基于随机森林模型的电信运营商外呼推荐系统

在电信运营商领域,外呼推荐是一种重要的推荐产品和服务的途径。实现了一种基于运营商大数据的自动外呼推荐系统,该系统能够挖掘用户的行为特征并且使用机器学习的方法预测用户对于被推荐产品的接受可能性。传统推荐系统使用的模型算法为矩阵分解、大规模稀疏特征分类、神经网络等。采用随机森林算法的主要原因是随机森林具有并行化程度高、训练速度快、生成的决策树可解释等诸多优点,适合于基于电信业数据的推荐系统。该外呼推荐系统基于Hadoop、Impala和Spark等大数据处理平台及工具,使用随机森林分类器作为核心算法,将用户最近的行为特征回归为接受外呼推荐产品的可能性。在线测试表明使用该系统与当前部署的人工随机外呼相比,能够提升约41%的用户接受率;同时,根据模型算法输出特征的重要性,进一步给出了两类用户的特征分析。     

面向图像分类的深度模型可解释性研究综述

深度学习目前在计算机视觉、自然语言处理、语音识别等领域得到了深入发展,与传统的机器学习算法相比,深度模型在许多任务上具有较高的准确率.然而,作为端到端的具有高度非线性的复杂模型,深度模型的可解释性没有传统机器学习算法好,这为深度学习在现实生活中的应用带来了一定的阻碍.深度模型的可解释性研究具有重大意义而且是非常必要的,近年来许多学者围绕这一问题提出了不同的算法.针对图像分类任务,将可解释性算法分为全局可解释性和局部可解释性算法.在解释的粒度上,进一步将全局解释性算法分为模型级和神经元级的可解释性算法,将局部可解释性算法划分为像素级特征、概念级特征以及图像级特征可解释性算法.基于上述分类框架,总结了常见的深度模型可解释性算法以及相关的评价指标,同时讨论了可解释性研究面临的挑战和未来的研究方向.认为深度模型的可解释性研究和理论基础研究是打开深度模型黑箱的必要途径,同时可解释性算法存在巨大潜力可以为解决深度模型的公平性、泛化性等其他问题提供帮助.     

随机森林的可解释性可视分析方法研究

由于随机森林算法在很多情况下都以“黑盒”的方式存在,对于用户而言,参数调整,训练甚至最终构建的模型细节是隐蔽的,这导致了随机森林模型的可解释性非常差,在一定程度上阻碍了该模型在一些诸如医学诊断、司法、安全领域等需要透明化和可解释需求比较高的领域使用。影响该模型可解释性挑战主要来源于特征选择和数据的随机性。同时随机森林包含许多决策树,用户很难理解和比较所有决策树的结构和属性。为了解决上述问题,设计并实现了可视分析系统FORESTVis,该系统包括树视图、部分依赖图、t-SNE投影图、特征视图等多个交互式可视化组件,借助该系统,相关研究人员和从业人员可以直观地了解随机森林的基本结构和工作机制,并协助用户对模型的性能进行评估。使用Kaggle公开数据集上进行案例分析,验证了该方法的可行性和有效性。     

预训练语言模型实体匹配的可解释性

实体匹配可以判断两个数据集中的记录是否指向同一现实世界实体,对于大数据集成、社交网络分析、网络语义数据管理等任务不可或缺.作为在自然语言处理、计算机视觉中取得大量成功的深度学习技术,预训练语言模型在实体识别任务上也取得了优于传统方法的效果,引起了大量研究人员的关注.然而,基于预训练语言模型的实体匹配技术效果不稳定、匹配结果不可解释,给这一技术在大数据集成中的应用带来了很大的不确定性.同时,现有的实体匹配模型解释方法主要面向机器学习方法进行模型无关的解释,在预训练语言模型上的适用性存在缺陷.因此,以Ditto、JointBERT等BERT类实体匹配模型为例,提出3种面向预训练语言模型实体匹配技术的模型解释方法来解决这个问题:(1)针对序列化操作中关系数据属性序的敏感性,对于错分样本,利用数据集元特征和属性相似度实现属性序反事实生成;(2)作为传统属性重要性衡量的补充,通过预训练语言模型注意力机制权重来衡量并可视化模型处理数据时的关联性;(3)基于序列化后的句子向量,使用k近邻搜索技术召回与错分样本相似的可解释性优良的样本,增强低置信度的预训练语言模型预测结果.在真实公开数据集上的实验结果...     

