We are happy to announce the release of DGL 2.1. In this release, we are making GNN data loading lightning fast. We introduce GPU acceleration for the whole GNN data loading pipeline in GraphBolt, including the graph sampling and feature fetching stages.

Learn more.

We’re thrilled to announce the release of DGL 2.0, a major milestone in our mission to empower developers with cutting-edge tools for Graph Neural Networks (GNNs). Traditionally, data loading has been a significant bottleneck in GNN training. Complex graph structures and the need for efficient sampling often lead to slow data loading times and resource constraints. This can drastically hinder the training speed and scalability of your GNN models. DGL 2.0 breaks free from these limitations with the introduction of dgl.graphbolt, a revolutionary data loading framework that supercharges your GNN training by streamlining the data pipeline.

Learn more.

We announces the launch of BSChecker, a new tool designed to identify subtle hallucinations in the outputs of large language models (LLMs). BSChecker provides

• a hallucination detection benchmark that covers three distinctive settings differentiated by the quantity and quality of context provided for LLM responses;
  
• a hallucination detection framework that checks hallucinations at finer granularity of knowledge triplets;
  
• leaderboards by both human annotations and auto-checkers to evaluate LLMs on hallucinations.
  

Learn More

Unsupervised Open-Vocabulary Object Localization in Videos: With the inductive bias of spatiotemporal consistency, we localize objects from videos in an unsupervised manner, further with a finetuned CLIP model we can also label semantics of each object.

Object-centric Multiple Object Tracking: OC-MOT significantly narrows the gap between the existing object-centric model and the fully supervised state-of-the-art MOT model by leveraging the versatile object-centric representation and the designed self-supervised memory module.

Coarse-to-Fine Amodal Segmentation with Shape Prior: we propose a novel approach, called Coarse-to-Fine Segmentation (C2F-Seg), to progressively model the amodal segmentation: initially reduces the learning space from the pixel-level image space to the VQ latent space, then refine with more details in pixel space.

Rethinking Amodal Video Segmentation from Learning Supervised Signals with Object-centric Representation: we propose a multi-view fusion layer based temporal module to fulfill sufficient object representation completion. Besides views from different frames, we design a translation module to project image features into the Bird’s-Eye View (BEV), which introduces 3D information to improve feature quality.

In the paper “An AMR-based Link Prediction Approach for Document-level Event Argument Extraction”, we proposed a method to build the Abstract Meaning Representation graph for a document whose nodes are normalized nouns and verbs. The edges are the pre-defined semantic roles, such as the subject and object of an action. We apply GNN+Transformer over an AMR graph to extract event arguments.

The idea in “Exploiting Abstract Meaning Representation for Open-Domain Question Answering” is similar to the above work but used in a different task with a different model architecture. Inspired by studies in Graph Learning area, we proposed a method to compress the information of edges as tokens and append these tokens to the text sequence. A transformer takes the concatenation of text tokens and graph tokens as inputs and generates the answer for the given question.

In the paper “Dual Cache for Long Document Neural Coreference Resolution”, we proposed a hybrid cache following LRU (Least recently used) and LFU (Least Frequently used) to maintain the candidate text span that is a reference of other entities when processing a document from left to right. Since the model only considers a fixed number of candidates in the cache, we save the memory and computation overhead without losing too much performance.

Since pre-defined and human-labeled discourse relations can be ambiguous when describing the semantics between sentences, we proposed to adopt a distributed representation to express the relationship between two sentences. We also propose an unsupervised learning method to obtain such distributed representations in the paper “Distributed Marker Representation for Ambiguous Discourse Markers and Entangled Relations”.

We are thrilled to announce the arrival of DGL 1.0, a cutting-edge machine learning framework for deep learning on graphs. A couple of highlights here:

• 100+ examples of state-of-the-art GNN models, 15+ top-ranked baselines on Open Graph Benchmark (OGB), available for learning and integration
• 150+ GNN utilities including GNN layers, datasets, graph data transform modules, graph samplers, etc. for building new model architectures or GNN-based solutions
• Flexible and efficient message passing and sparse matrix abstraction for developing new GNN building blocks
• Multi-GPU and distributed training capability to scale to graphs of billions of nodes and edges
  

We have three papers accepted at EMNLP 2022 (1 main conference and 2 findings).

