Benchmark Datasets Research Articles

ALDANER: Active Learning based Data Augmentation for Named Entity Recognition

Training Named Entity Recognition (NER) models typically necessitates the use of extensively annotated datasets. This requirement presents a significant challenge due to the labor-intensive and costly nature of manual annotation, especially in specialized domains such as medicine and finance. To address data scarcity, two strategies have emerged as effective: (1) Active Learning (AL), which autonomously identifies samples that would most enhance model performance if annotated, and (2) data augmentation, which automatically generates new samples. However, while AL reduces human effort, it does not eliminate it entirely, and data augmentation often leads to incomplete and noisy annotations, presenting new hurdles in NER model training. In this study, we integrate AL principles into a data augmentation framework, named Active Learning-based Data Augmentation for NER (ALDANER), to prioritize the selection of informative samples from an augmented pool and mitigate the impact of noisy annotations. Our experiments across various benchmark datasets and few-shot scenarios demonstrate that our approach surpasses several data augmentation baselines, offering insights into promising avenues for future research.

Hesitant fuzzy clustering with convolutional spiking neural network for movie recommendations

The movie recommender system is one of the most influential and practical tools for aiding individuals in quickly selecting films to watch. Despite numerous academic efforts to employ recommender systems for various purposes, such as movie-watching and book-buying, many studies have overlooked user-specific movie recommendations. This paper introduces a novel approach for movie recommendations that combines the hesitant fuzzy clustering with a convolutional spiking neural network movie recommender system. The initial step involves acquiring input data from benchmark datasets like MovieLens 100K and MovieLens 1M. Further, content-based features are extracted from the dataset using ternary pattern and discrete wavelet transforms. After that hesitant fuzzy linguistic Bi-objective clustering (HFLBC) is applied for cluster selection based on the extracted features. Subsequently, a movie recommender scheme utilizing a convolutional spiking neural network is introduced to predict user film preferences. The efficiency of the proposed model is compared to existing methods such as multi-modal trust-dependent recommender scheme and graph-dependent hybrid recommendation scheme. The results show a significant improvement, with the proposed model achieving 13.79% and 16.47% higher accuracy than the existing methods. The findings highlight the potential of proposed system in enhancing the accuracy and personalization of movie recommendations.

VotePLMs-AFP: Identification of antifreeze proteins using transformer-embedding features and ensemble learning

Antifreeze proteins (AFPs) are a unique class of biomolecules capable of protecting other proteins, cell membranes, and cellular structures within organisms from damage caused by freezing conditions. Given the significance of AFPs in various domains such as biotechnology, agriculture, and medicine, several machine learning methods have been developed to identify AFPs. However, due to the complexity and diversity of AFPs, the predictive performance of existing methods is limited. Therefore, there is an urgent need to develop an efficient and rapid computational method for accurately predicting AFPs. In this study, we proposed a novel predictor based on transformer-embedding features and ensemble learning for the identification of AFPs, termed VotePLMs-AFP. Firstly, three types of feature descriptors were extracted from pre-trained protein language models (PLMs) during the feature extraction process. Subsequently, we analyzed six combinations generated by these three embeddings to explore the optimal feature set, which was input into the soft voting-based ensemble learning classifier for the identification of AFPs. Finally, we evaluated the model on the two benchmark datasets. The experimental results show that our model achieves high prediction accuracy in 10-fold cross-validation (CV) and independent set testing, outperforming existing state-of-the-art methods. Therefore, our model could serve as an effective tool for predicting AFPs.

A novel deep anomaly detection approach for intrusion detection in futurisitic network

<span lang="EN-US">In an era where networks are increasingly heterogeneous and multi-domain, establishing robust security models to protect data and network infrastructure is becoming ever more complex. Traditional intrusion detection systems (IDS) often struggle with novel or variant attacks that fall outside predefined rule sets, resulting in significant detection challenges. This paper proposes a methodologically refined approach leveraging data-driven insights and statistically robust feature selection to enhance the training dataset. The study presents a long short-term memory-autoencoder (LSTM-AE) based learning model designed for multi-class anomaly detection. The model's novelty lies in its application of distance metrics to define distinct thresholds for varied attack classifications, a strategy that significantly amplifies detection precision. Experimental results validate the superior performance of the proposed system, achieving 94.82% accuracy rate, outperforming similar existing works. The study also proactively addresses common issues of class imbalance and skewed data representation in benchmark datasets by strategically training the model on normal traffic, enhancing its capability to generalize and identify anomalies effectively.</span>

