关键词： 多因子选股   随机森林   XGBOOST  

摘要： 量化投资是一种从历史数据提取出有效经验,并将这种经验应用于指导未来的一种投资方式。与传统的主观投资不同,量化投资不受人的主观情绪波动的影响,从而可以保持投资逻辑的稳定性和一致性。多因子选股作为权益类市场量化投资的重要策略一直以来都是学术界和业界关注的焦点。与传统的多因子选股模型通常关注标的资产的收益与各因子之间的线性关系不同,本文第一部分基于决策树模型可以发掘变量间的非线性和关系和解释性强的特点,对2006年4月至2018年6月的全A股市场数据建立了基于决策树的随机森林和XGBOOST两个集成算法模型的选股策略。并对上述策略分别进行了纯多头和市场中性的策略回测。其中18个月滚动训练的随机森林模型取得了纯多头23.8%的年化收益和0.673的夏普比率,市场中性26.7%的年化收益和1.447的夏普比率;18个月滚动训练的XGBOOST模型取得了纯多头26.8%的年化收益和0.782的夏普比率,市场中性28.9%的年化收益和1.594的夏普比率。纯多头策略和市场中性策略都远强于沪深300和中证500指数基准。第二部分首先提出了基于Focal-loss损失函数的XGBOOST模型以期在难分样本上取得更优效果,但结果并无明显变化。其次基于2017年市场风格切换导致的因子失效以及未来可能再次出现相同情况的现实,提出了基于不同长度训练集模型的模型融合方法,有效地提高了模型对于市场变化的反应能力。本文基于我国A股市场风格总体保持稳定的事实,放松了对于机器学习训练集和测试集样本来源同分布的假设。因此将时间序列数据截面化的思想的有效性仍有待市场样本外的表现加以验证。

关键词： 工业能源管理   工厂设备   能耗预测   机器学习   流式批处理  

摘要： 能源紧缺与环境污染等因素正在驱动制造行业探索更加高效的节能环保型生产。在第四次工业革命的背景下，将能耗预测功能用于工业能源管理的决策任务是改善工厂生产过程的有效途径。在赛博物理系统的层面上，可利用机器学习方法对广泛收集的制造信息数据进行分析，并将分析结果反馈给制造系统，作为一份决策报告用于生产过程优化。因此，基于机器学习方法的能耗预测系统具有重要的研究与应用意义。本文的研究目标旨在探索天气，设备工作时长和日期等因素对工厂设备能耗量的影响，并根据其历史数据建立一套可以用户自定义的模块化的能耗预测系统，使得系统在未来能够根据输入的当日天气等信息预测当日的设备能耗量，为能源管理提供前瞻性的决策支持。论文主要研究内容如下：　　1.基于CRISP-DM数据挖掘流程搭建一个针对设备能耗数据的数据处理流程，包括数据理解、数据预处理、建模和评估以及模型部署四个阶段。　　2.该系统实现了能耗信息的收集，在数据预处理和建模阶段设置可以用户自定义的模型及方法选择功能，使得用户可以在数据标准化，特征选择以及建模方法这三个步骤自行选择具体模型及方法，经选择后的模型在被导入Pipeline流式批处理管道后整合为一套机器学习模型。通过历史数据训练模型后，模型将自动生成各模块中的最优超参数。　　3.在用户可以自定义的步骤中对比各方法的性能表现，分析能耗的影响因素。包括特征选择阶段对缺失值比率、反向特征消除、主成分分析等方法的比较，以及在建模阶段对广义线性回归模型，决策树，支持向量机，多层感知机，决策树和集成学习方法的对比。　　4.系统在收集了2017年1月至2019年3月期间德国沃尔夫斯堡OpenHybrid实验室工厂所有设备每日的能耗数据，当地天气数据及日期信息后，将训练好的机器学习模型导入浏览器端的人机交互平台，用于未来对设备能耗的预测。