In the paper “DORE: Document Ordered Relation Extraction based on Generative Framework”, we develop a seq2seq approach for document-level relation extraction. Its main idea is converting the free text generation into a symbolic sequence generation problem and introducing the negative sampling to balance the training objective. We propose an reinforcement learning based approach in “ReLET: A Reinforcement Learning Based Approach for Explainable QA with Entailment Trees” for QA reasoning, which models the reasoning steps in a multi-step reinforcement learning framework, existing works model reasoning steps separately. A new dialog benchmark has been introduced in the paper “Dilogue Meaning Representation for Task-Oriented Dialogue Systems”, which demonstrates the graph structure to describe dialog states, enhancing the capability of representing multiple intents and nested structure.

Learn More：

• DORE: Document Ordered Relation Extraction based on Generative Framework 
  
• ReLET: A Reinforcement Learning Based Approach for Explainable QA with Entailment Trees 
  
• Dialogue Meaning Representation for Task-Oriented Dialogue Systems
  

In the latest DGL v0.9.1, we released a new pipeline for preprocess, partition and
dispatch graph of billions of nodes or edges for distributed GNN training. At
its core is a new data format called Chunked Graph Data Format (CGDF) which
stores graph data by chunks. The new pipeline processes data chunks in parallel
which not only reduces the memory requirement of each machine but also
significantly accelerates the entire procedure. 

In the latest issue of Nature Machine Intelligence, Yousun Jung and his research team at Korea Institute of Science and Technology (KAIST) have proposed a novel generalized-template-based graph neural network called LocalTransform for accurate organic reactivity prediction . The generalized reaction templates demonstrate the built-in interpretability in addition to state-of-the-art prediction accuracy. LocalTransform has been implemented with DGL, DGL-LifeSci and PyTorch. 

We're excited to announce a new major release DGL v0.9. Here are the highlights:

• New functionality to bring highly efficient graph analytics into GNN via NVIDIA RAPIDS cuGraph.
  
• Built-in FP16 & mixed precision support.
• New command lines for model inference and graph prediction task.

The new 0.8.2 update brings new additions for the GNN community. Here are the highlights:

• Synthetic Datasets for Developing GNN Explainability Approaches
• SIGN Diffusion Transform
• Label Propagation
• Directional Graph Network Layer
• Graph Isomorphism Network Layer with Edge Features
• Feature Masking
• Row-Normalizer of Features
  

V0.8.1 is a minor release that includes module updates, optimizations, new features and bug fixes.

The highlights:

• Grouped Reversible Residual Connection for GNNs
  
• Laplacian Positional Encoding
  
• Random Walk Positional Encoding
  
• GraphSAINT Samplers
  
• E(n) Equivariant Convolutional Layer
  
• Principal Neighbourhood Aggregation Layer
  

We are excited to announce the release of DGL v0.8, which brings many new features as well as improvement on system performance.

The highlights:

• A major update of the mini-batch sampling pipeline, better customizability, more optimizations.
• Significant acceleration and code simplification of popular heterogeneous graph NN modules; 11 new off-the-shelf NN modules for building models for link prediction, heterogeneous graph learning and GNN explanation.
• GNNLens: a DGL empowered tool to visualize and understand graph data using GNN explanation models.
• New functions to create, transform and augment graph datasets, making it easier to conduct research on graph contrastive learning or repurposing a graph.
• DGL-Go: a new GNN model training command line tool that utilizes a simple interface so that users can quickly apply GNNs to their problems and orchestrate experiments with state-of-the-art GNN models.
  

There is an accelerated pace of industry adoption of GNNs. At the just-concluded GTC November 2021, CEO Jensen Huang highlighted the importance of GNN to the industry and announced the comprehensive partnership between NVIDIA and the DGL community. “Graphs — a key data structure in modern data science — can now be projected into deep-neural network frameworks with Deep Graph Library, or DGL, a new Python package.”, said Jensen.

DGL team is very excited to see the thriving growth of the DGL community. We and welcome all partnerships about this open-source project. We are HIRING both full-time employees as well as interns.

Today, the GAMMA Lab of Beijing University of Posts and Telecommunications officially released OpenHGNN-0.1.0, an Open Source Toolkit for Heterogeneous Graph Neural Network. OpenHGNN is based on DGL [Deep Graph Library] and PyTorch, and includes SOTA models of heterogeneous graph.

