基于差异性评估对Co-training文本分类算法的改进

Co-training算法要求两个特征视图满足一致性和独立性假设,但是,许多实际应用中不存自然的划分且满足这种假设的两个视图,且直接评估两个视图的独立性有一定的难度.分析Co-training的理论假设,本文把寻找两个满足一致性和独立性特征视图的目标,转变成寻找两个既满足一定的正确性,又存在较大的差异性的两个基分类器的问题.首先利用特征评估函数建立多个特征视图,每个特征视图包含足够的信息训练生成一个基分类器,然后通过评估基分类器之间的差异性间接评估二者的独立性,选择两个满足一定的正确性和差异性比较大的基分类器协同训练.根据每个视图上采用的分类算法是否相同,提出了两种改进算法TV-SC和TV-DC.实验表明改进的TV-SC和TV-DC算法明显优于基于随机分割特征视图的Co-Rnd算法,而且TV-DC算法的分类效果要优于TV-SC算法.

基于图的Co-Training网页分类

本文充分利用网页数据的超链接关系和文本信息,提出了一种用于网页分类的归纳式半监督学习算法:基于图的Co-training网页分类算法(Graph based Co-training algorithmfor web page classification),简称GCo-training,并从理论上证明了算法的有效性.GCo-training在Co-training算法框架下,迭代地学习一个基于由超链接信息构造的图的半监督分类器和一个基于文本特征的Bayes分类器.基于图的半监督分类器只利用少量的标记数据,通过挖掘数据间大量的关系信息就可达到比较高的预测精度,可为Bayes分类器提供大量的标记信息;反过来学习大量标记信息后的Bayes分类器也可为基于图的分类器提供有效信息.迭代过程中,二者互相帮助,不断提高各自的性能,而后Bayes分类器可以用来预测大量未见数据的类别.在Web→KB数据集上的实验结果表明,与利用文本特征和锚文本特征的Co-training算法和基于EM的Bayes算法相比,GCo-training算法性能优越.

基于Co-training的用户属性预测研究

针对当前基于第三方应用数据进行用户属性预测算法研究,其较少考虑应用前台实际使用时长问题,由此,本文在应用的使用频率及使用时长的基础上,构造了应用前台均使用时长特征,该特征能进一步刻画用户对应用的兴趣度;同时,为充分利用大量未标注数据,从多角度特征对用户属性进行预测,由此本文采用了Co-training框架,该框架包含两个均由栈式自编码器与神经网络相结合的网络结构。实验过程中,对于栈式自编码算法,先利用未标注的数据对网络进行参数初始化,使得网络参数处于一个较优的位置,再利用有标注的数据,采用基于准确率的梯度下降算法,对网络参数进行更新,最终达到收敛。实验结果表明,本文算法在准确率、召回率、F1值上均有所提高。

半监督学习的Co-training算法研究

介绍一种基于半监督学习的协同训练(Co-training)分类算法,当可用的训练样本比较少时,使用传统的方法进行分类,如决策树分类,将无法得到用户满意的结果,而且它们需要大量的标记样本。事实上,获取有标签的样本的代价是相当昂贵的。于是,使用较少的已标记样本和大量的无标记样本进行协同训练的半监督学习,成为研究者首选。

Co-training机器学习方法在中文组块识别中的应用

采用半指导机器学习方法co training实现中文组块识别。首先明确了中文组块的定义,co training算法的形式化定义。文中提出了基于一致性的co training选取方法将增益的隐马尔可夫模型(TransductiveHMM)和基于转换规则的分类器(fnTBL)组合成一个分类体系,并与自我训练方法进行了比较,在小规模汉语树库语料和大规模未带标汉语语料上进行中文组块识别,实验结果要比单纯使用小规模的树库语料有所提高,F值分别达到了85 34%和83 4 1% ,分别提高了2 13%和7 2 1%。

