Nowadays, location-based services are widely used, requiring instant responses to a large volume of multiple spatial queries over massive road networks, i.e., single-pair shortest path (SPSP) query, k-nearest neighbor (kNN) query, and range query. Creating index-based structure for each kind of query is costly, hence it is important to handle multiple spatial queries within one efficient structure. Partition-based hierarchical approaches show promising potential to meet the requirement. However, existing approaches require large search space on massive road networks especially for long-distance queries, which is inefficient and hard to scale. To overcome the drawbacks, we propose the shortcut-enhanced graph hierarchy tree (SCG-tree), which leverages shortcuts to effectively prune the search space over a hierarchical structure. With the SCG-tree, a pruned shortcut-based method is designed to answer SPSP query, and a two-phase expansion strategy is proposed to leverage shortcuts for kNN and range queries. Theoretical analyses show the superiority of proposed shortcut-based query algorithms. Extensive experiments demonstrate that our approach can achieve three times speedup for kNN query and an order of magnitude speedup for SPSP and range queries over existing methods on real road networks that scale up to 24 million nodes and 58 million edges.
Table reasoning aims to generate inference results based on the user requirement and the provided table. Enhancing the table reasoning capability of the model can aid in obtaining information efficiently. Recent advancements have positioned Large Language Models (LLMs) as the predominant method for table reasoning, primarily due to their substantial reduction in annotation costs and superior performance compared to previous methods. However, existing research still lacks a summary of LLM-based table reasoning works. Therefore, questions about which techniques can improve table reasoning performance in the era of LLMs and how to enhance table reasoning abilities in the future, remain largely unexplored. This gap significantly limits progress in research. To answer the above questions and advance table reasoning research with LLMs, we present this survey to analyze existing research, inspiring future work. In this paper, we analyze the mainstream techniques used to improve table reasoning performance in the LLM era. Also, we provide research directions from the improvement of existing methods to inspire future research.
The learnware paradigm has been proposed as a new manner for reusing models from a market of various well-trained models, which can relieve users’ burden of training a new model from scratch. A learnware consists of a well-trained model and a specification which explains the purpose or specialty of the model without revealing data. By specification matching, the market can identify the most useful learnwares for users’ tasks. Prior art attempted to generate the specification by a reduced kernel mean embedding approach. However, such kind of specification is defined by some pre-designed kernel function, which lacks flexibility. In this paper, we advance a methodology for direct specification learning from data, introducing a novel neural network named SpecNet for this purpose. Our approach accepts unordered datasets as input and subsequently produces specification vectors in a latent space. Notably, the flexibility and efficiency of our learned specifications are underscored by their derivation from diverse tasks, rendering them particularly adept for learnware identification. Empirical studies provide validation for the efficacy of our proposed approach.
Estimating the Number of Distinct Values (NDVs) is a critical task in the fields of databases and data streams. Over time, various algorithms for estimating NDVs have been developed, each tailored to different requirements for time, I/O, and accuracy. These algorithms can be broadly categorized into two main types: sampling-based and sketch-based. Sampling-based NDV algorithms improve efficiency by sampling rather than accessing all items, often at the cost of reduced accuracy. In contrast, sketch-based NDV algorithms maintain a compact sketch using hashing to scan the entire dataset, typically offering higher accuracy but at the expense of increased I/O costs. When dealing with large-scale data, scanning the entire table may become infeasible. Thus, the challenge of efficiently and accurately estimating NDVs has persisted for decades. This paper provides a comprehensive review of the fundamental concepts, key techniques, and a comparative analysis of various NDV estimation algorithms. We first briefly examine traditional estimators in chronological order, followed by an in-depth discussion of the newer estimators developed over the past decade, highlighting the specific scenarios in which they are applicable. Furthermore, we illustrate how NDV estimation algorithms have been adapted to address the complexities of modern real-world data environments effectively. Despite significant progress in NDV estimation research, challenges remain in terms of theoretical scalability and practical application. This paper also explores potential future directions, including block sampling NDV estimation, learning-based NDV estimation, and their implications for database applications.
The landscape of financial transactions has grown increasingly complex due to the expansion of global economic integration and advancements in information technology. This complexity poses greater challenges in detecting and managing financial fraud. This review explores the role of Graph Neural Networks (GNNs) in addressing these challenges by proposing a unified framework that categorizes existing GNN methodologies applied to financial fraud detection. Specifically, by examining a series of detailed research questions, this review delves into the suitability of GNNs for financial fraud detection, their deployment in real-world scenarios, and the design considerations that enhance their effectiveness. This review reveals that GNNs are exceptionally adept at capturing complex relational patterns and dynamics within financial networks, significantly outperforming traditional fraud detection methods. Unlike previous surveys that often overlook the specific potentials of GNNs or address them only superficially, our review provides a comprehensive, structured analysis, distinctly focusing on the multifaceted applications and deployments of GNNs in financial fraud detection. This review not only highlights the potential of GNNs to improve fraud detection mechanisms but also identifies current gaps and outlines future research directions to enhance their deployment in financial systems. Through a structured review of over 100 studies, this review paper contributes to the understanding of GNN applications in financial fraud detection, offering insights into their adaptability and potential integration strategies.
