Latest ArticlesBlockchain enables traditional exchanges and lending houses to be extended to Depl (decentralized platforms), which allows anyone to access exchange and lending without the help of intermediaries. Most Depls are implemented as smart contracts running on Ethereum and interact with another smart contract, cryptocurrency (i.e. token), to achieve various functions. Although Depl involves more than 35 billion worth of tokens, little is known about whether the actual transfer of tokens is as consistent as Depl expects. The inconsistency between the actual transfer of tokens and what the decentralized platform expects is known as behavioral inconsistency, resulting in property damage and user confusion. In this work, we take the first step to investigate such inconsistency between Depl and tokens. We propose to automatically detect inconsistency by comparing the actual token transfer behavior with the behavior indicated by the internal records of Depl by monitoring the core data structure changes of Depl and token.The experimental results show that inconsistent behavior exists in 1012749 transactions with an accuracy of 98.0%, which involves 2871 pairs of Depl and tokens, and is related to 110 Depl and 2544 tokens.In addition, 10 main reasons behind the inconsistency are summarized, e.g., cheater Depl, inconsistent scale, unclear coin lock rules, etc.
Evaluation metrics are crucial in the field of code synthesis. Commonly used code evaluation metrics can be classified into three types: match-based, semantic-based, and execution-based. Among them, the execution-based Pass@k metric accurately assesses the functionality of predicted code by executing test cases. However, calculating this metric requires a significant amount of overhead, necessitating the design of an automated evaluation metric that can assess the functionality of predicted code without the need for test cases. Additionally, a good evaluation metric should be robust, that is the metric can maintain its accuracy even when the predicted code undergoes minor changes. To address these challenges, we propose an automated robust metric, called CodeScore-R, based on UniXcoder and contrastive learning, for evaluating the functionality of code synthesis. CodeScore-R employs techniques such as sketch-based processing, syntactic-equivalent transformations, and mutation testing to effectively mitigate the interference caused by identifiers, syntax structures, and operators on evaluation results. Experimental results demonstrate that in the tasks of code generation and migration in Java and Python, CodeScore-R outperforms other evaluation metrics and is more closely aligned with the Pass@k metric, while exhibiting stronger robustness.
Code summarization is a natural language description of source code, and high-quality code summaries help to improve developers’ program understanding efficiency. In recent years, research on code summarization has focused on generating summaries for method-grained code snippet. However, in an object-oriented language such as Java, class is the basic programming unit. Due to the above problems, we propose a class summarization generation method based on hierarchical representation and context enhancement, called HRCE, as well as constructs a class summarization dataset containing 358992 pairs of <Java class, content, summary>. HRCE uses code simplification strategy to remove non-critical code of class to shorten the code length. Then, HRCE models the class hierarchy, including class signature, attribute and method respectively, to obtain the semantic information and hierarchical structure information of the class. In addition, HRCE selects parent’s class signature and class summary to describe the context that the class depends on in the project. Experiments show that a generative model for class summarization based on hierarchical representation and context enhancement is able to characterize the semantics and hierarchical structure of the code, and obtain information from both inside and outside of the target class. As a result, HRCE outperforms all baseline models on evaluation metrics such as BLEU, METEOR, ROUGE-L, etc.
In the field of data mining, the issue of data imbalance impacting model prediction accuracy is widespread, and also the issue of user privacy protection is neglected. Fake dataset generation has come to light as a crucial remedy for these problems. However, because of the traits of high-dimensional and irrelevant features, it is difficult to generate high-quality data in circumstances where structured data predominate. Considering the successful applications of the diffusion model in image generation task, we aim to apply the diffusion model for the task of customer churn prediction, which is a typical scenario in data mining. we utilize the Gaussian diffusion model and polynomial diffusion model to generate data for numerical and categorical features in customer churn data. Research and analysis have been conducted on the predictive performance and data privacy protection capabilities of our model. We conduct extensive experiments on customer churn data from multiple domains to explore the potential of fusing synthetic dataset and real dataset for data reconstruction. The results demonstrate that the diffusion model can generate high-quality data and improve the performance of various prediction methods, which can help alleviate the issue of data imbalance. Additionally, the data produced by the diffusion model exhibit a distribution that is quite similar to the original dataset, which may be useful for protecting user privacy.
Requirements elicitation and modeling is a critical step in requirements engineering that affects subsequent system design and implementation. Traditional requirements elicitation and modeling methods usually involve multiple types of stakeholders, such as requirement providers and requirements analysts, working together in an iterative manner, which requires a lot of manpower. It is important to reduce the burden of requirements providers and requirements analysts and improve the efficiency of elicitation and modeling. Some of the existing efforts use knowledge bases to provide more knowledge to assist modeling or elicitation, and some use natural language processing techniques to automate the elicitation or modeling process. But they do not reduce the burden of requirements providers. Leveraging the generative capability of LLMs (large language models), we provide ChatModeler, a requirement elicitation and modeling framework for iterative human-machine collaboration. Specifically, based on the division of labor and collaboration relationship of real-world requirements teams, some work of requirements providers and requirements analysts is taken up by LLMs, while requirements providers only need to perform confirmation. In this paper, prompts are designed for the various roles played by the various LLMs, which varies with the meta-model of the requirements. ChatModeler is used in 14 sets of comparative experiments with state-of-the-art automated modeling approaches for three types of requirement models on seven requirement cases, demonstrating the superiority of ChatModeler in both reducing the number of interactions of the requirements provider and generating higher quality requirement models.
