Knowledge and Information Systems最新文献

Recently, the impact of product or service reviews on customers' purchasing decisions has become increasingly significant in online businesses. Consequently, manipulating reviews for fame or profit has become prevalent, with some businesses resorting to paying fake reviewers to post spam reviews. Given the importance of reviews in decision-making, detecting fake reviews is crucial to ensure fair competition and sustainable e-business practices. Although significant efforts have been made in the last decade to distinguish credible reviews from fake ones, it remains challenging. Our literature review has identified several gaps in the existing research: (1) most fake review detection techniques have been proposed for high-resource languages such as English and Chinese, and few studies have investigated low-resource and multilingual fake review detection, (2) there is a lack of research on deceptive review detection for reviews based on language code-switching (code-mix), (3) current multi-feature integration techniques extract review representations independently, ignoring correlations between them, and (4) there is a lack of a consolidated model that can mutually learn from review emotion, coarse-grained (overall rating), and fine-grained (aspect ratings) features to supplement the problem of sentiment and overall rating inconsistency. In light of these gaps, this study aims to provide an in-depth literature analysis describing strengths and weaknesses, open issues, and future research directions.

As deep neural networks for visual recognition gain momentum, many studies have modified the loss function to improve the classification performance on long-tailed data. Typical and effective improvement strategies are to assign different weights to different classes or samples, yielding a series of cost-sensitive re-weighting cross-entropy losses. Granted, most of these strategies only focus on the properties of the training data, such as the data distribution and the samples’ distinguishability. This paper works these strategies into a weighted cross-entropy loss framework with a simple production form ((text {WCEL}_{prod })), which takes into account different features of different losses. Also, there is this new loss function, misclassification-guided loss (MGL), that generalizes the class-wise difficulty-balanced loss and utilizes the misclassification rate on validation data to update class weights during training. In respect of MGL, a series of weighting functions with different relative preferences are introduced. Both softmax MGL and sigmoid MGL are derived to address the multi-class and multi-label classification problems. Experiments are undertaken on four public datasets, namely MNIST-LT, CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, and a self-built dataset of 4 main-classes, 44 sub-classes, and altogether 57,944 images, where the results show that on the self-built dataset, the exponential weighting function achieves higher balanced accuracy than the polynomial function does. Ablation studies also show that MGL sees better performance in combination with most of other state-of-the-art loss functions under the (text {WCEL}_{prod }) framework.

Many time series data mining algorithms work by reasoning about the relationships the conserved shapes of subsequences. To facilitate this, the Matrix Profile is a data structure that annotates a time series by recording each subsequence’s Euclidean distance to its nearest neighbor. In recent years, the community has shown that using the Matrix Profile it is possible to discover many useful properties of a time series, including repeated behaviors (motifs), anomalies, evolving patterns, regimes, etc. However, the Matrix Profile is limited to representing the relationship between the subsequence’s shapes. It is understood that, for some domains, useful information is conserved not in the subsequence’s shapes, but in the subsequence’s features. In recent years, a new set of features for time series called catch22 has revolutionized feature-based mining of time series. Combining these two ideas seems to offer many possibilities for novel data mining applications; however, there are two difficulties in attempting this. A direct application of the Matrix Profile with the catch22 features would be prohibitively slow. Less obviously, as we will demonstrate, in almost all domains, using all twenty-two of the catch22 features produces poor results, and we must somehow select the subset appropriate for the domain. In this work, we introduce novel algorithms to solve both problems and demonstrate that, for most domains, the proposed C²²MP is a state-of-the-art anomaly detector.

Large language models (LLMs) have transformed the interpretation and creation of human language in the rapidly developing field of computerized language processing. These models, which are based on deep learning techniques like transformer architectures, have been painstakingly trained on massive text datasets. This study paper takes an in-depth look into LLMs, including their architecture, historical evolution, and applications in education, healthcare, and finance sector. LLMs provide logical replies by interpreting complicated verbal patterns, making them beneficial in a variety of real-world scenarios. Their development and implementation, however, raise ethical concerns and have societal ramifications. Understanding the importance and limitations of LLMs is critical for guiding future research and ensuring the ethical use of their enormous potential. This survey exposes the influence of these models as they change, providing a roadmap for researchers, developers, and policymakers navigating the world of artificial intelligence and language processing.

Ensemble learning has gotten a lot of interest because of its capacity to increase predictive accuracy by merging numerous models. However, redundant data and a high level of computing complexity frequently plague ensembles. To choose a subset of models while maintaining the accuracy and diversity of the ensemble, ensemble pruning techniques are used to address these problems. Accuracy and diversity must coexist, even though their goals are conflicting. This is why we formulate the issue of ensemble pruning as a dual-objective maximization problem using the idea from information theory. Then, we propose a Comprehensive Ensemble Pruning Framework (CEPF) based on the dual-objective maximization (DOM) trade-off metric. Extensive evaluation of our framework on the exclusively collected PhysioSense dataset demonstrates the superiority of our method compared to existing pruning techniques. PhysioSense dataset was collected after getting approval from the Institutional Human Ethics Committee (IHEC) of Panimalar Medical College Hospital and Research Institute, Chennai, Tamil Nadu (Protocol No: PMCHRI-IHEC-059). The proposed framework not only preserves or improves ensemble accuracy and diversity but also achieves a significant reduction in actual ensemble size. Furthermore, the proposed method provides valuable insights into the dual-objective trade-off between accuracy and diversity paving the way for further research and advancements in ensemble pruning techniques.