基于样本条件价值改进的Co-training算法

Co-training是一种主流的半监督学习算法.该算法中两视图下的分类器通过迭代的方式,互为对方从无标记样本集中挑选新增样本,以更新对方训练集.Co-training以分类器的后验概率输出作为新增样本的挑选策略,该策略忽略了样本对于当前分类器的价值.针对该问题,本文提出一种改进的Co-training式算法—CVCOT(Conditional value-based co-training),即采用基于样本条件价值的挑选策略来优化Co-training.通过定义无标记样本的条件价值,各视图下的分类器以样本条件价值为依据来挑选新增样本,以此更新训练集.该策略既可保证新增样本的标记可靠性,又能优先将价值较高的富信息样本补充到训练集中,可以有效地优化分类器.在UCI数据集和网页分类应用上的实验结果表明:CVCOT具有较好的分类性能和学习效率.

用于在线产品评论质量分析的Co-training算法

在线评论广泛存在于电子商务网站平台,其中包含着客户对产品的评价及偏好.高效分析在线评论数据并满足客户需求,对许多谋求立足于竞争激烈的国际化市场的企业来说至关重要.但因在线评论的质量不一,使得如何分析在线评论的质量成为一项重要工作.从两个方面提取特征对在线评论进行描述,并构建了一种Co-training算法来判断评论的质量.通过对比实验验证了该算法相对于单一分类算法的优势.

基于Co-training训练CRF模型的评价对象识别

评价对象是指某段评论中评价词语所修饰的对象或对象的属性。为了识别评论中的评价对象,提出基于Co-training的训练CRF模型方法。该方法首先人工标注少量的原始数据集,使用Co-training方式对未标注数据进行自动识别,以扩大已标注训练数据。通过原始标注数据集和Co-training方式标注数据集,训练CRF模型。在汽车领域中,对待标注汽车评论语料中评价对象识别的精确率为67.483%,召回率为67.832%。

Feature selection for co-training

Co-training is a semi-supervised learning method, which employs two complementary learners to label the unlabeled data for each other and to predict the test sample together. Previous studies show that redundant information can help improve the ratio of prediction accuracy between semi-supervised learning methods and supervised learning methods. However, redundant information often practically hurts the performance of learning machines. This paper investigates what redundant features have effect on the semi-supervised learning methods, e.g. co-training, and how to remove the redundant features as well as the irrelevant features. Here, FESCOT （feature selection for co-training） is proposed to improve the generalization performance of co-training with feature selection. Experimental results on artificial and real world data sets show that FESCOT helps to remove irrelevant and redundant features that hurt the performance of the co-training method.

Semi-supervised LIBS quantitative analysis method based on co-training regression model with selection of effective unlabeled samples

The accuracy of laser-induced breakdown spectroscopy(LIBS) quantitative method is greatly dependent on the amount of certified standard samples used for training. However, in practical applications, only limited standard samples with labeled certified concentrations are available. A novel semi-supervised LIBS quantitative analysis method is proposed, based on co-training regression model with selection of effective unlabeled samples. The main idea of the proposed method is to obtain better regression performance by adding effective unlabeled samples in semisupervised learning. First, effective unlabeled samples are selected according to the testing samples by Euclidean metric. Two original regression models based on least squares support vector machine with different parameters are trained by the labeled samples separately, and then the effective unlabeled samples predicted by the two models are used to enlarge the training dataset based on labeling confidence estimation. The final predictions of the proposed method on the testing samples will be determined by weighted combinations of the predictions of two updated regression models. Chromium concentration analysis experiments of 23 certified standard high-alloy steel samples were carried out, in which 5 samples with labeled concentrations and 11 unlabeled samples were used to train the regression models and the remaining 7 samples were used for testing. With the numbers of effective unlabeled samples increasing, the root mean square error of the proposed method went down from 1.80% to 0.84% and the relative prediction error was reduced from 9.15% to 4.04%.

