Papers at international conferences related to AGI and knowledge information/graph data

Artificial General Intelligence

AGI, one of the main themes of this blog, stands for Artificial General Intelligence and refers to artificial intelligence with human-like general intelligence. While current artificial intelligence technology can exceed human performance in certain tasks, but requires the development of specialised models for each task, AGI is required to have a general intelligence that can respond to various tasks, like humans.

AGI (Artificial General Intelligence) can also be described as an artificial intelligence system that can comprehensively process various types of knowledge and information, and is closely related to knowledge information processing and graph data.

Knowledge information processing is one of the key technologies in artificial intelligence, which is used to process vast amounts of data and knowledge and perform tasks such as extraction, classification, inference and semantic interpretation, and AGI integrates these various types of knowledge information processing technologies to be able to perform a variety of human-like tasks It can be said that the aim is to be able to

In addition, graph data is a method for representing data and knowledge using nodes and edges, and is one of the most important data representation methods in artificial intelligence, as relationships and patterns can be easily grasped by using them In AGI, this graph data is effectively utilised to enable the acquisition of advanced knowledge from a vast amount of information, AGI can be said to make effective use of this graphical data and to enable the acquisition of advanced knowledge from vast amounts of information.

In this article, we will discuss a selection of recent (as of mid-2024) papers presented at prominent international conferences on topics centred on this knowledge information processing and machine learning of graph data.

ICML(International Conference on Machine Learning)

ICML will be recognised as one of the most prestigious conferences in the field of machine learning ICML covers a wide range of topics related to machine learning theory, algorithms and applications, including: deep learning, reinforcement learning, supervised and unsupervised learning, kernel methods, graphical models, and distributional estimation. Notable recent research results in ICML in this issue include.

• “COINs: Model-based Accelerated Inference for Knowledge Graphs“：To accelerate link prediction and knowledge graph query response models, we introduce COMmunity INformed Graph Embedding (COIN), which employs a community detection-based graph data augmentation procedure, followed by a two-stage prediction pipeline, namely community prediction for node localisation, and then localisation within the predicted community. We directly formulate the time complexity reduction and describe theoretically justified criteria to gauge the applicability of our approach in our setting. In addition, we provide numerical evidence of superior scalability in model evaluation costs
• “Generalizing Knowledge Graph Embedding with Universal Orthogonal Parameterization“：Recent advances in knowledge graph embedding (KGE) rely on Euclidean/hyperbolic relational orthogonalisation to model essential logical patterns and topological structures. However, existing approaches are limited to strict relational orthogonalisation with restricted dimensions and homogeneous geometry, and lack modelling capabilities. This work moves beyond these approaches in terms of both dimension and geometry by introducing a powerful framework named GoldE, which features a universal orthogonal parameterisation based on a generalised form of household head reflection. Such parameterisation naturally achieves dimensional extension and geometric unification with theoretical guarantees, allowing our framework to simultaneously capture important logical patterns and the inherent topological heterogeneity of knowledge graphs. Empirically, GoldE achieves state-of-the-art performance on three standard benchmarks.
• “Community-Invariant Graph Contrastive Learning“：Graph extensions have attracted a great deal of attention in recent years, with Graph Contrastive Learning (GCL) for learning well-generalised node/graph representations. However, mainstream GCL methods often prefer to randomly disrupt the graph for extensions. This exhibits limited generalisation and inevitably leads to corruption of high-level graph information, i.e. the graph community. Furthermore, current knowledge-based graph augmentation methods can only focus on topology or node functions, so the models lack robustness to different types of noise. To address these limitations, this study investigates the role of graph communities in graph augmentation and elucidates key advantages for learnable graph augmentation. Based on our observations, we propose a community-invariant GCL framework to preserve graph community structure during learnable graph augmentation. By maximising spectral variation, this framework unifies both topological and functional augmentation constraints and increases the robustness of the model; empirical evidence on 21 benchmark datasets demonstrates the exclusive benefits of our framework.

NeurIPS (Conference on Neural Information Processing Systems)

NeurIPS will be one of the world’s largest scientific conferences on neural information processing systems and recognised as one of the most important conferences in computer science NeurIPS will feature machine learning, deep learning, neural networks, reinforcement learning and natural language processing. Presentations have been made in a wide range of fields. In this issue, we have included reports on graph neural networks, natural language processing and molecular generation as papers that have attracted attention at recent NeurIPS.

• “Learning from Both Structural and Textual Knowledge for Inductive Knowledge Graph Completion“：Learning in rule-based systems plays a crucial role in Knowledge Graph Completion (KGC). Existing rule-based systems restrict the system’s input to structural knowledge only, potentially omitting knowledge useful for reasoning, such as textual knowledge. In this paper, we propose a two-stage framework for learning rule-based systems that imposes both structural and textual knowledge. In the first stage, a set of triples with a confidence score (called emph{soft triples}) is computed from a remotely supervised text corpus, and a textual inclusion model with multi-instance learning is used to estimate whether a given triple is accompanied by a sequence of sentences . In a second step, these soft triples are used to learn a rule-based model of KGC. To reduce the negative effects of noise caused by soft triples, we propose a new formalism of rules to be learnt, named emph{text enhanced rules} or emph{TE-rules} for short To learn TE-rules effectively, we propose a We propose a neural model that simulates Theoretically, we show that any set of TE-rules can always be interpreted by a particular parameter assignment of the neural model. Three new datasets are presented to assess the effectiveness of our method. Experimental results show that the introduction of soft triples and TE rules leads to significant performance improvements in guided link prediction.
• “UUKG: Unified Urban Knowledge Graph Dataset for Urban Spatiotemporal Prediction“：Accurate urban space-time prediction (USTP) is crucial for the development and operation of smart cities. As an emerging building block, multi-source urban data is usually integrated as an Urban Knowledge Graph (UrbanKG), which provides critical knowledge for urban spatio-temporal forecasting models. However, existing UrbanKGs are often tailored to specific downstream forecasting tasks and are not publicly available, limiting potential advances. In this paper, we present UUKG, a unified urban knowledge graph dataset for knowledge-enhanced urban spatio-temporal forecasting. Specifically, we first build an UrbanKG consisting of millions of triplets for two large cities by connecting heterogeneous urban entities such as boroughs, POIs and road segments. In addition, qualitative and quantitative analysis of the constructed UrbanKGs will be conducted to identify various higher-order structural patterns, such as hierarchies and cycles, that can be exploited to benefit downstream USTP tasks.To validate and facilitate the use of UrbanKGs, 15 KG embedding methods in the KG completion task are implemented and evaluated, and the learned KG embeddings are integrated into nine spatiotemporal models for five different USTP tasks. The extensive experimental results not only provide a benchmark for knowledge-enhanced USTP models under different task settings, but also highlight the potential of state-of-the-art higher-order structure-aware UrbanKG embedding methods. We hope that the proposed UUKG will facilitate research on urban knowledge graphs and a wide range of smart city applications.
• “GraphAdapter: Tuning Vision-Language Models with Dual Knowledge Graph“：Adapter-style Efficient Transfer Learning (ETL) shows excellent performance in tuning Vision Language Models (VLMs) under low data regimes, with only a few additional parameters introduced to unearth task-specific knowledge based on a general and powerful representation of the VLM The VLM is not a very good model. However, most adaptor-style works face two limitations. (i) modelling task-specific knowledge with only a single modality; and (ii) overlooking the utilisation of inter-class relationships in downstream tasks, thereby leading to sub-optimal solutions. To mitigate this, we propose an effective adapter-style tuning strategy called GraphAdapter. It implements text adapters by explicitly modelling dual-modality structural knowledge (i.e. the correlation of different semantics/classes of textual and visual modalities) in a dual-knowledge graph. In particular, the dual knowledge graph is established with two sub-graphs – a textual knowledge sub-graph and a visual knowledge sub-graph – where nodes and edges represent semantics/classes and their correlations in the two modalities respectively. This allows each prompt text feature to utilise task-specific structural knowledge from both the textual and visual modalities, resulting in a more effective classifier for downstream tasks.Extensive experimental results on 11 benchmark datasets show that GraphAdapter significantly outperforms previous adaptor-based methods.
• “Fair Graph Distillation“: Graph neural networks (GNNs) struggle with large graphs due to high computational demands, so data distillation of graph data alleviates this problem by distilling large real graphs into smaller distilled graphs, while maintaining comparable predictive performance of GNNs trained on both graphs promises to do so. However, we observe that GNNs trained on distillation graphs may exhibit more severe group equity problems than GNNs trained on real graphs. Motivated by this observation, we propose textit{fair graph distillation} (Algnameabbr), an approach for generating small distilled textit{fair and beneficial} graphs based on graph distillation methods. The challenge lies in the lack of sensitive attributes of the nodes of the distillation graph, which makes most debiasing methods (such as normalisation and adversarial debiasing) intractable for the distillation graph. A simple and effective bias metric called Coherence for distillation graphs is developed. Based on the proposed coherence metric, we introduce a framework for fair graph distillation using a bilevel optimisation algorithm. Extensive experiments show that the proposed algorithm can achieve a better prediction performance fairness trade-off across different datasets and GNN architectures.

CVPR (IEEE Conference on Computer Vision and Pattern Recognition)

CVPR is an international conference on computer vision and pattern recognition organised by the Institute of Electrical and Electronics Engineers (IEEE) and is regarded as one of the most important scientific conferences in the field of computer vision At CVPR a wide range of presentations in the fields of computer vision and pattern recognition, etc., including image recognition, image processing, machine learning, deep learning, 3D vision, robot vision, etc. In this issue, the papers that have attracted attention at recent CVPRs include the following.

• “EGTR: Extracting Graph from Transformer for Scene Graph Generation“：Scene graph generation (SGG) is a challenging task of detecting objects and predicting relationships between objects; since the development of DETR, single-stage SGG models based on single-stage object detectors have been actively studied. However, complex modelling is used to predict relations between objects, ignoring the inherent relations between object queries learnt in the multi-head self-attention layer of the object detector Extracting relation graphs from the various relations learnt in the multi-head self-attention layer of the DETR decoder We propose a lightweight one-stage SGG model that By fully utilising the self-attention by-product, the relation graph can be effectively extracted with a shallow relation extraction head. Considering the dependence of the relation extraction task on the object detection task, we propose a novel relation smoothing technique that adaptively adjusts the relation labels according to the quality of the detected objects. With relation smoothing, the model is trained according to a continuous curriculum that focuses on the object detection task at the start of training and performs multi-task learning as object detection performance gradually improves. In addition, a connection prediction task is proposed as an auxiliary task for relation extraction, which predicts whether a relation exists between object pairs. The effectiveness and efficiency of the method on the Visual Genome and Open Image V6 datasets will be demonstrated.
• “Dynamic Graph Representation with Knowledge-aware Attention for Histopathology Whole Slide Image Analysis“：Classification of whole histopathological slide images (WSI) has become a fundamental task in medical microscopy image processing. Common approaches learn WSIs as instance bag representations and highlight important instances, but struggle to capture the interactions between instances. Furthermore, traditional graph representation methods utilise explicit spatial locations to build topological structures, but limit flexible interaction capabilities between instances at arbitrary locations, especially when spatially distant. In response, we propose a new dynamic graph representation algorithm that conceptualises WSI as a form of knowledge graph structure. Specifically, it dynamically constructs neighbour and directed edge embeddings based on head and tail relationships between instances. We then devise an attention mechanism for knowledge recognition that can update the functionality of head nodes by learning the joint attention scores of each neighbour and edge. Finally, we obtain graph-level embeddings through a global pooling process of updated heads, which act as an implicit representation of the WSI classification. Our end-to-end graph representation learning approach outperforms state-of-the-art WSI analysis methods on three TCGA benchmark datasets and an in-house test set.
• “De-confounded Data-free Knowledge Distillation for Handling Distribution Shifts“：Data-free knowledge distillation (DFKD) is a promising task to train high-performance small models to enhance real deployments without relying on original training data. Existing methods typically avoid relying on individual data by utilising synthetic or sampled data. However, a long-overlooked problem is the severe distributional shift between those replacements and the original data, which manifests itself as large differences in image quality and class proportions. Detrimental shifts are essentially confounding factors that cause significant performance bottlenecks. To tackle this problem, this paper proposes a novel perspective involving causal reasoning to untangle student models from the effects of such shifts. By designing a customised causal graph, the causal relationships between variables in the DFKD task are first revealed. It then proposes a Knowledge Distilled Causal Intervention (KDCI) framework based on backdoor adjustment to disrupt confounding; KDCI can be flexibly combined with most existing state-of-the-art baselines; experiments in combination with six representative DFKD methods have demonstrated. It clearly helps existing methods in almost all settings textit{e.g.} and improves the baseline of the CIFAR-100 dataset by up to 15.54% accuracy.

ACM SIGKDD (Knowledge Discovery and Data Mining)

It will be a special interest group of the Association for Computing Machinery (ACM), established to promote research in the field of ACM SIGKDD (Knowledge Discovery and Data Mining). ACM SIGKDD is an important venue for researchers and engineers from all over the world to present their research results in the fields of machine learning, statistics, data mining, data analytics, big data and artificial intelligence, and to share information on the latest technologies and trends in data science. The ACM SIGKDD has become an important venue for researchers and engineers from all over the world. In this issue, the following papers were highlighted at the recent ACM SIGKDD: the importance of data augmentation, faster graph generation techniques and improved interactive dismantling tasks.

• “Exploiting Relation-aware Attribute Representation Learning in Knowledge Graph Embedding for Numerical Reasoning“：Numerical inference is an essential task to support machine learning applications such as recommendation and information retrieval. Numerical inference tasks utilise existing relational information and numerical attributes (e.g. height of an entity) in the knowledge graph to compare two items and infer new facts (e.g. higher). However, most existing methods rely on utilising attribute encoders or additional loss functions to predict numerical relationships. Hence, when attributes are sparsely observed, prediction performance is often not robust. In this paper, we propose a Relation-Aware Attribute Representation Learning-based Knowledge Graph Embedding (RAKGE) method for numerical reasoning tasks called RAKGE, which incorporates a newly proposed attribute representation learning mechanism that can exploit the association between relations and their corresponding numerical attributes RAKGE incorporates a newly proposed attribute representation learning mechanism that can exploit the association between relations and corresponding numerical attributes. In addition, a robust self-monitoring learning method is introduced to generate invisible positive and negative examples, making the approach more reliable when numerical attributes are sparsely available.In an evaluation of three real-world datasets, the proposed model outperformed state-of-the-art methods and compared to best competitors Hits@1で最大65.1% and achieved up to 52.6% improvement in MRR.
• “Knowledge Graph Self-Supervised Rationalization for Recommendation“：This paper presents a novel self-monitoring rationalisation method called KGRec for knowledge-aware recommender systems. An attentive knowledge rationalisation mechanism is proposed that generates rational scores of knowledge triplets in order to effectively identify useful knowledge connections. With these scores, KGRec integrates generative and contrastive self-supervision tasks for recommendations through rational masking. To emphasise the rationale of the knowledge graph, a new generative task is designed in the form of a masking reconstruction. By masking important knowledge with high rational scores, KGRec is trained to reconstruct and highlight useful knowledge connections that serve as a theoretical basis. To further rationalise the impact of collaborative interaction on knowledge graph learning, we introduce a contrastive learning task that adjusts the signal from the interaction view of knowledge and user items. To ensure noise-tolerant contrast, potentially noisy edges of both graphs, as judged by reasonable scores, are masked; extensive experiments on three real-world datasets show that KGRec outperforms state-of-the-art methods.
• “Exploring the Potential of Large Language Models (LLMs) in Learning on Graphs“：Learning on graphs has attracted significant attention due to its wide range of real-world applications. The most common pipelines for learning on graphs with text node attributes mainly rely on graph neural networks (GNNs), which utilise shallow text embeddings as initial node representations and are limited in their general knowledge and deep semantic understanding. Recently, Large Language Models (LLMs) have been proven to possess extensive common knowledge and strong semantic understanding capabilities, revolutionising existing workflows for processing textual data. This paper aims to explore the potential of LLMs in graph machine learning, particularly in node classification tasks, and investigate two possible pipelines, LLMs-as-Enhancers and LLMs-as-Predictors. The former utilises LLMs to enhance textual attributes of nodes with vast amounts of knowledge and generate predictions through GNNs. The latter seeks to employ LLMs directly as stand-alone predictors. We have conducted comprehensive and systematic research on these two pipelines in a variety of settings. From the comprehensive empirical results, we make original observations and find new insights that open up new possibilities and suggest promising directions for utilising LLMs for learning with graphs

ICLR (International Conference on Learning Representations)

ICLR is the world’s leading conference on representation learning in machine learning. representation learning is a branch of machine learning that refers to methods for extracting useful features from data. at ICLR, a variety of The ICLR has published papers in various fields of representation learning, such as deep learning, recurrent neural networks and convolutional neural networks. In this issue, the following papers are listed as recent ICLR highlights: performance improvement of large-scale language models, improvement of multilingual processing techniques, and innovation in image editing techniques.

• “GraphLLM: Boosting Graph Reasoning Ability of Large Language Models“：Advances in large-scale language modelling (LLM) have significantly pushed the boundaries for artificial general intelligence (AGI) due to its superior ability to understand diverse types of information, including but not limited to images and audio. Despite this progress, significant gaps remain in enabling LLMs to skilfully understand and reason about graph data. Recent research highlights the overwhelming performance of LLMs on basic graph inference tasks. This paper seeks to unearth the obstacles that hinder LLMs in graph reasoning and identifies the common practice of transforming graphs into natural language descriptions (Graph2Text) as a fundamental bottleneck. To overcome this obstacle, we introduce GraphLLM, a pioneering end-to-end approach that synergistically integrates graph learning models and LLM. This synergy provides LLMs with the ability to skilfully interpret and reason about graph data, utilising the superior expressive power of graph learning models. empirical evaluation across four basic graph inference tasks validates the effectiveness of GraphLLM. Results show a significant average accuracy improvement of 54.44%, along with a notable context reduction of 96.45% across the various graph inference tasks.
• “Think-on-Graph: Deep and Responsible Reasoning of Large Language Model on Knowledge Graph“：Large-scale language models (LLMs) have been used with great success in a variety of tasks, but often struggle with problems of illusions, especially in scenarios requiring deep and responsible reasoning. These problems can be partially solved by introducing external knowledge graphs (KGs) into LLM reasoning. In this paper, we propose a novel LLM-KG integration paradigm, where the LLM⊗KGLLM is treated as an agent, which interactively explores relevant entities and relations in the KG and performs inference based on the acquired knowledge; the LLM agent iteratively performs beam search in the KG to find the most We further implement this paradigm by introducing a new approach called Think-on-Graph (ToG), which discovers promising inference paths and returns the most likely inference results. We use a number of well-designed experiments to examine and illustrate the following advantages of ToG: 1) Compared to LLM, ToG has deeper reasoning power; 2) ToG has the capacity for knowledge traceability and knowledge accuracy by utilising LLM reasoning and expert feedback have; 3) ToG provides a flexible plug-and-play framework for different LLM, KG and prompting strategies without additional training costs; 4) ToG’s performance with small LLM models can exceed that of large, such as GPT-4 in certain scenarios; 5) ToG’s performance with small LLM models can potentially exceeds that of large LLMs, thereby reducing the cost of LLM deployment and application. As a training-free method with lower computational costs and better generality, ToG achieves overall SOTA on six of the nine datasets where most previous SOTAs relied on additional training.
• “Towards Foundation Models for Knowledge Graph Reasoning“：The foundational model of language and vision has the ability to perform inferences on textual and visual input, thanks to transferable representations such as a vocabulary of language tokens. Knowledge graphs (KGs) have different entity and relation vocabularies that do not generally overlap, and a key challenge in designing foundational models of KGs is to learn such transferable representations that allow inference in any graph with any entity and relation vocabulary. In this work, we take a step towards such a foundational model and present ULTRA, an approach for learning universal and transferable graph representations. ulTRA constructs relational representations as functions that are conditioned on interactions. Such a conditioning strategy allows pre-trained ULTRA models to be inductively generalised to invisible KGs with arbitrary relational vocabularies and fine-tuned on arbitrary graphs: link prediction experiments on 57 different KGs show that a single on an invisible graph of various sizes The zero-shot inductive inference performance of the pre-trained ULTRA model was found to be equal to or better than a strong baseline trained on a given graph in many cases. Fine-tuning can further improve performance.

TNNLS（IEEE Transactions on Neural Networks and Learning Systems）

TNNLS is a journal that publishes papers on neural networks and machine learning, and will include papers on a wide range of machine learning topics, including neural networks, deep learning, reinforcement learning, kernel methods, statistical learning theory, optimisation theory and evolutionary computation. In this issue, recent TNNLS papers of interest include new methods combining evolutionary algorithms described in “Overview of evolutionary algorithms and examples of algorithms and implementations” and multi-objective optimisation methods with deep learning, and applications of deep learning to graph data, as listed below.

• “RARE: Robust Masked Graph Autoencoder“：Mask Graph Autoencoder (MGAE) has emerged as a promising self-monitoring graph pre-training (SGP) paradigm due to its simplicity and effectiveness. However, existing efforts perform masking and reconstruction operations in raw data space, as is done in the computer vision (CV) and natural language processing (NLP) domains, while ignoring important non-Euclidean properties of graph data. The resulting highly unstable local connection structures significantly increase the uncertainty of guessing the masked data, reduce the reliability of exploited self-monitoring signals and lead to inferior representations for downstream evaluation. To address this problem, a robust mAsked gRaph autoEncoder (RARE), which further masks and reconstructs node samples in higher-order latent function space to improve the certainty of guessing masked data and the reliability of the self-monitoring mechanism, is A new SGP method is proposed. Through both theoretical and empirical analysis, it was found that implementing a joint mask reconstruction strategy on both the latent function and the raw data space could improve stability and performance. To this end, we elaborate a masked latent function completion scheme that predicts the latent function of masked nodes under the guidance of higher-order sample correlations, which are difficult to observe in terms of raw data. Specifically, we first employ a latent function predictor to predict the masked latent function from what is visible. Next, the raw data of the masked samples are encoded with a momentum graph encoder, and the resulting representation is then used to improve the prediction results through latent feature matching Extensive experiments on 17 datasets have shown that the state-of-the-art (SOTA) across three downstream tasks The effectiveness and robustness of RARE against competitors has been demonstrated.
• “Correlation-aware Spatial-Temporal Graph Learning for Multivariate Time-series Anomaly Detection“：Multivariate time series anomaly detection is critical in many applications, such as retail, transport, power grids and water treatment plants. Existing approaches to this problem mainly employ either statistical models that do not capture non-linear relationships well, or traditional deep learning models (such as CNNs and LSTMs) that do not explicitly learn pairwise correlations between variables. To overcome these limitations, a new method for time-series anomaly detection, called Correlation-Sensitive Spatial-Temporal Graph Learning (CST-GL), is proposed: via a multivariate time-series correlation learning module that allows spatial-temporal graph neural networks (STGNNs) to develop pairwise correlations are explicitly captured. By employing graph convolutional networks that exploit one-hop and multi-hop neighbour information, STGNN components can then encode rich spatial information from complex pairwise dependencies between variables. Using a time module consisting of extended convolutional functions, STGNN can further capture long-range dependencies over time. A new anomaly scoring component is further integrated into CST-GL to estimate the extent of anomalies in a purely unsupervised manner. Experimental results show that CST-GL effectively detects anomalies in general settings and enables early detection with different time delays.
• “Knowledge Graph Contrastive Learning based on Relation-Symmetrical Structure“：Knowledge Graph Embedding (KGE) aims at learning powerful representations that benefit different artificial intelligence applications. Contrastive learning, on the other hand, is widely used in graph learning as an effective mechanism to increase the discriminative power of learnt representations. However, the complex structure of KGs makes it difficult to construct suitable contrastive pairs. Only a few attempts have been made to integrate contrastive learning strategies with KGEs. However, most of them rely on language models (e.g. Bart) for contrastive pair construction rather than fully mining the information underlying the graph structure, hindering expressive capabilities. Surprisingly, entities within relational symmetric structures are usually found to be similar and correlated. To this end, we propose a knowledge graph contrast learning framework based on the relational symmetric structure KGE-SymCL, which mines KG symmetric structure information to enhance the discriminative power of KGE models. Specifically, a plug-and-play approach is proposed by taking entities in relational symmetric positions as positive pairs. Furthermore, self-monitoring alignment loss is designed to bring together positive pairs. Experimental results on link prediction and entity classification datasets show that KGE-SymCL can be easily employed in various KGE models for performance improvement. Furthermore, extensive experiments show that our model has the potential to outperform other state-of-the-art baselines.

SIGGRAPH (Special Interest Group on Computer GRAPHics)

The Special Interest Group on Computer GRAPHics, SIGGRAPH for short, is a professional organisation on computer graphics and interactive technologies and is part of the Association for Computing Machinery (ACM) SIGGRAPH is dedicated to advances in the fields of computer graphics, visualisation, animation, visual effects, virtual reality (VR), augmented reality (AR) and human-computer interaction.

• Zero Grads: Learning Local Surrogate Losses for Non-Differentiable Graphics

Gradient-based optimisation is now ubiquitous in the graphics world, but unfortunately it cannot be applied to problems with undefined or zero gradients. To avoid this problem, the loss function can be manually replaced by a similarly miniscule but differentiable “substitute”. Our proposed framework ZeroGrads automates this process by learning a neural approximation of the objective function. The surrogate learns an actively smoothed version of the objective function, encouraging locality and focusing the surrogate’s capabilities on what is important in the current learning episode. Fitting is performed online in parallel with parameter optimisation and is performed in a self-supervised manner without pre-computed data or pre-trained models. As objective sampling is expensive (requiring full rendering or simulator execution), we devise an efficient sampling scheme that allows tractable runtimes and competitive performance with little overhead. We demonstrate optimisation of a variety of non-convex, non-differentiable black-box problems in graphics, including visibility in rendering, discrete parameter space in procedural modelling and optimal control in physics-driven animation. In contrast to other non-differentiable algorithms, our approach is scalable to higher dimensions and is demonstrated on problems with up to 35k linked variables.

ISWC（International Semantic Web Conference）

The ISWC is an international conference for researchers on the Semantic Web, where the Semantic Web is a technology that aims to improve the utilisation of information on the Web by representing it in a format that machines can process. The papers that attracted attention at the recent ISWC are listed below, including methods that utilise BERT in knowledge graph completion tasks, methods that emphasise robustness and explanatory properties, and methods that improve accuracy by considering the structure of the knowledge graph.

• “PSYCHIC: A Neuro-Symbolic Framework for Knowledge Graph Question-Answering Grounding“：Answering Academic Questions in the Linked Data (Scholarly QALD) Challenge of the International Semantic Web Conference (ISWC) 2023 Challenge presents two subtasks to tackle Question Answering (QA) on Knowledge Graph (KG) questions: queries related to KG questions and Answering KGQA over DBLP (DBLP-QUAD) task by proposing a neuro-symbolic (NS) framework based on PSYCHIC, an extractive QA model that allows entities to be identified. Our system achieved an F1 score of 00.18% in answering the questions and ranked third in Entity Linking (EL) with a score of 71.00%.
• “Semantic Web Challenge on Tabular Data to Knowledge Graph Matching (SemTab 2023)“：The SemTab Challenge focused on benchmarking systems that interpret and annotate tabular data using semantic information from knowledge graphs. Various methods were evaluated, highlighting advances in semantic table interpretation using external knowledge bases such as wikidata.
• “TorchicTab: Semantic Table Annotation with Wikidata and Language Models“：A wealth of tabular data exists and is used in a wide range of applications. However, the majority of these data lack the semantic information needed for users and machines to understand them properly. This lack of semantic understanding of tables hinders their use in the data analysis pipeline. Solutions exist for semantically interpreting tables, but they focus on specific annotation tasks and table types and rely on large knowledge bases, making them difficult to re-use in real-world settings. Therefore, a more robust system is needed that generates more accurate annotations and adapts to different table types. The Semantic Web Challenge on Tabular Data and Knowledge Graph Matching (SemTab) was introduced to benchmark semantic table interpretation systems by evaluating them on diverse data sets and tasks. This paper introduces TorchicTab, a versatile semantic table interpretation system that can annotate tables with different structures using external knowledge graphs such as Wikidata or annotated tables containing pre-defined terms for training We evaluate the proposed system according to the different annotation tasks of the SemTab challenge. Results show that our system is able to generate accurate annotations for different tasks across different datasets.

AAAI（Association for the Advancement of Artificial Intelligence）

The AAAI is an important international conference for many researchers in artificial intelligence, and will be an organisation that publishes papers and reports on basic research, applied research, education and public policy in various artificial intelligence fields. In this issue, we list some of the recent AAAI papers that have attracted attention for their progress in self-supervised learning on graph data, improvement of the accuracy of few-shot learning by using graph structures and meta-learning, and development of fast graph classification methods that combine deep learning and Gaussian processes.

• “Lifelong Embedding Learning and Transfer for Growing Knowledge Graphs“：Existing knowledge graph (KG) embedding models focus primarily on static KG. However, real-world KGs do not remain static, but evolve and grow in parallel with the development of KG applications. As a result, new facts and previously unseen entities and relationships continually emerge, requiring an embedded model that can quickly learn and transfer new knowledge through growth. Motivated by this, we delve into an expanding field of KG embedding in this paper: lifetime KG embedding. We consider knowledge transfer and retention of learning about KG growth snapshots without having to learn embedding from scratch. The proposed model includes a masked KG autoencoder for embedding learning and updating, an embedding transfer strategy for injecting learned knowledge into new entity and relationship embeddings, and an embedding normalisation method to avoid catastrophic forgetting. investigating the impact of different aspects of KG growth To do so, four datasets are constructed to assess the performance of KG embedding across the lifespan. Experimental results show that the proposed model outperforms state-of-the-art induced and lifetime embedding baselines.
• “Few-Shot Knowledge Graph Completion“：Knowledge graphs (KGs) serve as a useful resource for a variety of natural language processing applications. Previous KG completion approaches require a large number of training instances (i.e. head-tail entity pairs) for every relation. The practical case is that for most relations there are few entity pairs available. Existing work on one-shot learning limits the generalisability of the method for several-shot scenarios and does not fully use supervisory information. However, the completion of a few-shot KG has not yet been fully researched. In this work, we propose a new one-shot relational learning model (FSRL) that aims to discover new relational facts in several-shot references.FSRL effectively captures knowledge from heterogeneous graph structures, performs an aggregate representation of several-shot references, and for all relations, it provides a reference set of It can match similar entity pairs, and extensive experiments on two public datasets show that FSRL outperforms the state-of-the-art.
• “Exploring Relational Semantics for Inductive Knowledge Graph Completion“：Knowledge Graph Completion (KGC) aims to infer missing information in an incomplete Knowledge Graph (KG). Most previous works only consider transductive scenarios where entities are present in the KG and do not work effectively in inductive scenarios involving emerging entities. Recently, several graph neural network-based methods have been proposed for inductive KGC by aggregating neighbourhood information to capture some uncertainty semantics from neighbouring auxiliary triples. However, these methods ignore the more general relational semantics underlying all known triples, which can provide richer information to represent emerging entities to satisfy inductive scenarios. In this paper, we propose a new model for induced KGC, called CFAG, which exploits two granularity levels of relational semantics in a coarse-grained aggregator (CG-AGG) and a fine-grained generating adversary net (FG-GAN) CG-AGG is first based on a hypergraph neural network-based global aggregator and a graph neural network-based local aggregator to generate entity representations with multiple semantics, while FG-GAN further enhances entity representations with specific semantics through a conditional generative adversary net. Experimental results on benchmark datasets show that our model outperforms state-of-the-art models in induced KGC.

IJCAI（International Joint Conference on Artificial Intelligence）

The IJCAI will be one of the world’s most important scientific conferences in the field of artificial intelligence, covering a wide range of fields in artificial intelligence, mainly artificial intelligence, machine learning, knowledge representation, natural language processing, planning, search, statistical inference, knowledge-based systems, multi-agent systems and robotics. The IJCAI was first held in 1957 and has been held every two years since then.

Recent papers at the IJCAI that focus on areas such as machine learning, transition learning, knowledge distillation, deep learning and recommendation systems, which are discussed in “An overview of transition learning and examples of algorithms and implementations“, are listed below.

• “A Canonicalization-Enhanced Known Fact-Aware Framework For Open Knowledge Graph Link Prediction” Open Knowledge Graph (OpenKG) link prediction aims to predict missing fact triples in the form of (head noun phrase, relation phrase, tail noun phrase). As triples are not normalised, previous methods either normalise noun phrases (NPs) to reduce graph sparsity or use textual forms to improve type compatibility. However, they fail to normalise relation phrases (RPs) and triples, and OpenKG maintains high sparsity, hindering performance. To address the above issues, we propose a Canicalisation-Enhanced Known Fact-Aware (CEKFA) framework that improves link prediction performance through RP and triple sparsity reduction. First, we propose a similarity-driven RP normalisation method to reduce RP sparsity by sharing knowledge of semantically similar RPs. Second, to reduce triple sparsity, a known-fact-aware triple normalisation method is designed to retrieve relevant known facts from the training data. Finally, these two types of normalisation information are integrated into a general two-stage re-ranking framework that can be applied to most existing knowledge graph embedding methods.Experimental results on two OpenKG datasets, ReVerb20K and ReVerb45K, show that our approach achieves state-of-the-art results, showing that our approach achieves state-of-the-art results. Extensive experimental analysis demonstrates the effectiveness and generalisability of the proposed framework.
• “Generative Diffusion Models on Graphs: Methods and Applications” As a new generative paradigm, diffusion models have achieved remarkable success in a variety of image generation tasks, such as image inpainting, image-to-text translation and video generation. Graph generation is an important computational task on graphs with numerous real-world applications. The aim is to learn the distribution of a given graph and generate a new graph. Given the great success of diffusion models in image generation, there has been a growing effort in recent years to utilise these techniques to advance graph generation. This paper first provides a comprehensive overview of generative diffusion models on graphs. In particular, representative algorithms for three variants of graph diffusion models are reviewed: score matching with Langevin dynamics (SMLD), de-noised diffusion probability models (DDPMs) and score-based generative models (SGMs). It then summarises the main applications of generative diffusion models on graphs, with a particular focus on modelling molecules and proteins. Finally, promising directions for generative diffusion modelling of graph-structured data are discussed. The survey also created a website for the GitHub project by collecting supporting resources for generative diffusion models on graphs.
• “Adaptive Path-Memory Network for Temporal Knowledge Graph Reasoning” Temporal Knowledge Graph (TKG) reasoning aims to predict future missing facts based on historical information and has recently attracted increased research interest. Much work has been done to model the historical structural and temporal features of the inference task. Most existing work models the graph structure primarily according to entity representation. However, the size of TKG entities in real-world scenarios is substantial and the number of new entities increases over time. Therefore, we propose a new architectural modelling with TKG relational capabilities: the aDAptivE path-MemOry Network (DaeMon), which adaptively models temporal path information between query subjects and each candidate object over history time. It models historical information independently of entity representation. Specifically, DaeMon uses path memory to record temporal path information obtained from path aggregation units across the timeline, taking into account memory passing strategies between neighbouring timestamps. extensive experiments conducted on four real-world TKG datasets show that the proposed model significantly performance improvements and show that it outperforms the absolute performance by up to 4.8% in MRR