关键词： 游戏产业   虚拟商品   定价策略   机器学习方法  

摘要： 2016年游戏产业收入已经达到千亿，用户规模达到6亿。整个游戏市场细分更为精细，游戏产品竞争亦愈加激烈。游戏产品收入主要取决于用户基础，用户付费率和用户人均付费值(PRAPA模型)。游戏内虚拟商品的定价对于游戏产品的用户付费率和用户人均付费值起着决定性作用。从价值决定价格来进行类比，游戏内虚拟道具商品的价值主要为用户提供的游戏体验，而游戏体验通常会取决于游戏道具的设计特征，因此游戏虚拟商品道具的价格取决于游戏产品和道具特征。市场反极好的产品说明其游戏特征和产品定价能够被市场接受。因此基于热门游戏产品特征，市场反馈和游戏内道具的道具特征和其商品价格的关系是能够被大部分市场用户所接受的。　　基于此提出一种全新的基于深度学习的价格模型deepmodelforest，该模型将产品设计特征，市场反馈和道具特征信息作为模型的训练数据集来训练模型。考虑到价格预测问题主要是一种回归问题，而现有的研究成果中深度学习模型大多基于图像识别，语义分析，语音识别等等主要解决的问题也是分类问题和判别分析，观察过kaggle网站上关于回归问题的方法发现使用决策树模型大多取得的成绩较好，因此考虑将决策树结合进深度学习框架从而提出一种新的模型deepmodelforest。在新模型deepmodelforest中以对回归问题拟合能力较强的模型树M5P算法为基础，利用bagging的思想建立N个模型树构建树层，其后连接若干个全连接层构成深度学习网络。在模型的求解方面，以反向传播算法为基础求解全连接层权重和偏置，以建立梯度残差树的方式来更新树层。在激活函数方面使用maxout方法来解决梯度饱和问题，使用均方误差作为损失函数。在解法优化方面，使用RMSprop来优化梯度求解，使用batchnormalization来优化层连接。　　整个模型的实现使用比较简洁且最近热门的Python语言进行开发，对比多种深度学习框架后，使用基于TensorFIow和theano两种backend的具有高度模块化且能够快速建模的Keras进行开发，并完成树层的自定义。最后模型对于训练集的拟合达到了极高的准确率，为了比较模型的拟合性能，文中对于相同的数据使用随机森林，梯度提升回归树，神经网络算法进行数据回归拟合，deepmode]forest相对于随机森林，梯度回归树，神经网络算法在准确率和均方误差上均有非常明显的提升。说明deepmodelforest模型对数据集的拟合相对于已有算法有所提升。限于实验环境缺少GPU和分布式的计算环境，相关实验仅做100次epoch但已经看到随着epoch增加模型的拟合准确率不断提升。

关键词： Burst   Learning   Machine   Networks   Optical   Quality   Security   Service   Switching  

摘要： The Optical Bust Switching (OBS) network has become the most promising switching technology for building the next generation of internet backbone infrastructure. However, OBS networks still face a number of security and Quality of Service (QoS) challenges, particularly from Burst Header Packet (BHP) flood attacks. In OBS, a core switch handles requests, reserving one of the unoccupied channels for incoming data bursts (DB) through BHP. An attacker can exploit this fact and send malicious BHP without the corresponding DB. If unresolved, threats such as BHP flooding attacks can result in low bandwidth utilization, limited network performance, a high burst loss rate, and eventually, DoS. In this dissertation, we focus our investigations on the network security and QoS in the presence of BHP flooding attacks. First, we proposed and developed a new security model that can be embedded into OBS core switch architecture to prevent BHP flooding attacks. The countermeasure security model allows the OBS core switch to classify the ingress nodes based on their behavior and the amount of reserved resources not being utilized. A malicious node causing a BHP flooding attack will be blocked by the developed model until the risk disappears or the malicious node redeems itself. Using our security model, we can effectively and preemptively prevent a BHP flooding attack regardless of the strength of the attacker. In the second part of this dissertation, we investigated the potential use of machine learning (ML) in countering the risk of the BHP flood attack problem. In particular, we proposed and developed a new series of rules, using the decision tree method to prevent the risk of a BHP flood attack. The proposed classification rule models were evaluated using different metrics to measure the overall performance of this approach. The experiments showed that using rules derived from the decision trees did indeed counter BHP flooding attacks, and enabled the automatic classification of