Minimax entropy-based co-training for fault diagnosis of blast furnace

Due to the problems of few fault samples and large data fluctuations in the blast furnace(BF)ironmaking process,some transfer learning-based fault diagnosis methods are proposed.The vast majority of such methods perform distribution adaptation by reducing the distance between data distributions and applying a classifier to generate pseudo-labels for self-training.However,since the training data is dominated by labeled source domain data,such classifiers tend to be weak classifiers in the target domain.In addition,the features generated after domain adaptation are likely to be at the decision boundary,resulting in a loss of classification performance.Hence,we propose a novel method called minimax entropy-based co-training(MMEC)that adversarially optimizes a transferable fault diagnosis model for the BF.The structure of MMEC includes a dual-view feature extractor,followed by two classifiers that compute the feature's cosine similarity to representative vector of each class.Knowledge transfer is achieved by alternately increasing and decreasing the entropy of unlabeled target samples with the classifier and the feature extractor,respectively.Transfer BF fault diagnosis experiments show that our method improves accuracy by about 5%over state-of-the-art methods.

Recognition of Chinese Organization Name Using Co-training

Chinese organization name recognition is hard and important in natural language processing. To reduce tagged corpus and use untagged corpus,we presented combing Co-training with support vector machines (SVM) and conditional random fields (CRF) to improve recognition results. Based on principles of uncorrelated and compatible,we constructed different classifiers from different views within SVM or CRF alone and combination of these two models. And we modified a heuristic untagged samples selection algorithm to reduce time complexity. Experimental results show that under the same tagged data,Co-training has 10% F-measure higher than using SVM or CRF alone; under the same F-measure,Co-training saves at most 70% of tagged data to achieve the same performance.

Network traffic classification based on ensemble learning and co-training

Classification of network traffic is the essential step for many network researches. However, with the rapid evolution of Internet applications the effectiveness of the port-based or payload-based identification approaches has been greatly diminished in recent years. And many researchers begin to turn their attentions to an alternative machine learning based method. This paper presents a novel machine learning-based classification model, which combines ensemble learning paradigm with co-training techniques. Compared to previous approaches, most of which only employed single classifier, multiple classifters and semi-supervised learning are applied in our method and it mainly helps to overcome three shortcomings： limited flow accuracy rate, weak adaptability and huge demand of labeled training set. In this paper, statistical characteristics of IP flows are extracted from the packet level traces to establish the feature set, then the classification model is crested and tested and the empirical results prove its feasibility and effectiveness.

Weighted Co-Training for Cross-Domain Image SentimentClassification

Image sentiment classification, which aims to predict the polarities of sentiments conveyed by the images, has gained a lot of attention. Most existing methods address this problem by training a general classifier with certain visual features, ignoring the discrepancies across domains. In this paper, we propose a novel weighted co-training method for cross-domain image sentiment classification, which iteratively enlarges the labeled set by introducing new high-confidence classified samples to reduce the gap between the two domains. We train two sentiment classifiers with both the images and the corresponding textual comments separately, and set the similarity between the source domain and the target domain as the weight of a classifier. We perform extensive experiments on a real Flickr dataset to evaluate the proposed method, and the empirical study reveals that the weighted co-training method significantly outperforms some baseline solutions.

Towards Defense Against Adversarial Attacks on Graph Neural Networks via Calibrated Co-Training

Graph neural networks(GNNs)have achieved significant success in graph representation learning.Nevertheless,the recent work indicates that current GNNs are vulnerable to adversarial perturbations,in particular structural perturbations.This,therefore,narrows the application of GNN models in real-world scenarios.Such vulnerability can be attributed to the model’s excessive reliance on incomplete data views(e.g.,graph convolutional networks(GCNs)heavily rely on graph structures to make predictions).By integrating the information from multiple perspectives,this problem can be effectively addressed,and typical views of graphs include the node feature view and the graph structure view.In this paper,we propose C^(2)oG,which combines these two typical views to train sub-models and fuses their knowledge through co-training.Due to the orthogonality of the views,sub-models in the feature view tend to be robust against the perturbations targeted at sub-models in the structure view.C^(2)oG allows sub-models to correct one another mutually and thus enhance the robustness of their ensembles.In our evaluations,C^(2)oG significantly improves the robustness of graph models against adversarial attacks without sacrificing their performance on clean datasets.

