RULEM: A novel heuristic rule learning approach for ordinal classification with monotonicity constraints

In many real world applications classification models are required to be in line with domain knowledge and to respect monotone relations between predictor variables and the target class, in order to be acceptable for implementation. This paper presents a novel heuristic approach, called RULEM, to induce monotone ordinal rule based classification models. The proposed approach can be applied in combination with any rule- or tree-based classification technique, since monotonicity is guaranteed in a post-processing step. RULEM checks whether a rule set or decision tree violates the imposed monotonicity constraints and existing violations are resolved by inducing a set of additional rules which enforce monotone classification. The approach is able to handle non-monotonic noise, and can be applied to both partially and totally monotone problems with an ordinal target variable. Two novel justifiability measures are introduced which are based on RULEM and allow to calculate the extent to which a classification model is in line with domain knowledge expressed in the form of monotonicity constraints. An extensive benchmarking experiment and subsequent statistical analysis of the results on 14 public data sets indicates that RULEM preserves the predictive power of a rule induction technique while guaranteeing monotone classification. On the other hand, the post-processed rule sets are found to be significantly larger which is due to the induction of additional rules. E.g., when combined with Ripper a median performance difference was observed in terms of PCC equal to zero and an average difference equal to −0.66%, with on average 5 rules added to the rule sets. The average and minimum justifiability of the original rule sets equal respectively 92.66% and 34.44% in terms of the RULEMF justifiability index, and 91.28% and 40.1% in terms of RULEMS, indicating the effective need for monotonizing the rule sets.     

Multi-objective rule mining using a chaotic particle swarm optimization algorithm

In this paper, classification rule mining which is one of the most studied tasks in data mining community has been modeled as a multi-objective optimization problem with predictive accuracy and comprehensibility objectives. A multi-objective chaotic particle swarm optimization (PSO) method has been introduced as a search strategy to mine classification rules within datasets. The used extension to PSO uses similarity measure for neighborhood and far-neighborhood search to store the global best particles found in multi-objective manner. For the bi-objective problem of rule mining of high accuracy/comprehensibility, the multi-objective approach is intended to allow the PSO algorithm to return an approximation to the upper accuracy/comprehensibility border, containing solutions that are spread across the border. The experimental results show the efficiency of the algorithm.     

对抗样本三元组约束的度量学习算法

针对已有三元组约束的度量学习算法大多利用先验知识构建约束,一定程度上制约了度量学习算法性能的问题,本文借鉴对抗训练中样本扰动的思想,在原始样本附近学习对抗样本以构造对抗三元组约束,基于对抗三元组和原始三元组约束构建了度量学习模型,提出了对抗样本三元组约束的度量学习算法(metric learning algorithm with adversarial sample triples constraints,ASTCML)。实验结果表明,提出的算法既克服了已有固定约束方法受先验知识影响大的问题,也提高了分类精度,说明区分更加难以区分的三元组约束能够提升算法的性能。     

Compositional metric learning for multi-label classification

Multi-label classification aims to assign a set of proper labels for each instance, where distance metric learning can help improve the generalization ability of instance-based multi-label classification models. Existing multi-label metric learning techniques work by utilizing pairwise constraints to enforce that examples with similar label assignments should have close distance in the embedded feature space. In this paper, a novel distance metric learning approach for multi-label classification is proposed by modeling structural interactions between instance space and label space. On one hand, compositional distance metric is employed which adopts the representation of a weighted sum of rank-1 PSD matrices based on component bases. On the other hand, compositional weights are optimized by exploiting triplet similarity constraints derived from both instance and label spaces. Due to the compositional nature of employed distance metric, the resulting problem admits quadratic programming formulation with linear optimization complexity w.r.t. the number of training examples.We also derive the generalization bound for the proposed approach based on algorithmic robustness analysis of the compositional metric. Extensive experiments on sixteen benchmark data sets clearly validate the usefulness of compositional metric in yielding effective distance metric for multi-label classification.     