No-reference image quality assessment based on visual explanation images and deep transfer learning

Quantifying image quality in the absence of a reference image continues to be a challenge despite the introduction of numerous no-reference image quality assessments (NR-IQA) in recent years. Unlike most existing NRIQA methods, this paper proposes an efficient NR-IQA method based on deep visual interpretations. Specifically, the main components of the proposed method are: i) generating a pseudo-reference image (PRI) for the input distorted images, ii) employing a pretrained convolutional network to extract feature maps from the distorted image and the corresponding PRI, iii) producing visual explanation images (VEIs) by using the feature maps of the distorted image and the corresponding PRI, iv) measuring the similarity between the two VEIs using an image similarity metric, and v) employing a non-linear mapping function for quality score alignment. In our experiments, we evaluated the efficacy of the proposed method across various forms of distortion using four benchmark datasets (LIVE, SIQAD, CSIQ, and TID2013). The proposed approach demonstrates parity with the latest methods, as evidenced by comparisons with both hand-crafted NR-IQA and deep learning-based approaches.

Preference learning based on adaptive graph neural network for multi-criteria decision support

Intelligent decision-making assists decision-makers (DMs) in making choices through data analysis, model prediction, and automated processes. Central to this field are two key concepts: multi-criteria decision making (MCDM) and preference learning (PL). While MCDM and PL both aim to develop decision models that rank alternatives based on observed or revealed preferences, they diverge in focus. MCDM concentrates on the DMs' perspectives, whereas PL emphasizes model-driven approaches. This divergence presents significant challenges in integrating these methodologies, particularly in ensuring the integrated method remains scalable and interpretable amidst the complexity of decision scenarios. To bridge this gap, our study introduces the use of graph structures to frame decision problems and proposes a novel PL method employing graph neural network (GNN) for multi-criteria decision support. This method is anchored in the Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS) technique and combines an adaptive GNN model with a weight determination model. The GNN model updates embeddings from the alternative's criteria and category features, utilizing an attention mechanism to adaptively assess their importance. Concurrently, the weight determination model contains a weight neural network module to set objective criterion weights and a game theory-based module for calculating combined criterion weights. The method not only inherits the interpretability and intuitive appeal of decision models but also leverages the computational efficiency and high accuracy of machine learning. In experiments conducted on benchmark datasets, our method exhibits significant performance improvements, especially in ranking-related evaluation metrics, outperforming the best baseline by 5.78%.

OOD-CV-v2 : An Extended Benchmark for Robustness to Out-of-Distribution Shifts of Individual Nuisances in Natural Images.

Enhancing the robustness of vision algorithms in real-world scenarios is challenging. One reason is that existing robustness benchmarks are limited, as they either rely on synthetic data or ignore the effects of individual nuisance factors. We introduce OOD-CV-v2, a benchmark dataset that includes out-of-distribution examples of 10 object categories in terms of pose, shape, texture, context and the weather conditions, and enables benchmarking of models for image classification, object detection, and 3D pose estimation. In addition to this novel dataset, we contribute extensive experiments using popular baseline methods, which reveal that: 1) Some nuisance factors have a much stronger negative effect on the performance compared to others, also depending on the vision task. 2) Current approaches to enhance robustness have only marginal effects, and can even reduce robustness. 3) We do not observe significant differences between convolutional and transformer architectures. We believe our dataset provides a rich test bed to study robustness and will help push forward research in this area.

Searching to Exploit Memorization Effect in Deep Learning With Noisy Labels.

Sample selection approaches are popular in robust learning from noisy labels. However, how to control the selection process properly so that deep networks can benefit from the memorization effect is a hard problem. In this paper, motivated by the success of automated machine learning (AutoML), we propose to control the selection process by bi-level optimization. Specifically, we parameterize the selection process by exploiting the general patterns of the memorization effect in the upper-level, and then update these parameters using predicting accuracy obtained from model training in the lower-level. We further introduce semi-supervised learning algorithms to utiilize noisy-labeled data as unlabeled data. To solve the bi-level optimization problem efficiently, we consider more information from the validation curvature by the Newton method and cubic regularization method. We provide convergence analysis for both optimization methods. Results show that while both methods can converge to an (approximately) stationary point, the cubic regularization method can find better local optimal than the Newton method with less time. Experiments on both benchmark and real-world data sets demonstrate that the proposed searching method can lead to significant improvements upon existing methods. Compared with existing AutoML approaches, our method is much more efficient on finding a good selection schedule.