关键词： 电力巡检   计算机视觉   机器学习   小样本学习  

摘要： 巡检无人机成本低廉,智能化程度高,可对输电线路进行快速且安全的巡检工作,具有广阔应用前景,是当前电力巡检领域的热点研究项目。然而,电力巡检中类别繁杂,难以为每个类别都收集到足够的训练样本,且为训练样本人工标注的代价过于高昂。小样本分类算法适合于解决此类给定了丰富基础类别的训练样例,却缺乏新类标记实例的情况。但针对电力巡检这一具体的业务场景,仍缺乏相应研究。因此,需要进一步研究小样本分类算法,理清其原理、明确其数据基础,以期能将其成功应用于智能电力巡检系统。本文在经典分类神经网络模型的基础上进行了两个方法改进:增加了基于注意力机制的小样本分类权重发生器;将卷积神经网络模型的分类器更改为测试样本特征向量与类向量之间的余弦相似性函数。该方法在保留了其它小样本分类方法能较好地减少类内方差这一优点的基础上,对于较高域差异的训练集也体现了较好的自适应能力。综上所述,本文首先总结了近年来小样本学习方法的进展与成果,并阐述各算法的特点、主要针对的问题,比较分析了几种典型的小样本学习方法,包括匹配网络、原型网络、关系网络以及模型无关元学习方法。其次,设计了一个一致性的实验方案,减少因各小样本算法复杂的算法结构以及差异化的训练细节而导致的性能差异,并以此定量分析了各算法实际性能。此后探讨了小样本学习方法的训练目标,证明了更深的骨干网络显著降低了各方法之间的性能差异,并通过实验发现类内方差是造成小样本学习性能差异的主要因素之一。最后,实验结果还表明随着域差异增大,增强对于少数新类实例的自适应能力将比减少类内方差更能提高算法的准确率,该结果也对电力巡检领域数据集的制备工作提供了参考。

关键词： Alzheimer's disease   classification   early diagnosis   machine learning   magnetic resonance imaging   mild cognitive impairment   predictive analytics  

摘要： Background: Predicting clinical course of cognitive decline can boost clinical trials' power and improve our clinical decision-making. Machine learning (ML) algorithms are specifically designed for the purpose of prediction;however. identifying optimal features or algorithms is still a challenge. Objective: To investigate the accuracy of different ML methods and different features to classify cognitively normal (CN) individuals from Alzheimer's disease (AD) and to predict longitudinal outcome in participants with mild cognitive impairment (MCI). Methods: A total of 1,329 participants from the Alzheimer's Disease Neuroimaging Initiative (ADNI) were included: 424 CN, 656 MCI, and 249 AD individuals. Four feature-sets at baseline (hippocampal volume and volume of 47 cortical and subcortical regions with and without demographics and APOE4) and six machine learning methods (decision trees, support vector machines, K-nearest neighbor, ensemble linear discriminant, boosted trees, and random forests) were used to classify participants with normal cognition from participants with AD. Subsequently the model with best classification performance was used for predicting clinical outcome of MCI participants. Results: Ensemble linear discriminant models using demographics and all volumetric magnetic resonance imaging measures as feature-set showed the best performance in classification of CN versus AD participants (accuracy = 92.8%, sensitivity = 95.8%, and specificity = 88.3%). Prediction accuracy of future conversion from MCI to AD for this ensemble linear discriminant at 6, 12, 24, 36, and 48 months was 63.8% (sensitivity = 74.4, specificity = 63.1), 68.9% (sensitivity = 75.9, specificity = 67.8), 74.9% (sensitivity = 71.5, specificity = 76.3), 75.3%, (sensitivity = 65.2, specificity = 79.7), and 77.0% (sensitivity = 59.6, specificity = 86.1), respectively. Conclusions: Machine learning models trained for classification of CN versus AD can improve our prediction ability of MCI