基于度量学习的无监督域适应方法及其在死亡风险预测上的应用

近年来,深度学习模型已在医疗领域的预测任务上得到广泛应用,并取得了不错的效果.然而,深度学习模型常会面临带标签训练数据不足、整体数据分布偏移和类别之间数据分布偏移的问题,导致模型预测的准确度下降.为解决上述问题,提出一种基于域对抗和加性余弦间隔损失的无监督域适应方法(additive margin softmax ba...     

A comparative analysis of classification algorithms in data mining for accuracy, speed and robustness

Classification algorithms are the most commonly used data mining models that are widely used to extract valuable knowledge from huge amounts of data. The criteria used to evaluate the classifiers are mostly accuracy, computational complexity, robustness, scalability, integration, comprehensibility, stability, and interestingness. This study compares the classification of algorithm accuracies, speed (CPU time consumed) and robustness for various datasets and their implementation techniques. The data miner selects the model mainly with respect to classification accuracy; therefore, the performance of each classifier plays a crucial role for selection. Complexity is mostly dominated by the time required for classification. In terms of complexity, the CPU time consumed by each classifier is implied here. The study first discusses the application of certain classification models on multiple datasets in three stages: first, implementing the algorithms on original datasets; second, implementing the algorithms on the same datasets where continuous variables are discretised; and third, implementing the algorithms on the same datasets where principal component analysis is applied. The accuracies and the speed of the results are then compared. The relationship of dataset characteristics and implementation attributes between accuracy and CPU time is also examined and debated. Moreover, a regression model is introduced to show the correlating effect of dataset and implementation conditions on the classifier accuracy and CPU time. Finally, the study addresses the robustness of the classifiers, measured by repetitive experiments on both noisy and cleaned datasets.     

基于最大相关熵准则的鲁棒度量学习算法

度量亦称距离函数,是度量空间中满足特定条件的特殊函数,一般用来反映数据间存在的一些重要距离关系.而距离对于各种分类聚类问题影响很大,因此度量学习对于这类机器学习问题有重要影响.受到现实存在的各种噪声影响,已有的各种度量学习算法在处理各种分类问题时,往往出现分类准确率较低以及分类准确率波动大的问题.针对该问题,本文提出一种基于最大相关熵准则的鲁棒度量学习算法.最大相关熵准则的核心在于高斯核函数,本文将其引入到度量学习中,通过构建以高斯核函数为核心的损失函数,利用梯度下降法进行优化,反复测试调整参数,最后得到输出的度量矩阵.通过这样的方法学习到的度量矩阵将有更好的鲁棒性,在处理受噪声影响的各种分类问题时,将有效地提高分类准确率.本文将在一些常用机器学习数据集（UCI）还有人脸数据集上进行验证实验.     

Do software models based on the UML aid in source-code comprehensibility? Aggregating evidence from 12 controlled experiments

In this paper, we present the results of long-term research conducted in order to study the contribution made by software models based on the Unified Modeling Language (UML) to the comprehensibility of Java source-code deprived of comments. We have conducted 12 controlled experiments in different experimental contexts and on different sites with participants with different levels of expertise (i.e., Bachelor’s, Master’s, and PhD students and software practitioners from Italy and Spain). A total of 333 observations were obtained from these experiments. The UML models in our experiments were those produced in the analysis and design phases. The models produced in the analysis phase were created with the objective of abstracting the environment in which the software will work (i.e., the problem domain), while those produced in the design phase were created with the goal of abstracting implementation aspects of the software (i.e., the solution/application domain). Source-code comprehensibility was assessed with regard to correctness of understanding, time taken to accomplish the comprehension tasks, and efficiency as regards accomplishing those tasks. In order to study the global effect of UML models on source-code comprehensibility, we aggregated results from the individual experiments using a meta-analysis. We made every effort to account for the heterogeneity of our experiments when aggregating the results obtained from them. The overall results suggest that the use of UML models affects the comprehensibility of source-code, when it is deprived of comments. Indeed, models produced in the analysis phase might reduce source-code comprehensibility, while increasing the time taken to complete comprehension tasks. That is, browsing source code and this kind of models together negatively impacts on the time taken to complete comprehension tasks without having a positive effect on the comprehensibility of source code. One plausible justification for this is that the UML models produced in the analysis phase focus on the problem domain. That is, models produced in the analysis phase say nothing about source code and there should be no expectation that they would, in any way, be beneficial to comprehensibility. On the other hand, UML models produced in the design phase improve source-code comprehensibility. One possible justification for this result is that models produced in the design phase are more focused on implementation details. Therefore, although the participants had more material to read and browse, this additional effort was paid back in the form of an improved comprehension of source code.     