A preordonance-based decision tree method and its parallel implementation in the framework of Map-Reduce

In supervised classification, decision trees are one of the most popular learning algorithms that are employed in many practical applications because of their simplicity, adaptability, and other perks. The development of effective and efficient decision trees remains a major focus in machine learning. Therefore, the scientific literature provides various node splitting measurements that can be utilized to produce different decision trees, including Information Gain, Gain Ratio, Average Gain, and Gini Index. This research paper presents a new node splitting metric that is based on preordonance theory. The primary benefit of the new split criterion is its ability to deal with categorical or numerical attributes directly without discretization. Consequently, the Preordonance-based decision tree” (P-Tree) approach, a powerful technique that generates decision trees using the suggested node splitting measure, is developed. Both multiclass classification problems and imbalanced data sets can be handled by the P-Tree decision tree strategy. Moreover, the over-partitioning problem is addressed by the P-Tree methodology, which introduces a threshold ϵ as a stopping condition. If the percentage of instances in a node falls below the predetermined threshold, the expansion of the tree will be halted. The performance of the P-Tree procedure is evaluated on fourteen benchmark data sets with different sizes and contrasted with that of five already existing decision tree methods using a variety of evaluation metrics. The results of the experiments demonstrate that the P-Tree model performs admirably across all of the tested data sets and that it is comparable to the other five decision tree algorithms overall. On the other hand, an ensemble technique called “ensemble P-Tree” offers a reliable remedy to mitigate the instability that is frequently associated with tree-based algorithms. This ensemble method leverages the strengths of the P-Tree approach to enhance predictive performance through collective decision-making. The ensemble P-Tree strategy is comprehensively evaluated by comparing its performance to that of two top-performing ensemble decision tree methodologies. The experimental findings highlight its exceptional performance and competitiveness against other decision tree procedures. Despite the excellent performance of the P-Tree approach, there are still some obstacles that prevent it from handling larger data sets, such as memory restrictions, time complexity, or data complexity. However, parallel computing is effective in resolving this kind of problem. Hence, the MR-P-Tree decision tree technique, a parallel implementation of the P-Tree strategy in the Map-Reduce framework, is further designed. The three parallel procedures MR-SA-S, MR-SP-S, and MR-S-DS for choosing the optimal splitting attributes, choosing the optimal splitting points, and dividing the training data set in parallel, respectively, are the primary basis of the MR-P-Tree methodology. Furthermore, several experimental studies are carried out on ten additional data sets to illustrate the viability of the MR-P-Tree technique and its strong parallel performance.

Multi-view contrastive clustering for cancer subtyping using fully and weakly paired multi-omics data

The identification of cancer subtypes is crucial for advancing precision medicine, as it facilitates the development of more effective and personalized treatment and prevention strategies. With the development of high-throughput sequencing technologies, researchers now have access to a wealth of multi-omics data from cancer patients, making computational cancer subtyping increasingly feasible. One of the main challenges in integrating multi-omics data is handling missing data, since not all biomolecules are consistently measured across all samples. Current computational models based on multi-omics data for cancer subtyping often struggle with the challenge of weakly paired omics data. To address this challenge, we propose a novel unsupervised cancer subtyping model named Subtype-MVCC. This model leverages graph convolutional networks to extract and represent low-dimensional features from each omics data type, using intra-view and inter-view contrastive learning approaches. By incorporating a weighted average fusion strategy to unify the dimension of each sample, Subtype-MVCC effectively handles weakly paired multi-omics datasets. Comprehensive evaluations on established benchmark datasets demonstrate that Subtype-MVCC outperforms nine leading models in this domain. Additionally, simulations with varying levels of missing data highlight the model's robust performance in handling weakly paired omics data. The clinical relevance and survival outcomes associated with the identified subtypes further validate the interpretability and reliability of the clustering results produced by Subtype-MVCC.

SSR-2D: Semantic 3D Scene Reconstruction From 2D Images.