基于用户注意力与视觉注意力的社交图像描述

图像描述是机器学习和计算机视觉的重要研究领域,但现有方法对于视觉特征和模型架构之间存在的语义信息关联性探索还存在不足.本文提出了一种基于用户标签、视觉特征的注意力模型架构,能够有效地结合社交图像特征和图像中用户标签生成更加准确的描述.我们在MSCOCO数据集上进行了实验来验证算法性能,实验结果表明本文提出的基于用户标签、视觉特征的注意力模型与传统方法相比具有明显的优越性.     

强化学习可解释性基础问题探索和方法综述

强化学习是一种从试错过程中发现最优行为策略的技术,已经成为解决环境交互问题的通用方法.然而,作为一类机器学习算法,强化学习也面临着机器学习领域的公共难题,即难以被人理解.缺乏可解释性限制了强化学习在安全敏感领域中的应用,如医疗、驾驶等,并导致强化学习在环境仿真、任务泛化等问题中缺乏普遍适用的解决方案.为了克服强化学习的这一弱点,涌现了大量强化学习可解释性(explainable reinforcement learning,XRL)的研究.然而,学术界对XRL尚缺乏一致认识.因此,探索XRL的基础性问题,并对现有工作进行综述.具体而言,首先探讨父问题——人工智能可解释性,对人工智能可解释性的已有定义进行了汇总;其次,构建一套可解释性领域的理论体系,从而描述XRL与人工智能可解释性的共同问题,包括界定智能算法和机械算法、定义解释的含义、讨论影响可解释性的因素、划分解释的直观性;然后,根据强化学习本身的特征,定义XRL的3个独有问题,即环境解释、任务解释、策略解释;之后,对现有方法进行系统地归类,并对XRL的最新进展进行综述;最后,展望XRL领域的潜在研究方向.     

目标检测模型的决策依据与可信度分析

目标检测模型已经在很多领域得到广泛应用, 但是, 作为一种机器学习模型, 对人类来说仍然是一个黑盒. 对模型进行解释有助于我们更好地理解模型, 并判断其可信度. 针对目标检测模型的可解释性问题, 提出将其输出改造为关注每一类物体存在性概率的具体回归问题, 进而提出分析目标检测模型决策依据与可信度的方法. 由于原有图像分割方法的泛用性较差, 解释目标检测模型时, LIME所生成解释的忠诚度较低、有效特征数量较少. 提出使用DeepLab代替LIME的图像分割方法, 以对其进行改进. 改进后的方法可以适用于解释目标检测模型. 实验的对比结果证明了所提出改进方法在解释目标检测模型时的优越性.     

InterDroid:面向概念漂移的可解释性Android恶意软件检测方法

针对Android恶意软件检测存在特征引入过程主观性高、特征选择过程可解释性差、训练模型检测效果不具备时间稳定性的问题,提出了一种面向概念漂移的可解释性Android恶意软件检测方法InterDroid,该方法首先通过高质量的人工Android恶意软件分析报告引入权限、API包名、意图、Dalvik字节码4种特征.并通过自动化机器学习算法TPOT(tree-based tipeline optimization tool)获得InterDroid训练及对比算法,从而摒弃传统方法中繁复的模型选择与参数调整过程.其后,融入模型解释算法SHAP(shapley additive explanations)改进传统的特征包装方法,从而获得对分类结果具有高贡献度的特征组合用于检测模型训练最后,通过曼-惠特尼U(Mann-Whitney U,MWU)与机器学习模型的双重检验证明概念漂移现象在Android恶意软件检测中的存在性.并基于联合分布适配(joint distribution adaptation,JDA)算法提高检测模型对新时期Android恶意软件的检测准确率.实验表明:InterDroid筛选出的特征组合具备稳定性与可解释性.同时,InterDroid中的特征迁移模块可将自身对2019年、2020年新兴Android恶意软件的检测准确率分别提高46％,44％.     