关键词： 准确性   高效性   目标候选区域   行人检测  

摘要： 目标候选框可以完成对目标物体的初步定位,是目标检测和目标识别的基础,既适合对多类目标的定位也适合对单类目标的定位。在实时监控系统、门禁系统、刷脸支付等实际环境中具有重要作用。所以,目标候选框的准确高效定位具有重要研究意义。本文主要从目标候选框生成的高效性与准确性两个方面进行研究。(1)本文主要针对BING方法定位准确率不高的问题,首先使用扩展模板提取特征增强物体边缘特征的判别力,从而改进原梯度特征定位偏差的问题。然后采用高斯软权重非最大值抑制算法,解决了最大目标候选框附近非最大目标候选区域被强制归零的问题,提高了BING方法定位的准确率。最后在公开的数据集上对改进算法进行验证。实验结果表明,改进的BING算法在不增加算法计算复杂度的基础上,提高了BING方法定位的准确率。(2)本文基于具有较高准确率的选择性搜索方法进行两方面的改进,一是在初始分割阶段使用汉明距离计算相邻像素之间的相似性,二是针对选择性搜索方法合并策略计算复杂的问题,使用一种基于颜色的哈希策略的方法计算图像初始分割区域的相似度,从而判断相邻区域是否可以合并。这种方法大大地减少了选择性搜索算法在区域合并时计算的迭代次数,并在公开的测试集上进行验证。结果表明,本文改进的选择性搜索方法在尽可能不降低准确率的基础上,效率提高了60%。基于目标候选框在DPM行人检测算法中的应用。本文使用改进BING算法和改进选择性搜索算法代替DPM行人检测算法中的滑动窗口完成对行人的检测。通过行人数据集INRIA对检测结果进行验证,实验结果表明,使用改进目标候选框结合DPM的方法检测行人,加快了DPM行人检测的速度,实现了对行人的高效准确定位。

关键词： Medical imaging   Biomechanics   Artificial intelligence  

摘要： Machine learning with deep neural networks has demonstrated high performance for high dimensionality prediction tasks across multiple domains with sufficient sample data. Cardiovascular disease is a pertinent public health issue that has the potential to be better understood and addressed via deep learning approaches. In this work, we study machine learning approaches toward diagnosing various forms of cardiovascular disease and predicting its biomechanical behavior across multiple scales. We begin by training deep learning models for an initial classification objective in echocardiography, a ubiquitous imaging modality for cardiologists. For view classification, we are able to demonstrate physician-level performance. We then expand the work from a methods and clinical application perspective. We address the high cost of annotation in medical imaging by examining data-efficient supervised and semi-supervised algorithms. In addition, we expand our prediction tasks towards the ultimate goal of automated, accurate cardiovascular disease diagnosis by predicting left ventricular hypertrophy. To understand the nature of cardiovascular disease and develop treatments, a close look at the underlying biomechanics is important. For atherosclerosis, a leading cause of morbidity and mortality, we bridge finite element methods and machine learning to predict arterial tissue stress. Likewise for cytoskeletal proteins, which are the structural building blocks of human biology and influence cardiovascular health, we develop graph neural network algorithms to predict force response and conformational dynamics in calponin homology domains. Moreover, we hope to lay the groundwork to advance the intersection of machine learning, biomechanics, and cardiovascular disease.

关键词： Crisis   Event Detection   Text Mining   Classification   Ranking  

摘要： The use of social media is expanding significantly and can serve a variety of purposes. Over the last few years, users of social media have played an increasing role in the dissemination of emergency and disaster information. It is becoming more common for affected populations and other stakeholders to turn to Twitter to gather information about a crisis when decisions need to be made, and action is taken. However, social media platforms, especially on Twitter, presents some drawbacks when it comes to gathering information during disasters. These drawbacks include information overload, messages are written in an informal format, the presence of noise and irrelevant information. These factors make gathering accurate information online very challenging and confusing, which in turn may affect public, communities, and organizations to prepare for, respond to, and recover from disasters. To address these challenges, we present an integrated three parts (clustering-classification-ranking) framework, which helps users choose through the masses of Twitter data to find useful information. In the first part, we build standard machine learning models to automatically extract and identify topics present in a text and to derive hidden patterns exhibited by a dataset. Next part, we developed a binary and multi-class classification model of Twitter data to categorize each tweet as relevant or irrelevant and to further classify relevant tweets into four types of community engagement: reporting information, expressing negative engagement, expressing positive engagement, and asking for information. In the third part, we propose a binary classification model to categorize the collected tweets into high or low priority tweets. We present an evaluation of the effectiveness of detecting events using a variety of features derived from Twitter posts, namely: textual content, term frequency-inverse document frequency, Linguistic, sentiment, psychometric, temporal, and spatial. Our framework