In recent years, the emergence and development of large language models (LLMs) have revolutionized the field of natural language processing and even artificial intelligence. With the increasing number of model parameters and training data, the perplexity of language models decreases in a predictable manner, which implies the improvement of performance on various natural language processing tasks. Therefore, scaling up language models has been a promising way to improve the system intelligence. In this survey, we first review the definition and scope of LLMs and provide a scale standard to distinguish “large” language models from the perspectives of performance and computing. Then, we review the development and representative work of LLMs in three dimensions: data, algorithm, and model architecture, showing how up-scaling in these dimensions drives the development of LLMs at different stages. Next, we discuss the emergent abilities of LLMs and possible interpretations behind them. We highlight three key emergent abilities, i.e., chain-of-thought prompting, in-context learning, and instruction-following, introducing their related advances and applications. Finally, we outline some potential directions and challenges of LLMs.
In recent years, graph neural networks have achieved remarkable results in application fields such as recommendation systems and natural language processing with the help of large amounts of data and supercomputing power, and they mainly deal with graph data with pairwise relationships. However, in many real-world networks, the relationships between objects are more complex and beyond pairwise, such as scientific collaboration networks, protein networks, and others. If we directly use a graph to represent the complex relationships as pairwise relations, which will lead to a loss of information. Hypergraph is a flexible modeling tool, which shows higher-order relationships that cannot be fully described by a graph, making up for the shortage of graph. In light of this, scholars begin to care about how to develop neural networks on hypergraph, and successively put forward many hypergraph neural network models. Therefore, we overview the existing hypergraph neural network models. Firstly, we comprehensively review the development of the hypergraph neural network in the past three years. Secondly, we propose a new classification method according to the design method of hypergraph neural networks, and elaborate on representative models. Then, we introduce the application areas of hypergraph neural networks. Finally, the future research direction of hypergraph neural networks are summarized and discussed.
With the increasing popularity of intelligent mobile terminals, people increasingly use social network platforms (such as Twitter, Sina Weibo, etc.) for information acquisition, comments, and exchanges. Although GPS devices can accurately obtain location information, many users do not directly share their location information for privacy and security considerations. Therefore, obtaining the geographic location of online users has become an important topic in both academia and industry and is the foundation of many downstream applications, such as location-based targeted advertising, event/location recommendations, early warning of natural disasters or diseases, and criminal tracking, etc. We survey in detail the methods, data types, evaluation metrics, and fundamental algorithms for predicting the geographic location of social network users. First, we discriminate different online user geolocation tasks and corresponding evaluation protocols. Subsequently, we assess the data structures and fusion methods used for individual geolocation tasks. Besides, we analyze the existing information extraction and feature selection approaches, as well as their advantages and disadvantages. Moreover, we provide a taxonomy to categorize existing user geolocation models and algorithms, followed by a thorough analysis of different methods from three aspects: geographic dictionary, traditional machine learning, deep learning and graph neural networks. Finally, we summarize the difficulties and challenges in user location prediction while outlining the possible research trend and opportunities to shed light on future work in this field.
Cardinality estimation is the basis and core of query optimizer for the database management system (DBMS). With the development of artificial intelligence (AI) technology, AI technology has shown superior performance in data processing and extracting the relationship from the data. In recent years, the research of the cardinality estimation method based on machine learning has made significant progress and received wide attention from the academic community. Firstly, we introduce the technical background and development status of cardinality estimation methods based on machine learning. Secondly, we give the definition and the feature encoding technology of the related concepts of cardinality estimation. Then, we expound on the classification structure of cardinality estimation technology from two aspects: traditional cardinality estimation and cardinality estimation based on machine learning. Then, we further subdivide cardinality estimation based on machine learning into three types of cardinality estimation techniques: query-driven, data-driven, and hybrid models. Then, we focus on analyzing the modeling flow, typical methodologies, and characteristics of each type of model. In addition, we analyze and summarize the application of cardinality estimation in SQL and NoSQL. Finally, we discuss the challenges and future research directions on cardinality estimation methods based on machine learning.
Fingerprint features and localization model are two key factors affecting the localization accuracy in fingerprint localization methods. In terms of fingerprint feature, the stability of visible light intensity is high, but the discrimination of position features is low. The radio signal strength is highly distinguishable, but has strong volatility. Meanwhile, the localization models based on convolutional neural network (CNN) can not effectively highlight important features during feature extraction. In view of the above problems, a hybrid fingerprint indoor localization method based on ECA-CNN (ECACon-HF) is proposed in this paper. First, we use visible light intensity and received signal strength indication (RSSI) of Bluetooth low energy (BLE) to construct hybrid fingerprints, reduce the influence of instability of BLE fingerprints, and enhance the discrimination between different positions. Meanwhile, the CNN localization model is improved by the efficient channel attention (ECA). ECA can adaptively extract important information from fingerprints through cross-channel interaction strategies, suppress environmental interference in fingerprints, enhance the expression ability of hybrid fingerprint features, and make more effective use of the advantages of hybrid fingerprints. The experimental results show that ECACon-HF proposed in this paper achieves the localization precision of 0.316 m, higher accuracy than on the single fingerprint. Meanwhile, based on the same fingerprint database, the proposed method outperforms other related indoor localization methods.