特征挖掘算法与不同机器学习方法在回归问题中的应用研究

近年来,机器学习方法在各个领域内的应用十分广泛,而特征挖掘则是机器学习过程中十分重要的数据预处理过程。Boruta算法是一种基于随机森林算法的特征筛选算法,该算法可以对机器学习过程中的自变量进行筛选,并且进行重要性排序。随机森林、支持向量机、k最近邻算法和梯度提升模型则是机器学习中的经典算法,在解决回归和分类问题中均有较广泛的应用。本文通过红葡萄酒案例,将Boruta算法与4种机器学习方法相结合,探讨Boruta算法的作用及几种不同机器学习方法的区别,为将特征挖掘算法和机器学习在回归问题中的应用提供参考。     

一种预测流程剩余时间的可解释特征分层方法

流程剩余时间预测对于业务异常的预防和干预有着重要的价值和意义.现有的剩余时间预测方法通过深度学习技术达到了更高的准确率,然而大多数深度模型结构复杂难以解释预测结果,即不可解释问题.此外,剩余时间预测除了活动这一关键属性还会根据领域知识选择若干其他属性作为预测模型的输入特征,缺少通用的特征选择方法,对于预测的准确率和模型的可解释性存在一定的影响.针对上述问题,提出基于可解释特征分层模型(explainable feature-based hierarchical model,EFH model)的流程剩余时间预测框架.具体而言,首先提出特征自选择策略,通过基于优先级的后向特征删除和基于特征重要性值的前向特征选择,得到对预测任务具有积极影响的属性作为模型输入.然后提出可解释特征分层模型架构,通过逐层加入不同特征得到每层的预测结果,解释特征值与预测结果的内在联系.采用LightGBM (light gradient boosting machine)和LSTM (long short-term memory)算法实例化所提方法,框架是通用的,不限于选用算法.最后在8个真实事件日志上与最新方法进行比较.实验结果表明所提方法能够选取出有效特征,提高预测的准确率,并解释预测结果.     

Minimal-unsatisfiable-core-driven Local Explainability Analysis for Random Forest

As Machine Learning (ML) is widely applied in security-critical fields, the requirements for the interpretability of ML also increase. The interpretability aims at helping people understand internal operation principles and decision principles of models, so as to improve models'' credibility. However, research on the interpretability of ML models such as Random Forest (RF) is still in the infant stage. Considering the strict and standardized characteristics of formal methods and their wide application in the field of ML in recent years, this study leverages formal methods and logical reasoning to develop an ML interpretability method for interpreting the prediction of RF. Specifically, the decision-making process of RF is encoded into a first-order logical formula, and a Minimal Unsatisfiable Core (MUC) is taken as the core. Local interpretation of feature importance and counterfactual sample generation methods are provided. Experimental results on several public datasets illustrate the high quality of the proposed feature importance measurement, and the counterfactual sample generation method outperforms the existing state-of-the-art methods. Moreover, from the perspective of user-friendliness, the user report can be generated according to the analysis results of counterfactual samples, which can provide suggestions for users to improve their situation in real-life applications.     

An efficient semantic – Related image retrieval method

Many previous techniques were designed to retrieve semantic images in a certain neighborhood of the query image and thus bypassing the semantically related images in the whole feature space. Several recently methods were designed to retrieve semantically related images in the entire feature space but with low precision. In this paper, we propose a Semantic – Related Image Retrieval method (SRIR), which can retrieve semantic images spread in the entire feature space with high precision. Our method takes advantage of the user feedback to determine the semantic importance of each query and the importance of each feature. In addition, the retrieval time of our method does not increase with the number of user feedback. We also provide experimental results to demonstrate the effectiveness of our method.     