Most deep learning approaches to comprehensive semantic modeling of 3D indoor spaces require costly dense annotations in the 3D domain. In this work, we explore a central 3D scene modeling task, namely, semantic scene reconstruction without using any 3D annotations. The key idea of our approach is to design a trainable model that employs both incomplete 3D reconstructions and their corresponding source RGB-D images, fusing cross-domain features into volumetric embeddings to predict complete 3D geometry, color, and semantics with only 2D labeling which can be either manual or machine-generated. Our key technical innovation is to leverage differentiable rendering of color and semantics to bridge 2D observations and unknown 3D space, using the observed RGB images and 2D semantics as supervision, respectively. We additionally develop a learning pipeline and corresponding method to enable learning from imperfect predicted 2D labels, which could be additionally acquired by synthesizing in an augmented set of virtual training views complementing the original real captures, enabling more efficient self-supervision loop for semantics. As a result, our end-to-end trainable solution jointly addresses geometry completion, colorization, and semantic mapping from limited RGB-D images, without relying on any 3D ground-truth information. Our method achieves state-of-the-art performance of semantic scene completion on two large-scale benchmark datasets MatterPort3D and ScanNet, surpasses baselines even with costly 3D annotations in predicting both geometry and semantics. To our knowledge, our method is also the first 2D-driven method addressing completion and semantic segmentation of real-world 3D scans simultaneously.

Structured orthogonal random features based on DCT for kernel approximation

Random Fourier features are a popular technique used to improve nonlinear kernel methods in large-scale problems. Recent studies have shown that replacing random Gaussian matrices of random feature maps with appropriately scaled random orthogonal matrices, such as SORF, can significantly improve kernel approximation performance. However, the SORF method theoretically requires zero-padding for datasets whose input feature dimension does not meet d=2p for some p∈N+, resulting in increased memory and computation time. To address this limitation, we propose a new structured orthogonal random features method based on discrete cosine transform (SORF-DCT) to approximate Gaussian kernel functions. The SORF-DCT method does not require the input feature dimension to be d=2p. The key to SORF-DCT is that we combine the DCT matrix with diagonal ”sign-flipping” matrices and scale them appropriately using a diagonal matrix whose diagonal elements follow the chi distribution, resulting in a structured orthogonal matrix whose elements follow the standard Gaussian distribution instead of a random Gaussian matrix. SORF-DCT generates D-dimensional structured orthogonal random features in O(Dlogd) time and O(D) memory by employing fast discrete cosine transform, where d and D represent the input feature dimension and expansion dimension, respectively. We prove that SORF-DCT is an unbiased estimator of the Gaussian kernel function and analyze the concentration error bounds. Experimental results on eleven benchmark datasets show that SORF-DCT achieves lower kernel approximation errors, requires less time, and has comparable nonlinear learning capabilities.

Granular intents learning via mutual information maximization for knowledge-aware recommendation

Knowledge-aware recommender systems, which utilize knowledge graphs (KGs) to enrich item information, have been shown to improve the accuracy and explainability of recommendations. Besides, KGs are further explored to determine the intent of choosing items (i.e., the reason why users select items of interest). Conventional methods represent intents either as sets of relations in a KG or as KG entities. However, such approaches fail to fully leverage the combined information provided by both entities and relations. To address this issue, we propose a new KG-based user Intent Extraction Framework (KIEF) to capture user intents at a more fine-grained level for recommendation. Specifically, we propose a novel intent representation constructed with relation-aware entity representation, encouraging finer granularity for user intents. Furthermore, since a KG may contain noisy information that impairs the quality of user intent, it is compulsory to consider which factors in a KG are important to represent a user’s intent. Thus, we introduce global intent which are comprehensive features for the entire interactions of all users and local intent, which are empirical features of individual users from personal history. By maximizing mutual information between global and local intents, KIEF captures user preference for items. Through extensive experiments on four real-world benchmark datasets, we prove the superior performance of KIEF over the state-of-the-art and analyze interpretable explanations for understanding user intents.

Label Deconvolution for Node Representation Learning on Large-Scale Attributed Graphs Against Learning Bias.

Node representation learning on attributed graphs-whose nodes are associated with rich attributes (e.g., texts and protein sequences)-plays a crucial role in many important downstream tasks. To encode the attributes and graph structures simultaneously, recent studies integrate pre-trained models with graph neural networks (GNNs), where pre-trained models serve as node encoders (NEs) to encode the attributes. As jointly training large NEs and GNNs on large-scale graphs suffers from severe scalability issues, many methods propose to train NEs and GNNs separately. Consequently, they do not take feature convolutions in GNNs into consideration in the training phase of NEs, leading to a significant learning bias relative to the joint training. To address this challenge, we propose an efficient label regularization technique, namely Label Deconvolution (LD), to alleviate the learning bias by a novel and highly scalable approximation to the inverse mapping of GNNs. The inverse mapping leads to an objective function that is equivalent to that by the joint training, while it can effectively incorporate GNNs in the training phase of NEs against the learning bias. More importantly, we show that LD converges to the optimal objective function values by the joint training under mild assumptions. Experiments demonstrate LD significantly outperforms state-of-the-art methods on Open Graph Benchmark datasets.

Attention enhanced machine instinctive vision with human-inspired saliency detection

Salient object detection (SOD) enables machines to recognize and accurately segment visually prominent regions in images. Despite recent advancements, existing approaches often lack progressive fusion of low and high-level features, effective multi-scale feature handling, and precise boundary detection. Moreover, the robustness of these models under varied lighting conditions remains a concern. To overcome these challenges, we present Attention Enhanced Machine Instinctive Vision framework for SOD. The proposed framework leverages the strategy of Multi-stage Feature Refinement with Optimal Attentions-Driven Framework (MFRNet). The multi-level features are extracted from six stages of the EfficientNet-B7 backbone. This provides effective feature fusions of low and high-level details across various scales at the later stage of the framework. We introduce the Spatial-optimized Feature Attention (SOFA) module, which refines spatial features from three initial-stage feature maps. The extracted multi-scale features from the backbone are passed from the convolution feature transformation and spatial attention mechanisms to refine the low-level information. The SOFA module concatenates and upsamples these refined features, producing a comprehensive spatial representation of various levels. Moreover, the proposed Context-Aware Channel Refinement (CACR) module integrates dilated convolutions with optimized dilation rates followed by channel attention to capture multi-scale contextual information from the mature three layers. Furthermore, our progressive feature fusion strategy combines high-level semantic information and low-level spatial details through multiple residual connections, ensuring robust feature representation and effective gradient backpropagation. To enhance robustness, we train our network with augmented data featuring low and high brightness adjustments, improving its ability to handle diverse lighting conditions. Extensive experiments on four benchmark datasets—ECSSD, HKU-IS, DUTS, and PASCAL-S- validate the proposed framework’s effectiveness, demonstrating superior performance compared to existing SOTA methods in the domain. Code, qualitative results, and trained weights will be available to the research community at the link: https://github.com/habib1402/MFRNet-SOD.

Detection and identification of un-uniformed shape text from blurred video frames

<p><span lang="EN-GB">The identification and recognition of text from video frames have received a lot of attention recently, that makes many computer vision-based applications conceivable. In this study, we modify the picture mask and the original identification of the mask region convolution neural network and permit detection in three levels, including holistic, sequence, and at the level of pixels. To identify the texts and determine the text forms, semantics at the pixel and holistic levels can be used. With masking and detection, existences of the character and the word are separated and recognised. In addition, text detection using the results of 2-D feature space instance segmentation is done. Moreover, we explore text recognition using an attention-based optical character recognition (OCR) method with mask</span><span lang="EN-US"> r</span><span lang="EN-GB">egion convolution neural networks (R-CNN) to address and detect the problem of smaller and blurrier texts at the sequential level. Using attribute maps of the word occurrences in sequence to seq, the OCR method calculates the character sequence. At last, a fine-grained learning strategy is proposed to constructs models at word level using the annotated datasets, resulting in the training of a more precise and reliable model. The well-known benchmark datasets ICDAR 2013 and ICDAR 2015 are used to test our suggested methodology.</span></p>

Deep one-class classification model assisted by radius constraint for anomaly detection of industrial control systems

Anomaly detection of industrial control systems (ICS) based on sensor data analytic is of utmost importance because ICS may suffer from various attacks leading to anomaly behaviors and even equipment failures. As an emerging deep learning technique, deep support vector data description (DeSVDD) has been successfully applied to ICS anomaly detection. Its advantage lies in only requiring normal data to train one-class classifier, which is suitable for actual industrial scenarios with extremely data imbalance of abundant normal data and scare abnormal data. However, this also results in its shortcoming of ignoring the valuable classification information hidden in the abnormal data. In order to overcome this issue, this paper proposes an improved DeSVDD method, called radius constraint DeSVDD (RC-DeSVDD). The proposed model constructs an anomaly detection model framework of abnormal data assisted DeSVDD, where a radius constraint is designed by considering the difference between normal and abnormal data. Further, considering the limited amount of abnormal data, a bi-directional generative adversarial network (BiGAN) is introduced to generate abnormal data. Experimental results on three benchmark datasets demonstrate the superiority of the proposed RC-DeSVDD anomaly detection method.

A class-incremental learning approach for learning feature-compatible embeddings

Humans have the ability to constantly learn new knowledge. However, for artificial intelligence, trying to continuously learn new knowledge usually results in catastrophic forgetting, the existing regularization-based and dynamic structure-based approaches have shown great potential for alleviating. Nevertheless, these approaches have certain limitations. They usually do not fully consider the problem of incompatible feature embeddings. Instead, they tend to focus only on the features of new or previous classes and fail to comprehensively consider the entire model. Therefore, we propose a two-stage learning paradigm to solve feature embedding incompatibility problems. Specifically, we retain the previous model and freeze all its parameters in the first stage while dynamically expanding a new module to alleviate feature embedding incompatibility questions. In the second stage, a fusion knowledge distillation approach is used to compress the redundant feature dimensions. Moreover, we propose weight pruning and consolidation approaches to improve the efficiency of the model. Our experimental results obtained on the CIFAR-100, ImageNet-100 and ImageNet-1000 benchmark datasets show that the proposed approaches achieve the best performance among all the compared approaches. For example, on the ImageNet-100 dataset, the maximal accuracy improvement is 5.08%. Code is available at https://github.com/ybyangjing/CIL-FCE.

Fine-grained semantic oriented embedding set alignment for text-based person search

Text-based person search aims to retrieve images of a person that are highly semantically relevant to a given textual description. The difficulty of this retrieval task is modality heterogeneity and fine-grained matching. Most existing methods only consider the alignment using global features, ignoring the fine-grained matching problem. The cross-modal attention interactions are popularly used for image patches and text markers for direct alignment. However, cross-modal attention may cause a huge overhead in the reasoning stage and cannot be applied in actual scenarios. In addition, it is unreasonable to perform patch-token alignment, since image patches and text tokens do not have complete semantic information. This paper proposes an Embedding Set Alignment (ESA) module for fine-grained alignment. The module can preserve fine-grained semantic information by merging token-level features into embedding sets. The ESA module benefits from pre-trained cross-modal large models, and it can be combined with the backbone non-intrusively and trained in an end-to-end manner. In addition, an Adaptive Semantic Margin (ASM) loss is designed to describe the alignment of embedding sets, instead of adapting a loss function with a fixed margin. Extensive experiments demonstrate that our proposed fine-grained semantic embedding set alignment method achieves state-of-the-art performance on three popular benchmark datasets, surpassing the previous best methods.

Tensorized and Compressed Multi-View Subspace Clustering via Structured Constraint.

Multi-view learning has raised more and more attention in recent years. However, traditional approaches only focus on the difference while ignoring the consistency among views. It may make some views, with the situation of data abnormality or noise, ineffective in the progress of view learning. Besides, the current datasets have become high-dimensional and large-scale gradually. Therefore, this paper proposes a novel multi-view compressed subspace learning method via low-rank tensor constraint, which incorporates the clustering progress and multi-view learning into a unified framework. First, for each view, we take the partial samples to build a small-size dictionary, which can reduce the effect of both redundancy information and computation cost greatly. Then, to find the consistency and difference among views, we impose a low-rank tensor constraint on these representations and further design an auto-weighted mechanism to learn the optimal representation. Last, due to the non-square of the learned representation, the bipartite graph has been introduced, and under the structured constraint, the clustering results can be obtained directly from this graph without any post-processing. Extensive experiments on synthetic and real-world benchmark datasets demonstrate the efficacy and efficiency of our method, especially for the views with noise or outliers.