Mining software repositories for comprehensible software fault prediction models

Software managers are routinely confronted with software projects that contain errors or inconsistencies and exceed budget and time limits. By mining software repositories with comprehensible data mining techniques, predictive models can be induced that offer software managers the insights they need to tackle these quality and budgeting problems in an efficient way. This paper deals with the role that the Ant Colony Optimization (ACO)-based classification technique AntMiner+ can play as a comprehensible data mining technique to predict erroneous software modules. In an empirical comparison on three real-world public datasets, the rule-based models produced by AntMiner+ are shown to achieve a predictive accuracy that is competitive to that of the models induced by several other included classification techniques, such as C4.5, logistic regression and support vector machines. In addition, we will argue that the intuitiveness and comprehensibility of the AntMiner+ models can be considered superior to the latter models.     

Metric learning for weather image classification

Image classification is a core task in many applications of computer vision. Recognition of weather conditions based on large-volume image datasets is a challenging problem. However, there has been little research on weather-related recognition using color images, particularly with large datasets. In this study, we proposed a metric learning framework to investigate a two-class weather classification problem. We improve the classification accuracy using metric learning approaches. Extracting features from images is a challenging task and practical requirements such as domain knowledge and human intervention. In this paper, we define several categories of weather feature cures based on observations of outdoor images captured under different weather conditions. Experimental results show that a classifier based on metric learning framework is effective in weather classification and outperforms the previous approach when using the same dataset.     

Asymmetric and Category Invariant Feature Transformations for Domain Adaptation

-1We address the problem of visual domain adaptation for transferring object models from one dataset or visual domain to another. We introduce a unified flexible model for both supervised and semi-supervised learning that allows us to learn transformations between domains. Additionally, we present two instantiations of the model, one for general feature adaptation/alignment, and one specifically designed for classification. First, we show how to extend metric learning methods for domain adaptation, allowing for learning metrics independent of the domain shift and the final classifier used. Furthermore, we go beyond classical metric learning by extending the method to asymmetric, category independent transformations. Our framework can adapt features even when the target domain does not have any labeled examples for some categories, and when the target and source features have different dimensions. Finally, we develop a joint learning framework for adaptive classifiers, which outperforms competing methods in terms of multi-class accuracy and scalability. We demonstrate the ability of our approach to adapt object recognition models under a variety of situations, such as differing imaging conditions, feature types, and codebooks. The experiments show its strong performance compared to previous approaches and its applicability to large-scale scenarios.     

A genetic algorithm-based method for feature subset selection

As a commonly used technique in data preprocessing, feature selection selects a subset of informative attributes or variables to build models describing data. By removing redundant and irrelevant or noise features, feature selection can improve the predictive accuracy and the comprehensibility of the predictors or classifiers. Many feature selection algorithms with different selection criteria has been introduced by researchers. However, it is discovered that no single criterion is best for all applications. In this paper, we propose a framework based on a genetic algorithm (GA) for feature subset selection that combines various existing feature selection methods. The advantages of this approach include the ability to accommodate multiple feature selection criteria and find small subsets of features that perform well for a particular inductive learning algorithm of interest to build the classifier. We conducted experiments using three data sets and three existing feature selection methods. The experimental results demonstrate that our approach is a robust and effective approach to find subsets of features with higher classification accuracy and/or smaller size compared to each individual feature selection algorithm.     