Character motion synthesis by principal component analysis and motion control interface by hands

In this paper, we propose an interactive character motion control interface that uses hands. Using their hands and fingers, the user can control a large number of degrees of freedom at the same time. We applied principal component analysis to a set of sample poses and assigned the extracted principal components to each degree of freedom of the hands (such as the hand positions and finger bending/extending angles). The user can control the blending weights of the principal components and deform the character's pose by moving their hands and bending/extending their fingers. We introduced pose and action controls, so that we can alter the standing pose and perform various actions with deformations. So that various types of actions were possible, we constructed a number of action models in advance. We introduced action model selection and action execution mechanisms. We developed methods for computing the feature vector, for applying principal component analysis, and for pose and action synthesis. In addition, we introduced a pose transition method for performing a step motion when necessary to prevent foot sliding. We present our experimental results and demonstrate the effectiveness of our interface. Copyright © 2015 John Wiley & Sons, Ltd.     

Multi-scale hierarchical sampling change detection using Random Forest for high-resolution satellite imagery

ABSTRACT

High-resolution imagery provides rich information useful for land-use and land-cover change detection; however, methods to exploit these data lag behind data collection technologies. In this article, we propose a novel object-oriented multi-scale hierarchical sampling (MSHS) change detection method for high-resolution satellite imagery. In our method, MSHS is carried out to automatically obtain multi-scale training samples and different sample combinations. The training sample spectra, texture, and shape features are fused to build feature space after MSHS. Sample combinations and corresponding feature spaces are input into Random Forest (RF) to train multiple change classifiers. An optimal RF change detection classifier is selected when the out-of-bag error parameter in RF is at the minimum. In order to validate the proposed method, we applied it to high-resolution satellite image data and compared the detection results from our method and the single-scale sampling change detection method. These experimental results show that false alarm rates and missed detection of changed objects using our method were lower than the single-scale sampling change detection method. To demonstrate the scalability of the algorithm, different change detection methods were applied to three study sites. Experimental results show that our method delivered high overall accuracy and F₁-scores. Compared to traditional methods, our method makes full use of the multi-scale characteristics of ground objects. Our approach does not extend multi-scale feature vectors directly, but instead automatically increases the amount of the training samples at multiple scales, without increasing the volume of manual processing, thus improving the ability of the algorithm to generalize features from the RF model, making it more robust.     

Decision Boundary Visualization for Counterfactual Reasoning

Machine learning algorithms are widely applied to create powerful prediction models. With increasingly complex models, humans' ability to understand the decision function (that maps from a high-dimensional input space) is quickly exceeded. To explain a model's decisions, black-box methods have been proposed that provide either non-linear maps of the global topology of the decision boundary, or samples that allow approximating it locally. The former loses information about distances in input space, while the latter only provides statements about given samples, but lacks a focus on the underlying model for precise ‘What-If'-reasoning. In this paper, we integrate both approaches and propose an interactive exploration method using local linear maps of the decision space. We create the maps on high-dimensional hyperplanes—2D-slices of the high-dimensional parameter space—based on statistical and personal feature mutability and guided by feature importance. We complement the proposed workflow with established model inspection techniques to provide orientation and guidance. We demonstrate our approach on real-world datasets and illustrate that it allows identification of instance-based decision boundary structures and can answer multi-dimensional ‘What-If'-questions, thereby identifying counterfactual scenarios visually.     

Ontology based automatic feature recognition framework

AFR has long been realized as a key technology for design automation. A significant shortcoming in AFR is that most of them are individual systems that are isolated from each other, due to the absence of a standard feature library or feature modeling techniques. Few studies attempted to overcome this problem by allowing a certain degree of user customization or extension, which are still far from success. In order to address this issue, this paper proposes an ontology-based feature recognition framework. In the framework, features are captured transparently and hierarchically within a formal OWL ontology, and the feature recognition is achieved by applying an efficient backward-chained ontology reasoner to reason through the ontology. The resulting feature recognition system shows a high level of flexibility, maintainability, and explainability, for both representing and recognizing features. The effectiveness of the framework is finally demonstrated with three case studies.     