Sentiment Summarization is an automated technology that extracts important features of sentences and then reorganizes selected words or sentences by their aspect class and sentiment polarity. This emerging research area wields considerable influence, where a sentiment-based summary can provide insight into users’ subjective opinions, creating social engagement that benefits industry players and entrepreneurs. Meanwhile, systematic studies examining sentiment-based summarization, particularly those delving into aspect levels, are still limited. Whereas aspects are crucial to obtain a comprehensive assessment of a product or service for improving sentiment summarization results. Hence, we conducted a comprehensive survey of aspect extraction techniques in sentiment summarization by classifying techniques based on sentiment analysis levels and features. This work analyzes the current research trends and challenges in the research domain from a different perspective. More than 150 literature published from 2004 to 2023 are collected mainly from credible academic databases. We summarized and performed a comparative analysis of the sentiment summarization approaches and tabulated their performance based on different domains, sentiment levels, and features. We also derived a thematic taxonomy of aspect extraction techniques in sentiment summarization from the analysis and illustrated its usage in various applications. Finally, this study presents recommendations for the challenges and opportunities for future research development.

This paper identifies trends in the application of big data in the transport sector and categorizes research work across scientific subfields. The systematic analysis considered literature published between 2012 and 2022. A total of 2671 studies were evaluated from a dataset of 3532 collected papers, and bibliometric techniques were applied to capture the evolution of research interest over the years and identify the most influential studies. The proposed unsupervised classification model defined categories and classified the relevant articles based on their particular scientific interest using representative keywords from the title, abstract, and keywords (referred to as top words). The model’s performance was verified with an accuracy of 91% using Naïve Bayesian and Convolutional Neural Networks approach. The analysis identified eight research topics, with urban transport planning and smart city applications being the dominant categories. This paper contributes to the literature by proposing a methodology for literature analysis, identifying emerging scientific areas, and highlighting potential directions for future research.

Few-shot knowledge graph completion (FKGC) refers to the task of inferring missing facts in a knowledge graph by utilizing a limited number of reference entities. Most FKGC methods assume a single similarity metric, which leads to a single feature space and makes it difficult to separate positive and negative samples effectively. Therefore, we propose a multi-scale relational metric network (MSRMN) specifically designed for FKGC, which integrates multiple scales of measurement methods to learn a more comprehensive and compact feature space. In this study, we design a complete neighbor random sampling algorithm to sample complete one-hop neighbor information, and aggregate both one-hop and multi-hop neighbor information to enhance entity representations. Then, MSRMN adaptively obtains prototype representations of relations and integrates three different scales of measurement methods to learn a more comprehensive feature space and a more discriminative feature mapping, enabling positive query entity pairs to obtain higher measurement scores. Evaluation of MSRMN on two public datasets for link prediction demonstrates that MSRMN attains top-performing outcomes across various few-shot sizes on the NELL dataset.

The identification of both closed frequent high average utility itemsets (CFHAUIs) and generators of frequent high average utility itemsets (GFHAUIs) has substantial significance because they play an essential and concise role in representing frequent high average utility itemsets (FHAUIs). These concise summaries offer a compact yet crucial overview that can be much smaller. In addition, they allow the generation of non-redundant high average utility association rules, a crucial factor for decision-makers to consider. However, difficulty arises from the complexity of discovering these representations, primarily because the average utility function does not satisfy both monotonic and anti-monotonic properties within each equivalence class, that is for itemsets sharing the same subset of transactions. To tackle this challenge, this paper proposes an innovative method for efficiently extracting CFHAUIs and GFHAUIs. This approach introduces novel bounds on the average utility, including a weak lower bound called (wlbau) and a lower bound named (auvlb). Efficient pruning strategies are also designed with the aim of early elimination of non-closed and/or non-generator FHAUIs based on the (wlbau) and (auvlb) bounds, leading to quicker execution and lower memory consumption. Additionally, the paper introduces a novel algorithm, CG-FHAUI, designed to concurrently discover both GFHAUIs and CFHAUIs. Empirical results highlight the superior performance of the proposed algorithm in terms of runtime, memory usage, and scalability when compared to a baseline algorithm.

Different embeddings capture various linguistic aspects, such as syntactic, semantic, and contextual information. Taking into account the diverse linguistic facets, we propose a novel hybrid model. This model hinges on the amalgamation of multiple embeddings through an attention encoder, subsequently channeled into an LSTM framework for sentiment classification. Our approach entails the fusion of Paragraph2vec, ELMo, and BERT embeddings to extract contextual information, while FastText is adeptly employed to capture syntactic characteristics. Subsequently, these embeddings were fused with the embeddings obtained from the attention encoder which forms the final embeddings. LSTM model is used for predicting the final classification. We conducted experiments utilizing both the Twitter Sentiment140 and Twitter US Airline Sentiment datasets. Our fusion model’s performance was evaluated and compared against established models such as LSTM, Bi-directional LSTM, BERT and Att-Coder. The test results clearly demonstrate that our approach surpasses the baseline models in terms of performance.