关键词： 手写数字识别   机器学习   SVM   人工神经网络  

摘要： 手写数字识别是光学数字识别技术的一个重要分支,本质上是图片分类问题,利用机器学习的方法解决手写数字识别问题已经达到非常高的精度。本文主要介绍了手写数字识别的背景、国内外研究现状、手写数字识别的一般步骤。文章第三分部对手写数字识别步骤做了重点介绍,包括支持向量机算法和人工神经网络算法。

关键词： 文本分类   机器学习   中文分词   特征向量提取  

摘要： 文本分类是处理和组织文本信息的关键技术,能够帮助有效地组织信息,快速区分有效信息和无用信息,满足用户的个性化需求。本文主要介绍了文本分类的背景、国内外的研究现状以及利用机器学习方法解决文本分类问题的一般步骤。文章第二部分对中文分词、特征向量提取、分类器训练和评估原理做了重点介绍,包括了自然语言处理的统计语言模型,机器学习的KNN、SVM、神经网络算法。

关键词： 机器学习   药物不良反应   多任务学习   多核函数学习   伽玛泊松衰减   多变量线性回归   相似度计算  

摘要： 目前,医疗安全正日益得到重视。其中,对药物不良反应信号的预测,在药物安全研究领域与新药研发领域具有重要的意义。药物不良反应每年会造成全球三分之一的医疗事故,以及上千亿美元的经济损失。因此,对药物不良反应的研究越来越受到世界各国的重视。为了监测药物不良反应,学者们基于医疗数据库提出了一些统计学模型和数据挖掘方法来监测/预测<药物-不良反应>关联关系。一方面,相关的数据挖据方法大都是基于关联规则或模式匹配算法,挖掘<药物-不良反应>的关联关系。由于这些数据挖掘算法只关注高频繁度的关联关系,因此存在两类缺点:(1)无法监测到低频的关联关系;(2)由于忽略了关联规则中混杂因子的影响,因此无法控制伴随药物对监测关联关系的影响,从而造成较高的错误率。另一方面,统计学对药物不良反应问题的建模,大都在小样本数据上基于列关联表计算<药物,不良反应>关联关系的强度,因此存在两类缺点:(1)小样本数据造成关联关系预测信号的偏差较大;(2)无法基于不同病人进行个性化预测。为解决上述问题,本文通过收集处理海量医疗数据,并基于机器学习方法针对药物不良反应监测的相关问题进行了研究。首先,本文的第一个研究问题是病人药物不良反应的个性化预测问题。药物不良反应的个性化,是指由于病人体质的差异,相同药物在不同病人体内会产生特定的不良反应,因此需要根据不同病人的特征来预测相应的药物不良反应。本文提出了一种基于病人特征相似度计算的多任务学习模型。该工作基于美国食品药品管理局公开的药物不良反应数据库FAERS,对病人信息构建特征空间,并首次提出将推荐系统中个性化推荐的方法应用在药物安全研究领域。本文基于FAERS数据提出多任务学习模型,通过计算病人与各种不良反应的关联强度,建立对应的个性化药物不良反应排序表。此外,本文原创性提出了一种新的衡量药物不良反应关联强度的验证标准HitRate@nn。通过实验表明,该模型在预测病人个性化药物不良反应问题上,有较高的准确率。其次,本文的第二个研究问题是药物多频率不良反应的预测问题。由于不同频率的不良反应造成的问题和危害不同,尤其是低频药物不良反应在临床测试阶段很难被检测出来,因此需要根据不同病人特征及药物属性来预测不同频率的不良反应。为此,本文提出了一种基于多核函数学习的多任务学习模型。该工作通过分析FAERS结构化数据,提出根据药物分子结构差异进行特征分类,并构建多核函数池。该模型通过核函数学习方法找出每一类特征对应的最优核函数,并根据不同特征构建对应的凸优化限制条件以及规则化函数,令相同特征在不同的病人不良反应预测任务中实现权重自动调整,达到预测不同频率不良反应的要求。同时,该模型将基于历史任务学习中训练的药物特征权重及不同药物特征核函数间的关系权重,用于对病人新药组合的预测任务中,可以实现对病人新药组合不同频率不良反应的预测。在验证标准HitRate@n的基础上,本文提出了一种衡量不同频率不良反应的验证标准overall-HitRate@n。实验表明,该模型在预测病人不同频率不良反应的问题上,尤其是低频药物不良反应,均优于所比较的方法。最后,本文的第三个研究问题是计算<药物,不良反应>关联关系的正阳性/因果性问题。由于小样本数据以及混杂因子(即伴随药物)的影响,造成所监测的关联关系的准确率较低,即关联关系的正阳性/因果性较弱。本文基于收集处理的4百万条FAERS病人数据,提出了一种伽玛泊松衰减多变量线性回归模型。本文使用伽玛泊松共轭先验分布对<药物,不良反应>的出现频率建模,并提出在监测单一药物不良反应关联关系时,将特征空间中其他所有药物视作伴随药物,通过建立回归模型减少混杂因子对关联关系的影响。本文对药物不良反应监测相关问题进行了深入研究,针对特定问题,给出了针对性的解决方案。本文的创新点和贡献如下:(1)针对药物不良反应个性化预测问题,本文首次提出了一种基于病人特征相似度计算的多任务学习模型;(2)针对数据挖掘方法难以监测低频药物不良反应的问题,本文提出基于多核函数多任务学习模型,通过多核函数多任务凸优化学习来预测药物不同频率的不良反应,并能有效的监测到低频药物的不良反应;(3)针对<药物,不良反应>关联关系正阳性/因果性的准确率问题,本文提出一种基于伽玛泊松衰减多变量线性回归模型,通过控制混杂因子来提高预测<药物,不良反应>关联关系的正阳性/因果性。综上所述,本文在基础理论和关键技术方面的研究成果为监测药物不良反应问题提供了新的途径。

