one text in each line)
│   trn_X_Y.txt (trn labels in spmat format)
|   tst_X_Y.txt (tst labels in spmat format)
|   filter_labels_test.txt (filter labels where label and test documents are same)
│
└───XXCondensedData (embeddings for tst, trn documents and labels, for benchmark datasets, XX=DX[Astec])
    │   trn_point_embs.npy (2D numpy matrix for trn document embeddings)
    │   tst_point_embs.npy (2D numpy matrix for tst document embeddings)
    |   label_embs.npy     (2D numpy matrix for label embeddings)

We have provided the DX(embeddings from Module 1 of Astec) embeddings for public benchmark datasets for ease of use. Got better(higher recall) embeddings from somewhere? Just plug the new ones and GalaXC will have better preformance, no need to make any code change! These files for LF-AmazonTitles-131K, LF-WikiSeeAlsoTitles-320K and LF-AmazonTitles-1.3M can be found here. Except the files in DXCondensedData, all other files are copy of the datasets from The Extreme Classification Repository.

Sample Runs

To reproduce the numbers on public benchmark datasets reported in the paper, the sample runs are

LF-AmazonTitles-131K

python -u -W ignore train_main.py --dataset /your/path/to/data/LF-AmazonTitles-131K --save-model 0  --devices cuda:0  --num-epochs 30  --num-HN-epochs 0  --batch-size 256  --lr 0.001  --attention-lr 0.001 --adjust-lr 5,10,15,20,25,28  --dlr-factor 0.5  --mpt 0  --restrict-edges-num -1  --restrict-edges-head-threshold 20  --num-random-samples 30000  --random-shuffle-nbrs 0  --fanouts 4,3,2  --num-HN-shortlist 500   --embedding-type DX  --run-type NR  --num-validation 25000  --validation-freq -1  --num-shortlist 500 --predict-ova 0  --A 0.6  --B 2.6

LF-WikiSeeAlsoTitles-320K

python -u -W ignore train_main.py --dataset /your/path/to/data/LF-WikiSeeAlsoTitles-320K --save-model 0  --devices cuda:0  --num-epochs 30  --num-HN-epochs 0  --batch-size 256  --lr 0.001  --attention-lr 0.05 --adjust-lr 5,10,15,20,25,28  --dlr-factor 0.5  --mpt 0  --restrict-edges-num -1  --restrict-edges-head-threshold 20  --num-random-samples 32000  --random-shuffle-nbrs 0  --fanouts 4,3,2  --num-HN-shortlist 500  --repo 1  --embedding-type DX --run-type NR  --num-validation 25000  --validation-freq -1  --num-shortlist 500  --predict-ova 0  --A 0.55  --B 1.5

LF-AmazonTitles-1.3M

python -u -W ignore train_main.py --dataset /your/path/to/data/LF-AmazonTitles-1.3M --save-model 0  --devices cuda:0  --num-epochs 24  --num-HN-epochs 15  --batch-size 512  --lr 0.001  --attention-lr 0.05 --adjust-lr 4,8,12,16,18,20,22  --dlr-factor 0.5  --mpt 0  --restrict-edges-num 5  --restrict-edges-head-threshold 20  --num-random-samples 100000  --random-shuffle-nbrs 1  --fanouts 3,3,3  --num-HN-shortlist 500   --embedding-type DX  --run-type NR  --num-validation 25000  --validation-freq -1  --num-shortlist 500 --predict-ova 0  --A 0.6  --B 2.6

SiameseXML results reproduction

Great work!

I want to reproduce the results of SiameseXML on the dataset Q2BP, how can I get this dataset?

And, it would be nice if the "Appendix" of the paper could be downloaded (only the main body of the paper, excluding the appendices in here.

Looking forward to the reply, ^^.

opened by nbcc 1

GalaXC: Graph Neural Networks with Labelwise Attention for Extreme Classification

Related tags

Overview

GalaXC

GalaXC: Graph Neural Networks with Labelwise Attention for Extreme Classification

Setup GalaXC

Dataset Structure

Sample Runs

YOU MAY ALSO LIKE

You might also like...

Some tentative models that incorporate label propagation to graph neural networks for graph representation learning in nodes, links or graphs.

A PyTorch implementation of "Graph Classification Using Structural Attention" (KDD 2018).

An implementation demo of the ICLR 2021 paper Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural Networks in PyTorch.

A PyTorch implementation of "Semi-Supervised Graph Classification: A Hierarchical Graph Perspective" (WWW 2019)

Official implementation of Self-supervised Graph Attention Networks (SuperGAT), ICLR 2021.

Code for the paper "How Attentive are Graph Attention Networks?"

Official PyTorch implementation of "AASIST: Audio Anti-Spoofing using Integrated Spectro-Temporal Graph Attention Networks"

Danfeng Hong, Lianru Gao, Jing Yao, Bing Zhang, Antonio Plaza, Jocelyn Chanussot. Graph Convolutional Networks for Hyperspectral Image Classification, IEEE TGRS, 2021.

Semi-Supervised Graph Prototypical Networks for Hyperspectral Image Classification, IGARSS, 2021.

Comments

SiameseXML results reproduction

Owner

Extreme Classification

HeatNet is a python package that provides tools to build, train and evaluate neural networks designed to predict extreme heat wave events globally on daily to subseasonal timescales.

DECAF: Deep Extreme Classification with Label Features

《LightXML: Transformer with dynamic negative sampling for High-Performance Extreme Multi-label Text Classiﬁcation》(AAAI 2021) GitHub:

Reimplementation of the paper `Human Attention Maps for Text Classification: Do Humans and Neural Networks Focus on the Same Words? (ACL2020)`

Implementation of Transformer in Transformer, pixel level attention paired with patch level attention for image classification, in Pytorch

Implementation of STAM (Space Time Attention Model), a pure and simple attention model that reaches SOTA for video classification

PocketNet: Extreme Lightweight Face Recognition Network using Neural Architecture Search and Multi-Step Knowledge Distillation

We have implemented shaDow-GNN as a general and powerful pipeline for graph representation learning. For more details, please find our paper titled Deep Graph Neural Networks with Shallow Subgraph Samplers, available on arXiv (https//arxiv.org/abs/2012.01380).

A static analysis library for computing graph representations of Python programs suitable for use with graph neural networks.

The source code of the paper "Understanding Graph Neural Networks from Graph Signal Denoising Perspectives"