Counterfactual inference to predict causal knowledge graph for relational transfer learning by assimilating expert knowledge --Relational feature transfer learning algorithm

Transfer learning (TL) is a machine learning (ML) method in which knowledge is transferred from the existing models of related problems to the model for solving the problem at hand. Relational TL enables the ML models to transfer the relationship networks from one domain to another. However, it has two critical issues. One is determining the proper way of extracting and expressing relationships among data features in the source domain such that the relationships can be transferred to the target domain. The other is how to do the transfer procedure. Knowledge graphs (KGs) are knowledge bases that use data and logic to graph-structured information; they are helpful tools for dealing with the first issue. The proposed relational feature transfer learning algorithm (RF-TL) embodies an extended structural equation modelling (SEM) as a method for constructing KGs. Additionally, in fields such as medicine, economics, and law related to people’s lives and property safety and security, the knowledge of domain experts is a gold standard. This paper introduces the causal analysis and counterfactual inference in the TL domain that directs the transfer procedure. Different from traditional feature-based TL algorithms like transfer component analysis (TCA) and CORelation Alignment (CORAL), RF-TL not only considers relations between feature items but also utilizes causality knowledge, enabling it to perform well in practical cases. The algorithm was tested on two different healthcare-related datasets — sleep apnea questionnaire study data and COVID-19 case data on ICU admission — and compared its performance with TCA and CORAL. The experimental results show that RF-TL can generate better transferred models that give more accurate predictions with fewer input features.     

动态数据流环境下的自适应性行为识别算法设计

传统的行为识别系统大多是建立在一个静态模型上,对样本特征值有非常强的依赖性,而对个别用户的行为习惯缺乏灵活适应性.基于三轴加速度计设计了一种在动态数据流环境下能够对用户行为进行增量学习的自适应性行为识别算法,该算法提出了一种新的提取加速度特征的方法,通过将三轴加速度计采集到的合成加速度数据集合看做物质,来提取物质的物理特征并训练这些物理特征用于投票分类,然后再通过增量学习来更新样本特征值,使其逐渐趋向于用户的行为习惯,从而达到更高的识别率.实验结果表明,该算法具有很高的识别率和对用户很好的适应性.     

面向可解释性的物体拓扑结构骨架表征方法

目的 模式识别中，通常使用大量标注数据和有效的机器学习算法训练分类器应对不确定性问题。然而，这一过程缺乏知识表征和可解释性。认知心理学和实验心理学的研究表明，人类往往不使用代价如此巨大的机制，而是使用表征、归纳、推理、解释和约束传播等与符号主义人工智能方法类似的手段来应对物体识别中的不确定性并提供可解释性。因此本文旨在从传统的符号计算出发，利用骨架拓扑结构表征提供一种可解释性的思路。方法 以骨架树为基本手段来形成物体拓扑结构特征和几何特征的形式化表征，并基于泛化框架对少量同类表征进行知识抽取来形成关于物体类别的知识概括显式化表征。结果 在形成物体类别的概括表征实验中，通过路径重建直观展示了同类属物体上得到的最一般表征的几何物理意义。在可解释性验证实验中，通过跨数据的拓扑应用展示了新测试样本相对于概括表征的特定差异，表明该表征具有良好的可解释性。最后在形状补全的不确定性推理实验中，不仅可以得到识别结论，而且清晰展示了识别背后做出的判断依据，进一步验证了该表征的可解释性。结论 实验表明一般化的形式表征能够应对尺寸、颜色和形状等不确定性问题，本文方法避免了基于纹理特征所带来的不确定性，适用于任意基于基元的表征方式，具有更好的鲁棒性、普适性和可解释性，计算代价更小。     

Multimedia information retrieval based on pairwise comparison and its application to visual search

Finding a way to elicit user preferences in the context of multimedia information retrieval is an important issue that remains to be solved. Users are not usually able to find a sought after image or provide an example of what they want. One of several possible methods that might be used to solve this problem involves reasoning about user queries through the assessment of several samples. In this article we propose a method by which user queries are retrieved based on the pairwise comparison of sample alternatives. Pairwise comparison was originally designed for the ranking of alternatives. In our method we rank criteria according to their importance for the user and then use this information to retrieve relevant records from the database. The method was implemented in Matlab and tested on the Microsoft Research Cambridge Image Database.