关键词： Accelerometers   Algebra   Algorithms   Amniotes   Animal behavior   Animals   Applied mathematics   Artificial intelligence   Behavior   Biology and life sciences   Cognitive science   Computer and information sciences   Earth sciences   Electronics   Engineering and technology   Foraging   Machine learning   Machine learning algorithms   Mammals   Marine and aquatic sciences   Marine biology   Marine mammals   Mathematics   Neuroscience   Organisms   Physical sciences   Polynomials   Research and analysis methods   Research Article   Sea lions   Seals   Simulation and modeling   Vertebrates   Zoology  

摘要： Constructing activity budgets for marine animals when they are at sea and cannot be directly observed is challenging, but recent advances in bio-logging technology offer solutions to this problem. Accelerometers can potentially identify a wide range of behaviours for animals based on unique patterns of acceleration. However, when analysing data derived from accelerometers, there are many statistical techniques available which when applied to different data sets produce different classification accuracies. We investigated a selection of supervised machine learning methods for interpreting behavioural data from captive otariids (fur seals and sea lions). We conducted controlled experiments with 12 seals, where their behaviours were filmed while they were wearing 3-axis accelerometers. From video we identified 26 behaviours that could be grouped into one of four categories (foraging, resting, travelling and grooming) representing key behaviour states for wild seals. We used data from 10 seals to train four predictive classification models: stochastic gradient boosting (GBM), random forests, support vector machine using four different kernels and a baseline model: penalised logistic regression. We then took the best parameters from each model and cross-validated the results on the two seals unseen so far. We also investigated the influence of feature statistics (describing some characteristic of the seal), testing the models both with and without these. Cross-validation accuracies were lower than training accuracy, but the SVM with a polynomial kernel was still able to classify seal behaviour with high accuracy (>70%). Adding feature statistics improved accuracies across all models tested. Most categories of behaviour-resting, grooming and feeding were all predicted with reasonable accuracy (52-81%) by the SVM while travelling was poorly categorised (31-41%). These results show that model selection is important when classifying behaviour and that by using animal charact