Cloud-native data warehouses have revolutionized data analysis by enabling elasticity, high availability and lower costs. And the increasing popularity of artificial intelligence (AI) drives data warehouses to provide predictive analytics besides the existing descriptive analytics. Consequently, more vendors start to support training and inference of AI models in data warehouses, exploiting the benefits of near-data processing for fast model development and deployment. However, most of the existing solutions are limited by a complex syntax or slow data transportation across engines.
In this paper, we present GaussDB-AISQL, a composable SQL system with AI capabilities. GaussDB-AISQL adopts a composable system design that decouples computing, storage, caching, DB engine and AI engine. Our system offers all the functionality needed by end-to-end model training and inference during the model lifecycle. It also enjoys the simplicity and efficiency by providing a SQL-like syntax and removes the burden of manual model management. When training an AI model, GaussDB-AISQL benefits from highly parallel data transportation by concurrent data pulling from the distributed shared memory. The feature selection algorithms in GaussDB-AISQL make the training more data-efficient. When running model inference, GaussDB-AISQL registers the trained model object in the local data warehouse as a user-defined-function, which avoids moving inference data out of the data warehouse to an external AI engine. Experiments show that GaussDB-AISQL is up to 19× faster than baseline approaches.
The interconnection between query processing and data partitioning is pivotal for the acceleration of massive data processing during query execution, primarily by minimizing the number of scanned block files. Existing partitioning techniques predominantly focus on query accesses on numeric columns for constructing partitions, often overlooking non-numeric columns and thus limiting optimization potential. Additionally, these techniques, despite creating fine-grained partitions from representative queries to enhance system performance, experience from notable performance declines due to unpredictable fluctuations in future queries. To tackle these issues, we introduce LRP, a learned robust partitioning system for dynamic query processing. LRP first proposes a method for data and query encoding that captures comprehensive column access patterns from historical queries. It then employs Multi-Layer Perceptron and Long Short-Term Memory networks to predict shifts in the distribution of historical queries. To create high-quality, robust partitions based on these predictions, LRP adopts a greedy beam search algorithm for optimal partition division and implements a data redundancy mechanism to share frequently accessed data across partitions. Experimental evaluations reveal that LRP yields partitions with more stable performance under incoming queries and significantly surpasses state-of-the-art partitioning methods.
Many medical image segmentation methods have achieved impressive results. Yet, most existing methods do not take into account the shape-intensity prior information. This may lead to implausible segmentation results, in particular for images of unseen datasets. In this paper, we propose a novel approach to incorporate joint shape-intensity prior information into the segmentation network. Specifically, we first train a segmentation network (regarded as the teacher network) on class-wise averaged training images to extract valuable shape-intensity information, which is then transferred to a student segmentation network with the same network architecture as the teacher via knowledge distillation. In this way, the student network regarded as the final segmentation model can effectively integrate the shape-intensity prior information, yielding more accurate segmentation results. Despite its simplicity, experiments on five medical image segmentation tasks of different modalities demonstrate that the proposed Shape-Intensity Knowledge Distillation (SIKD) consistently improves several baseline models (including recent MaxStyle and SAMed) under intra-dataset evaluation, and significantly improves the cross-dataset generalization ability. The source code will be publicly available after acceptance.
We present an animatable 3D Gaussian representation for synthesizing high-fidelity human videos under novel views and poses in real time. Given multi-view videos of a human subject, we learn a collection of 3D Gaussians in the canonical space of the rest pose. Each Gaussian is associated with a few basic properties (i.e., position, opacity, scale, rotation, spherical harmonics coefficients) representing the average human appearance across all video frames, as well as a latent code and a set of blend weights for dynamic appearance correction and pose transformation. The latent code is fed to an Multi-layer Perceptron (MLP) with a target pose to correct Gaussians in the canonical space to capture appearance changes under the target pose. The corrected Gaussians are then transformed to the target pose using linear blend skinning (LBS) with their blend weights. High-fidelity human images under novel views and poses can be rendered in real time through Gaussian splatting. Compared to state-of-the-art NeRF-based methods, our animatable Gaussian representation produces more compelling results with well captured details, and achieves superior rendering performance.