关键词： Plant sciences   Bioinformatics   Computer science  

摘要： Biological systems contain data of high dimensions and magnitudes, including biochemistry, cellular patterning, transcriptomics, and genomics. Here I combined network analyses and machine learning, to identify developmental patterns that may be amenable to the improvement of drought tolerance. Cuticles comprise the hydrophobic lipid layer covering the aboveground plant body, and have long been a research focus into water conservation in plants. However, no prior studies have examined cuticle development across a temporal and spatial gradient in a crop plant. I used gene network analyses to correlate the biochemical/developmental gradient of cuticle components with the underlying transcriptomic transitions to identify the role of PHYTOCHROME B-mediated light signaling in cuticle development. Subsequent statistical and biochemical analysis revealed LIPID-TRANSFER PROTEINs as evolutionary novelties contributing to the emergence of cuticles in land plants. Additionally, combining the power of genome- and transcriptome-wide association studies (GWAS and TWAS), vesicular trafficking was implicated in the regulation of cuticular evaporation rate. Water loss through the leaf surface is also moderated by specialized cell types (bulliform cells) in maize. Bulliform cell ontogeny was investigated in the developing maize leaf, and a machine learning approach (convolutional neural networks) was employed to conduct high-throughput phenotyping of microscopic bulliform cell traits in 60,780 leaf epidermal glue-impression images. A subsequent GWAS analysis on bulliform cell column number and column width identified a set of gene candidates implicated to function in cell division and DNA methylation. Overall this dissertation demonstrates a multidisciplinary approach combining developmental biology, transcriptomics, quantitative genetics, machine learning, and statistical data analysis, toward a more holistic understanding of the mechanisms of water conservation in maize.

关键词： Statistics   Computer science  

摘要： We present contributions to the application areas of photography and social sciences using a statistical machine learning approach, as well as methodological work in dimension reduction techniques. We first present a machine-learning based photography project in which we analyzed patterns of rule break among high-quality photographs. While high-quality photographs generally adhere to high-level rules set forth in photography, some high-quality photographs purposefully bend the rules to achieve a greater effect. Our work in photography solidifies the patterns of conditions that allow for such bends in rules. To do so, we used a machine learning and image processing based approach to find and evaluate rule breaking photographs in order to discover the patterns that justify their rule breaking. Next, we present an analysis of whether gender bias exists in the field of statistics. Specifically, we quantify the differences in citation counts between men and women in the statistics field, using a sample of papers from the top statistics journals. We first identified the gender of the first author of each paper in our dataset. We also added important covariates such as job title and country of employment of the first author, in order to account for factors that may possibly affect or possibly explain any disparity in citation counts between men and women. Our analysis found that, controlling for covariates and replications, that there is no evidence to suggest that there is a citation disparity between men and women authors in statistics. We also developed a modification of the locally linear embedding (LLE) dimension reduction algorithm that is designed to handle additive noise. This new modification is termed LLEAN, short for locally linear embedding with additive noise, and has been shown to perform better in the presence of noise distortion. LLEAN seeks to recover the noiseless data from the noisy data by exploiting the relationship between local linearity and reconstr

关键词： Autonomous   Latency   Self-adaptive   Tactic  

摘要： From self-driving cars to self-adaptive websites, the world is increasingly becoming more reliant on autonomous systems. Similar to many other domains, the system's behavior is often determined by its requirements. For example, a self-adaptive web service is likely to have some maximum value that response time should not surpass. To maintain this requirement, the system uses tactics, which may include activating additional computing resources. In real-world environments, tactics will frequently experience volatility, known as tactic volatility. This can include unstable time required to execute the tactic or frequent fluctuations in the cost to execute the tactic. Unfortunately, current self-adaptive approaches do not account for tactic volatility in their decision-making processes, and merely assume that tactics have static attributes. To address the limitations in current processes, we propose a Tactic Volatility Aware (TVA) solution. Our approach focuses on providing a volatility aware solution that enables the system to properly maintain requirements. Specifically, TVA utilizes a Autoregressive Integrated Moving Average Model (ARIMA) to estimate potential future values for requirements, while also using a Multiple Regression Analysis (MRA) model to make predictions of tactic latency and tactic cost at runtime. This enables the system to both better estimate the true behavior of its tactics and it allows the system to properly maintain its requirements. Using data containing real-world volatility, we demonstrate the effectiveness of using TVA with both statistical analysis methods and self-adaptive experiments. In this work, we demonstrate (I) The negative impact of not accounting for tactic volatility (II) The benefits of a ARIMA-modeling approach in monitoring system requirements (III) The effectiveness of MRA in predicting tactic volatility (IV) The overall benefits of TVA to the self-adaptive process. This work also presents the first known publicly available