An ensemble method for fuzzy rule-based classification systems

Fuzzy rule-based classification systems are very useful tools in the field of machine learning as they are able to build linguistic comprehensible models. However, these systems suffer from exponential rule explosion when the number of variables increases, degrading, therefore, the accuracy of these systems as well as their interpretability. In this article, we propose to improve the comprehensibility through a supervised learning method by automatic generation of fuzzy classification rules, designated SIFCO–PAF. Our method reduces the complexity by decreasing the number of rules and of antecedent conditions, making it thus adapted to the representation and the prediction of rather high-dimensional pattern classification problems. We perform, firstly, an ensemble methodology by combining a set of simple classification models. Subsequently, each model uses a subset of the initial attributes: In this case, we propose to regroup the attributes using linear correlation search among the training set elements. Secondly, we implement an optimal fuzzy partition thanks to supervised discretization followed by an automatic membership functions construction. The SIFCO–PAF method, analyzed experimentally on various data sets, guarantees an important reduction in the number of rules and of antecedents without deteriorating the classification rates, on the contrary accuracy is even improved.     

关联词约束的半监督文本分类方法

提出了一种没有训练集情况下实现对未标注类别文本文档进行分类的问题。类关联词是与类主体相关、能反映类主体的单词或短语。利用类关联词提供的先验信息,形成文档分类的先验概率,然后组合利用朴素贝叶斯分类器和EM迭代算法,在半监督学习过程中加入分类约束条件,用类关联词来监督构造一个分类器,实现了对完全未标注类别文档的分类。实验结果证明,此方法能够以较高的准确率实现没有训练集情况下的文本分类问题,在类关联词约束下的分类准确率要高于没有约束情况下的分类准确率。     

基于角度特征的分类网络

卷积神经网络(Convolutional Neural Networks,CNN)在图像分类任务中的卓越表现,使得其被广泛应用于计算机视觉的各个领域。图像分类模型精度与效率的提升,除了归功于网络结构的改变外,还有很大一部分原因来自于归一化技术以及分类损失函数的改进。在人脸识别任务中,随着精度的不断提升,分类损失函数从Softmax Loss到Triplet Loss,又从L-Softmax Loss到Arcface Loss,度量方式从几何度量发展到角度度量。度量方式的改变实际上是特征形式的变化,即特征形式从一般特征转变为角度特征。在Mnist数据集上,使用角度度量损失函数训练得到的特征点呈角度分布,同时准确率比几何度量高;将角度度量方式用更直接的角度特征来表示,训练得到的同类特征点呈直线分布,准确度也比一般角度度量更高。这不禁令人思考,在CNN分类模型中是否可以使用角度特征来代替一般特征。在CNN分类模型中,其主要架构往往由多个卷积层和一个或多个全连接层组成,通过统一卷积层与全连接层的归一化操作,得到角度卷积层与角度全连接层。在普通分类网络的基础上,用角度卷积层替换卷积层,用角度全连接层替换全连接层,可以得到一个由角度特征组成的角度分类网络。在Cifar-100数据集上,基于ResNet-32构造的角度分类网络相比原分类网络,分类准确率提高了2%,从而论证了角度特征在分类网络中的有效性。     

Predictive and comprehensible rule discovery using a multi-objective genetic algorithm

We present a multi-objective genetic algorithm for mining highly predictive and comprehensible classification rules from large databases. We emphasize predictive accuracy and comprehensibility of the rules. However, accuracy and comprehensibility of the rules often conflict with each other. This makes it an optimization problem that is very difficult to solve efficiently. We have proposed a multi-objective evolutionary algorithm called improved niched Pareto genetic algorithm (INPGA) for this purpose. We have compared the rule generation by INPGA with that by simple genetic algorithm (SGA) and basic niched Pareto genetic algorithm (NPGA). The experimental result confirms that our rule generation has a clear edge over SGA and NPGA.     

融合注意力机制的恶意代码家族分类研究

近年来,随着恶意代码家族变种的多样化和混淆等对抗手段的不断加强,传统的恶意代码检测方法难以取得较好的分类效果。鉴于此,提出了一种融合注意力机制的恶意代码家族分类模型。首先,使用逆向反汇编工具获取恶意样本的各区段特征,并利用可视化技术将各区段转化为RGB彩色图像的各通道;其次,引入通道域和空间域注意力机制来构建基于混合域注意力机制的深度可分离卷积网络,从通道和空间两个维度提取恶意样本的图像纹理特征;最后,选取九类恶意代码家族对模型进行训练和测试。实验结果表明,使用单一区段特征对恶意代码家族分类的准确率较低,采用融合特征能够有效地区分各类恶意代码家族,同时该模型相比于传统的神经网络模型取得了更好的分类效果,模型的分类准确率达到了98.38%。     

基于加权分类损失和核范数的领域自适应模型

领域自适应将源域上学习到的知识迁移到目标域上,使得在带标签数据少的情况下也可以有效地训练模型。采用伪标签的领域自适应模型未考虑错误伪标签的影响,并且在决策边界处样本的分类准确率较低,针对上述问题提出了基于加权分类损失和核范数的领域自适应模型。该模型使用带有伪标签的可信样本特征与带有真实标签的源域样本特征构建辅助域,在辅助域上设计加权分类损失函数,降低错误伪标签在训练过程中产生的影响;加入批量核范数最大化损失,提高决策边界处样本的分类准确率。在Office31、Office-Home、Image-CLEFDA基准数据集上与之前模型的对比实验表明,该模型有更高的精确度。     

The Role of Occam's Razor in Knowledge Discovery

Many KDD systems incorporate an implicit or explicit preference for simpler models, but this use of Occam's razor has been strongly criticized by several authors (e.g., Schaffer, 1993; Webb, 1996). This controversy arises partly because Occam's razor has been interpreted in two quite different ways. The first interpretation (simplicity is a goal in itself) is essentially correct, but is at heart a preference for more comprehensible models. The second interpretation (simplicity leads to greater accuracy) is much more problematic. A critical review of the theoretical arguments for and against it shows that it is unfounded as a universal principle, and demonstrably false. A review of empirical evidence shows that it also fails as a practical heuristic. This article argues that its continued use in KDD risks causing significant opportunities to be missed, and should therefore be restricted to the comparatively few applications where it is appropriate. The article proposes and reviews the use of domain constraints as an alternative for avoiding overfitting, and examines possible methods for handling the accuracy–comprehensibility trade-off.     

Improving learning accuracy of fuzzy decision trees by hybridneural networks

Although the induction of fuzzy decision tree (FDT) has been a very popular learning methodology due to its advantage of comprehensibility, it is often criticized to result in poor learning accuracy. Thus, one fundamental problem is how to improve the learning accuracy while the comprehensibility is kept. This paper focuses on this problem and proposes using a hybrid neural network (HNN) to refine the FDT. This HNN, designed according to the generated FDT and trained by an algorithm derived in this paper, results in a FDT with parameters, called weighted FDT. The weighted FDT is equivalent to a set of fuzzy production rules with local weights (LW) and global weights (GW) introduced in our previous work (1998). Moreover, the weighted FDT, in which the reasoning mechanism incorporates the trained LW and GW, significantly improves the FDTs' learning accuracy while keeping the FDT comprehensibility. The improvements are verified on several selected databases. Furthermore, a brief comparison of our method with two benchmark learning algorithms, namely, fuzzy ID3 and traditional backpropagation, is made. The synergy between FDT induction and HNN training offers new insight into the construction of hybrid intelligent systems with higher learning accuracy