Key Features:

• Easy-to-Use: OpenHGNN provides easy-to-use interfaces to run experiments with models and datasets, using optuna as hyperparameter optimization framework.
• Extensibility: User can define customized task/model/dataset to apply new models to new scenarios.
• Efficiency: Efficient APIs using DGL as the backend.

In collaboration with Shanghai Jiaotong University faculty and students, our new paper "Bag of Tricks for Node Classification with Graph Neural Networks" introduces and motivates several effective tricks-of-the-trade for training graph neural networks (GNN) models on node classification.

These relate to label usage, loss function formulations, and model design tweaks that can significantly improve performance, often to an extent that outweighs the gains from more dramatic changes in the underlying GNN architecture. Among other things, many of the top-ranked models on the Open Graph Bench-mark (OGB) leaderboard benefit from these techniques. This paper has also recently been awarded by the Best Paper at the Deep Learning on Graphs Workshop at KDD 2021.

Our paper "Learning Hierarchical Graph Neural Networks for Image Clustering" has been accepted by the International Conference on Computer Vision (ICCV) 2021. In this paper, we propose a hierarchical graph neural network (GNN) model that learns how to cluster a set of images into an unknown number of identities using a training set of images annotated with labels belonging to a disjoint set of identities.
Our hierarchical GNN uses a novel approach to merge connected components predicted at each level of the hierarchy to form a new graph at the next level. And unlike fully unsupervised hierarchical clustering, the choice of grouping and complexity criteria stems naturally from the supervision in the training set. The resulting method, Hi-LANDER, achieves an average of 54% improvement in F-score and 8% increase in Normalized Mutual Information (NMI) relative to current GNN-based clustering algorithms while also achieving a three-fold decrease in computational cost.

Our paper "Graph Neural Networks Inspired by Classical Iterative Algorithms" has been accepted by International Conference on Machine Learning (ICML) 2021 as the long oral talks. In this paper, we propose a novel graph learning framework TWIRLS from the energy minimization perspective. This framework interprets the message passing process on the graph as the energy minimization process with gradient descent methods.

We also derived a novel attention mechanism on the graph through this perspective, by adding the robustness constraint to the energy on graph and optimizing it by iterative reweighted least squares method. Experiments show that TWIRLS can effectively accomplish long-range information propagation problem without oversmoothing, and also has good performance on heterophily graph data and graph adversarial attack.

The recent DGL 0.6 release is a major update on many aspects of the project including documentation, APIs, system speed, and scalability. This article highlights some of the new features and enhancements.

We participated in WebNLG Challenge 2020 and ranked #1 in automatic evaluation for both graph-to-text (35 submissions) and text-to-graph (4 submissions) tracks; and our unsupervised approach cycleGT ranked #10 and #3, respectively. Besides, our graph-to-text submission also ranked #1 in the human evaluation. Combined with our first place finish in the automatic evaluation, it indicates that we achieved the top ranking in every possible category.

We also had two submissions accepted to the WebNLG 2020 workshop: our unsupervised approach CycleGT and our systems' report which won the challenge mentioned above. Both unsupervised and supervised system are open-sourced, unsupervised, supervised .

Dec 8, 2020, re:invent20, Amazon announced Amazon Neptune ML, a new capability of Amazon Neptune that uses Graph Neural Networks (GNNs), a machine learning (ML) technique purpose-built for graphs, to make easy, fast, and accurate predictions using graph data. Using the Deep Graph Library (DGL), an open-source library to which Amazon contributes, Neptune ML allows you to improve the accuracy of most predictions for graphs by over 50% when compared to making predictions using non-graph methods.

This is a patch release mainly for supporting CUDA 11.0. Now DGL supports CUDA 11.0 and PyTorch 1.7 on Linux/Windows/Mac.

The recent DGL 0.5 release is a major update on many aspects of the project including documentation, APIs, system speed and scalability. This article highlights some of the new features and enhancements.

Use the new Drug Repurposing Knowledge Graph (DRKG) for repurposing drugs for fighting COVID-19. A step-by-step tutorial on how to use knowledge graph embeddings learned by DGL-KE to make prediction and rank drugs, and some news coverage. DRKG is in SAIL’2021 top30 list.