Cross-lingual implicit discourse relation recognition with co-training

A lack of labeled corpora obstructs the research progress on implicit discourse relation recognition （DRR） for Chinese, while there are some available discourse corpora in other languages, such as English. In this paper, we propose a cross-lingual implicit DRR framework that exploits an available English corpus for the Chinese DRR task. We use machine translation to generate Chinese instances from a labeled English discourse corpus. In this way, each instance has two independent views： Chinese and English views. Then we train two classifiers in Chinese and English in a co-training way, which exploits unlabeled Chinese data to implement better implicit DRR for Chinese. Experimental results demonstrate the effectiveness of our method.

Towards making co-training suffer less from insufficient views

Co-training is a famous semi-supervised learning algorithm which can exploit unlabeled data to improve learning performance.Generally it works under a two-view setting (the input examples have two disjoint feature sets in nature),with the assumption that each view is sufficient to predict the label.However,in real-world applications due to feature corruption or feature noise,both views may be insufficient and co-training will suffer from these insufficient views.In this paper,we propose a novel algorithm named Weighted Co-training to deal with this problem.It identifies the newly labeled examples that are probably harmful for the other view,and decreases their weights in the training set to avoid the risk.The experimental results show that Weighted Co-training performs better than the state-of-art co-training algorithms on several benchmarks.

RFID indoor positioning based on semi-supervised actor-critic co-training

For large-scale radio frequency identification(RFID) indoor positioning system, the positioning scale is relatively large, with less labeled data and more unlabeled data, and it is easily affected by multipath and white noise. An RFID positioning algorithm based on semi-supervised actor-critic co-training(SACC) was proposed to solve this problem. In this research, the positioning is regarded as Markov decision-making process. Firstly, the actor-critic was combined with random actions and the unlabeled best received signal arrival intensity(RSSI) data was selected by co-training of the semi-supervised. Secondly, the actor and the critic were updated by employing Kronecker-factored approximation calculate(K-FAC) natural gradient. Finally, the target position was obtained by co-locating with labeled RSSI data and the selected unlabeled RSSI data. The proposed method reduced the cost of indoor positioning significantly by decreasing the number of labeled data. Meanwhile, with the increase of the positioning targets, the actor could quickly select unlabeled RSSI data and updates the location model. Experiment shows that, compared with other RFID indoor positioning algorithms, such as twin delayed deep deterministic policy gradient(TD3), deep deterministic policy gradient(DDPG), and actor-critic using Kronecker-factored trust region(ACKTR), the proposed method decreased the average positioning error respectively by 50.226%, 41.916%, and 25.004%. Meanwhile, the positioning stability was improved by 23.430%, 28.518%, and 38.631%.

基于K-means算法的Co-trainning的研究

在Co-trainning算法中通过两个训练集互相校正来达成分类,这里两个训练集所用的特征集对结果影响很大,选取两个好的特征集也就可以使Co-trainning算法结果更优。K-means算法是一种聚类算法,在对K-means算法研究和实现时,设计并实验将K-means算法思想运用到Co-trainning算法特征集选取上,效果较好。

基于情感标签的极性分类

情感极性分析是文本挖掘中一种非常重要的技术.然而在不同领域中,很多情感极性分类系统存在分类精度低和缺少大量标注数据的缺陷.针对这些问题,提出了一种基于情感标签的极性分类方法.首先通过所有文本建立Sentiment-Topic模型,抽取出文本的情感标签;然后利用情感标签将文本划分为两个子文本,并通过Co-training算法对子文本进行分类;最后合并两个子文本的分类结果,并确定文本的情感极性.实验结果表明该方法具有较高的分类精度,而且不需要大量的分类样本.