摘要： The Dishevelled/EGL-10/Pleckstrin (DEP) domain-containing (DEPDC) proteins have seven members. However, whether this superfamily can be distinguished from other proteins based only on the amino acid sequences, remains unknown. Here, we describe a computational method to segregate DEPDCs and non-DEPDCs. First, we examined the Pfam numbers of the known DEPDCs and used the longest sequences for each Pfam to construct a phylogenetic tree. Subsequently, we extracted 188-dimensional (188D) and 20D features of DEPDCs and non-DEPDCs and classified them with random forest classifier. We also mined the motifs of human DEPDCs to find the related domains. Finally, we designed experimental verification methods of human DEPDC expression at the mRNA level in hepatocellular carcinoma (HCC) and adjacent normal tissues. The phylogenetic analysis showed that the DEPDCs superfamily can be divided into three clusters. Moreover, the 188D and 20D features can both be used to effectively distinguish the two protein types. Motif analysis revealed that the DEP and RhoGAP domain was common in human DEPDCs, human HCC and the adjacent tissues that widely expressed DEPDCs. However, their regulation was not identical. In conclusion, we successfully constructed a binary classifier for DEPDCs and experimentally verified their expression in human HCC tissues.

关键词： ocean   eddy   machine learning   eddy lifetime   remote sensing  

摘要： We report a novel altimetry-based machine learning approach for eddy identification and characterization. The machine learning models use daily maps of geostrophic velocity anomalies and are trained according to the phase angle between the zonal and meridional components at each grid point. The trained models are then used to identify the corresponding eddy phase patterns and to predict the lifetime of a detected eddy structure. The performance of the proposed method is examined at two dynamically different regions to demonstrate its robust behavior and region independency.

关键词： Ferry traffic   Traffic prediction   Machine learning   Periodic analysis   Algorithm composition  

摘要： Ferry service plays an important role in several cities with waterfront areas. Transportation authorities often need to forecast volumes of vehicular traffic in queues waiting to board ships at ferry terminals to ensure sufficient capacity and establish schedules that meet demand. Several previous studies have developed models for long-term vehicle queue length prediction at ferry terminals using terminal operation data. Few studies, however, have been undertaken for short-term vehicular queue length prediction. In this study, machine learning methods including the artificial neural network (ANN) and support vector machine (SVM) are applied to predict vehicle waiting queue lengths at ferry terminals. Through time series analysis, the existence of a periodic queue-length pattern is established. Hence, methodologies used in this study take into account periodic features of vehicle queue data at terminals for prediction. To further consider the cyclical characteristics of vehicle queue data at ferry terminals, a prediction approach is proposed to decompose vehicle waiting queue length into two components: a periodic part and a dynamic part. A trigonometric regression function is introduced to capture the periodic component, and the dynamic part is modeled by SVM and ANN models. Moreover, an assembly technique for combining SVM and ANN models is proposed to aggregate multiple prediction models and in turn achieve better results than could be attained from a lone predictive method. The prediction results suggest that for multi-step ahead vehicle queue length prediction at ferry terminals, the ensemble model outperforms the separate prediction models and the hybrid models, especially as prediction step size increases. This research has important practical significance to both traffic service management interests and the travelers in cities along waterfront areas.

关键词： Materials informatics   Density functional theory   Regression  

摘要： The design of new and improved materials for different applications of interest is boosted by combining computations or experiments with machine learning techniques. Materials scientists seek to use learning algorithms that can easily and efficiently be applied to their data in order to obtain quantitative property prediction models. