Dictionary Multi-Modal Temporal Graph Learning.

Temporal Graph Clustering (TGC) is a new task with little attention, focusing on node clustering in temporal graphs. Compared with existing static graph clustering, it can find the balance between time requirement and space requirement (Time-Space Balance) through the interaction sequence-based batch-processing pattern. However, there are two major challenges that hinder the development of TGC, i.e., inapplicable clustering techniques and inapplicable datasets. To address these challenges, we propose a comprehensive benchmark, called BenchTGC. Specially, we design a BenchTGC Framework to illustrate the paradigm of temporal graph clustering and improve existing clustering techniques to fit temporal graphs. In addition, we also discuss problems with public temporal graph datasets and develop multiple datasets suitable for TGC task, called BenchTGC Datasets. According to extensive experiments, we not only verify the advantages of BenchTGC, but also demonstrate the necessity and importance of TGC task. We wish to point out that the dynamically changing and complex scenarios in real world are the foundation of temporal graph clustering.

Representation learning on temporal graphs has drawn considerable research attention owing to its fundamental importance in a wide spectrum of real-world applications. Though a number of studies succeed in obtaining time-dependent representations, it still faces significant challenges. On the one hand, most of the existing methods restrict the embedding space with a certain curvature. However, the underlying geometry in fact shifts among the positive curvature hyperspherical, zero curvature Euclidean and negative curvature hyperbolic spaces in the evolvement over time. On the other hand, these methods usually require abundant labels to learn temporal representations, and thereby notably limit their wide use in the unlabeled graphs of the real applications. To bridge this gap, we make the first attempt to study the problem of self-supervised temporal graph representation learning in the general Riemannian space, supporting the time-varying curvature to shift among hyperspherical, Euclidean and hyperbolic spaces. In this paper, we present a novel self-supervised Riemannian graph neural network (SelfRGNN). Specifically, we design a curvature-varying Riemannian GNN with a theoretically grounded time encoding, and formulate a functional curvature over time to model the evolvement shifting among the positive, zero and negative curvature spaces. To enable the self-supervised learning, we propose a novel reweighting self-contrastive approach, exploring the Riemannian space itself without augmentation, and propose an edge-based self-supervised curvature learning with the Ricci curvature. Extensive experiments show the superiority of SelfRGNN, and moreover, the case study shows the time-varying curvature of temporal graph in reality.

TANGO: A temporal spatial dynamic graph model for event prediction

Recent years, there is a growing interest in knowledge graph embedding (KGE), which maps symbolic entities and relations into low-dimensional vector space to effectively represent structured data from the knowledge graph. In addition, the concept of temporal knowledge graph is proposed to document dynamically changing facts in the real world. Existing works attempt to incorporate temporal information into static KGE methods to accomplish temporal knowledge representations. However, existing static or temporal KGE approaches focus on the single query fact and ignore the query-relevant contextual information in the graph structure. This paper moves beyond the traditional way of scoring facts in distinct vector space and proposes a unified framework with pre-trained language models (PLM) to learn dynamic <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">co</b>ntextualized <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">s</b>tatic/<bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">t</b>emporal <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</b>nowledge <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">g</b>raph <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">e</b>mbeddings, called CoS/TKGE. Given the query-specific subgraph, our model transforms it into an input sequence and uses the PLM to obtain the contextualized knowledge representations, which is flexible adaptive to the input graph contexts. We reformulate the link prediction task as a mask prediction problem to fine-tune the pre-trained language model. And the contrastive learning technique is employed to align dynamic contextual embeddings with static global embeddings. Experimental results on three widely used static and temporal KG datasets show the superiority of our model.

DyHDGE: Dynamic heterogeneous transaction graph embedding for safety-centric fraud detection in financial scenarios

Trajectory planning and tracking of lane change are critical technologies for autonomous buses. Characteristics of the buses susceptible to stability problems resulting from the high height, large passenger capacity and long lengths, coupling the dynamic traffic with the dynamic changes in the states of adjacent vehicles and road adhesion coefficient, put forward great challenges in lane change for autonomous buses (ABs). To cope with the foregoing challenges, a framework is proposed to achieve the trajectory planning and tracking of dynamic lane change for ABs. For trajectory planning approach, the trajectory planning and replanning is optimized in the safe range of longitudinal length of the lane-changing trajectory to obtain the real-time reference trajectory, combining consideration of vehicle yaw, roll stability and lane-changing efficiency. The minimum longitudinal length of lane-changing trajectory determined by the yaw stability and roll stability of ABs, combined with the maximum length formed by the adjacent vehicles with dynamic states, form the real-time safe range for lane-changing trajectory planning. For trajectory tracking approach, a tracking approach using model predictive control based on multipoint preview is proposed to achieve the real-time planned trajectory tracking considering buses stability. The effectiveness of the proposed strategy is evaluated by simulating an experimentally validated Trucksim model in different dynamic traffic scenarios to demonstrate the capability of the strategy in trajectory planning and tracking, and guaranteeing vehicle stability for dynamic lane change of ABs.

Temporal interaction graphs (TIGs), defined by sequences of timestamped interaction events, have become ubiquitous in real-world applications due to their capability to model complex dynamic system behaviors. As a result, temporal interaction graph representation learning (TIGRL) has garnered significant attention in recent years. TIGRL aims to embed nodes in TIGs into low-dimensional representations that effectively preserve both structural and temporal information, thereby enhancing the performance of downstream tasks such as classification, prediction, and clustering within constantly evolving data environments. In this paper, we begin by introducing the foundational concepts of TIGs and emphasizing the critical role of temporal dependencies. We then propose a comprehensive taxonomy of state-of-the-art TIGRL methods, systematically categorizing them based on the types of information utilized during the learning process to address the unique challenges inherent to TIGs. To facilitate further research and practical applications, we curate the source of datasets and benchmarks, providing valuable resources for empirical investigations. Finally, we examine key open challenges and explore promising research directions in TIGRL, laying the groundwork for future advancements that have the potential to shape the evolution of this field.

The development of activity recognition based on multi-modal data makes it possible to reduce human intervention in the process of monitoring. This paper proposes an efficient and cost-effective multi-modal sensing framework for activity monitoring, it can automatically identify human activities based on multi-modal data, and provide help to patients with moderate disabilities. The multi-modal sensing framework for activity monitoring relies on parallel processing of videos and inertial data. A new supervised adaptive multi-modal fusion method (AMFM) is used to process multi-modal human activity data. Spatio-temporal graph convolution network with adaptive loss function (ALSTGCN) is proposed to extract skeleton sequence features, and long short-term memory fully convolutional network (LSTM-FCN) module with adaptive loss function is adapted to extract inertial data features. An adaptive learning method is proposed at the decision level to learn the contribution of the two modalities to the classification results. The effectiveness of the algorithm is demonstrated on two public multi-modal datasets (UTD-MHAD and C-MHAD) and a new multi-modal dataset H-MHAD collected from our laboratory. The results show that the performance of the AMFM approach on three datasets is better than the performance of the video or the inertial-based single-modality model. The class-balanced cross-entropy loss function further improves the model performance based on the H-MHAD dataset. The accuracy of action recognition is 91.18%, and the recall rate of falling activity is 100%. The results illustrate that using multiple heterogeneous sensors to realize automatic process monitoring is a feasible alternative to the manual response.

SummaryAnalysis of multi-modal datasets can identify multi-scale interactions underlying biological systems, but can be beset by spurious connections due to indirect impacts propagating through an unmapped biological network. For example, studies in macaques have shown that BCG vaccination by an intravenous route protects against tuberculosis, correlating with changes across various immune data modes. To eliminate spurious correlations and identify critical immune interactions in a public multi-modal dataset (systems serology, cytokines, cytometry) of vaccinated macaques, we applied Markov Fields (MF), a data-driven approach that explains vaccine efficacy and immune correlations via multivariate network paths, without requiring large numbers of samples (i.e. macaques) relative to multivariate features. Furthermore, we find that integrating multiple data modes with MFs helps to remove spurious connections. Finally, we used the MF to predict outcomes of perturbations at various immune nodes, including a B-cell depletion that induced network-wide shifts without reducing vaccine protection, which we validated experimentally.

Link prediction in Online Social Networks—OSNs—has been the focus of numerous studies in the machine learning community. A successful machine learning-based solution for this task needs to (i) leverage global and local properties of the graph structure surrounding links; (ii) leverage the content produced by OSN users; and (iii) allow their representations to change over time, as thousands of new links between users and new content like textual posts, comments, images and videos are created/uploaded every month. Current works have successfully leveraged the structural information but only a few have also taken into account the textual content and/or the dynamicity of network structure and node attributes. In this paper, we propose a methodology based on temporal graph neural networks to handle the challenges described above. To understand the impact of textual content on this task, we provide a novel pipeline to include textual information alongside the structural one with the usage of BERT language models, dense preprocessing layers, and an effective post-processing decoder. We conducted the evaluation on a novel dataset gathered from an emerging blockchain-based online social network, using a live-update setting that takes into account the evolving nature of data and models. The dataset serves as a useful testing ground for link prediction evaluation because it provides high-resolution temporal information on link creation and textual content, characteristics hard to find in current benchmark datasets. Our results show that temporal graph learning is a promising solution for dynamic link prediction with text. Indeed, combining textual features and dynamic Graph Neural Networks—GNNs—leads to the best performances over time. On average, the textual content can enhance the performance of a dynamic GNN by 3.1% and, as the collection of documents increases in size over time, help even models that do not consider the structural information of the network.

In this paper, we systemically disentangle BHG, a graph deep learning framework for daily users' ads recommendations. BHG mainly relies on two pillars: (1) graph tokenization to convert the input temporal heterogeneous graph into sequences of tokens, and (2) graph MLP-Mixer neural architecture to learn node representations on sequences of tokens via a mini-batch manner. In general, BHG embraces three advantages: (1) flexibility, i.e., BHG can be seamlessly integrated with any existing industrial recommendation model by treating the learned node embeddings as additional features that encode interactions, (2) efficiency, i.e., the graph tokenization allows sampling the neighborhood both locally and globally, and reduces the number of nodes considered for aggregations, and (3) model simplicity, i.e., the graph MLP-Mixer does not require self-attention for aggregating nodes and hence enjoys the simplicity. We demonstrate the superior performance of the proposed BHG on two internal datasets and one public dataset. We hope this paper can share insights and explain large-scale graph deep learning deployments for researchers, engineers, and practitioners.

Hypothesis Generation (HG) aims to expedite biomedical researches by generating novel hypotheses from existing scientific literature. Most existing studies focused on modeling static snapshots of the corpus, neglecting the temporal evolution of scientific terms. Despite recent efforts to learn term evolution from Knowledge Bases (KBs) for HG, the temporal information from multi-source KBs is still overlooked, which contains important, up-to-date knowledge. In this paper, an innovative Temporal Contrastive Learning (TCL) framework is introduced to uncover latent associations between entities by jointly modeling their co-evolution across multi-source temporal KBs. Specifically, we first construct a temporal relation graph based on PubMed papers and a biomedical relation database (such as Comparative Toxicogenomics Database (CTD)). Then the constructed temporal relation graph and a temporal concept graph (such as Medical Subject Headings (MeSH)) are used to train two GCN-based recurrent networks for learning the entity temporal evolutional embeddings, respectively. Finally, a cross-view temporal prediction task is designed for learning knowledge enriched temporal embeddings by contrasting the temporal embeddings learned from the two Temporal Knowledge Graphs (TKGs). Findings from experiments conducted on three real-world biomedical term relationship datasets demonstrate that the proposed approach is clearly superior to approaches based on single TKG, achieving the state-of-the-art performance.

Since the static Knowledge Graph cannot meet the dynamics of knowledge in the real world, Temporal Knowledge Graph has become a potential method for processing temporal knowledge. However, Temporal Knowledge Graph still faces the problem of incompleteness. Hence knowledge completion has become the main reasoning task for computing the missing facts in Knowledge Graph. In order to capture more richer historical information for predicting future events, we propose a new event forecasting model called Global-Recent Historical Network (GRHNet). In GRHNet, we use statistical methods to simulate the evolution of events. Furthermore, we design a global history learner to capture the repeatability of knowledge, and a recent history learner to capture time-variability of knowledge. GRHNet is evaluated on two benchmark datasets. The results demonstrate that, compared with state-of-the-art baselines, GRHNet achieves at least a 3% relative improvement in MRR.

Recently, large-scale knowledge graphs (KGs) have become a key asset for search, analytics, recommendations and data integration. Large-scale KGs provide millions of facts about the real world. Each fact is composed as (subject, relation, object), e.g., the triplet (“Cristiano Ronaldo”, playFor, “Real Madrid”). However, these facts are blind to the temporal dimension. Actually, knowledge in practice is time-variant and many relations are only valid for a certain period of time. This phenomenon highlights the importance of building temporal knowledge graphs. In particular, knowledge in temporal KG is represented as (subject, relation, object, valid time), e.g., (“Cristiano Ronaldo”, playFor, “Real Madrid”, “[2009, 2018]”. However, research on temporal KG is very current and there are still many problems needed to be addressed. One obvious problem is that the size of temporal KG is still very small. For example, only 6.6% of the facts are time-aware in one of the largest knowledge graphs, YAGO3. In addition, 71% of people have no known place of birth, and 75% have no known nationality in Freebase. Furthermore, over 87.7% of the facts are uncovered in Japanese DBpedia compared with English DBpedia. Therefore, in this thesis, we study how to enlarge and enrich temporal knowledge graphs from three aspects, namely, temporal KG enrichment. In particular, we study the enrichment problem from the following aspects: (1) volume, (2) completeness, and (3) coverage.Our first solution is temporal knowledge harvesting which extracts temporal knowledge from free text directly. However, text corpus is noisy, and extracting structured temporal facts with high accuracy and coverage is very challenging. Inspired by pattern-based systems, we propose a temporal knowledge harvesting framework. In particular, we propose various techniques to extract temporal patterns, including corpus annotation, pattern generation, scoring and clustering. These techniques can reduce ambiguity in the text corpus and can improve both the accuracy and coverage of the extracted patterns. Second, we leverage the extracted patterns to harvest temporal knowledge. To improve the accuracy, we propose a parse-tree-based method. And to increase the coverage, we consider the relationships between tree components, including part of speech (POS), clause types, constituency and dependency. Experiments on real-world datasets verify the effectiveness of our proposed framework.Our second solution is temporal knowledge graph completion. Temporal KG completion is the task of inferring unobserved edges between entity pairs. Generally, temporal KG completion relies on the temporal KG embeddings technique, which learns the low-dimensional representations of all KG components. As research on temporal KG embedding is very current, we study both the embedding and the completion problems. In particular, we observe that contexts are extremely useful for learning the representations of entities and for inferring the unknown time intervals. As a result, we propose a context-aware embedding model for KG embeddings and a context-based temporal inference model for KG completion. In our embedding model, we not only capture factual plausibility as traditional methods did, but also propose a new measure on contexts, named temporal consistency. It measures how well the target entity interacts with its surrounding contexts on the temporal dimension. Our completion model is based on the embedding model, and further captures the interactions on the entity dimension. Extensive experiments verify the effectiveness of our models.Our third solution is temporal knowledge graph alignment which aims to discover the SameAs edges across two temporal KGs. Not like previous attribute-based alignment models, we further divide attribute facts into character facts, digit facts and time facts. In particular, we observe that the context information is extremely useful for the identification of the same entities. Therefore, we propose an alignment model on leveraging temporal contexts to represent entities. However, contexts of the same object can be very different. For example, the value of career predicate changed from “football player” to “coach” for the football player “Zidane”. This is because entities are evolving over time and thus the predicate values can be different. In our framework, we propose an alignment module that simulates the entity evolving process. Specifically, this module captures the interactions between contexts and aggregates context information to represent the entity. Lastly, we found that not all contexts are relevant, e.g., height v.s. career. Actually, only the relevant contexts are useful for capturing the evolving. Therefore, we propose a clustering approach for grouping relevant contexts together. Our experimental results validate the superiority of our proposed alignment model.

Temporal Knowledge Graph (TKG) is composed of a series of facts related to timestamps in the real world and has become the basis of many artificial intelligence applications. However, the existing TKG is usually incomplete. It has become a hot research task to infer missing facts based on existing facts in a TKG; namely, Temporal Knowledge Graph Completion (TKGC). The current mainstream TKGC models are embedded models that predict missing facts by representing entities, relations and timestamps as low-dimensional vectors. In order to deal with the TKG structure information, there are some models that try to introduce attention mechanism into the embedding process. But they only consider the structure information of entities or relations, and ignore the structure information of the whole TKG. Moreover, most of them usually treat timestamps as a general feature and cannot take advantage of the potential time series information of the timestamp. To solve these problems, wo propose a new Hierarchical Self-Attention Embedding (HSAE) model which inspired by self-attention mechanism and diachronic embedding technique. For structure information of the whole TKG, we divide the TKG into two layers: entity layer and relation layer, and then apply the self-attention mechanism to the entity layer and relation layer respectively to capture the structure information. For time series information of the timestamp, we capture them by combining positional encoding and diachronic embedding technique into the above two self-attention layers. Finally, we can get the embedded representation vectors of entities, relations and timestamps, which can be combined with other models for better results. We evaluate our model on three TKG datasets: ICEWS14, ICEWS05-15 and GDELT. Experimental results on the TKGC (interpolation) task demonstrate that our model achieves state-of-the-art results.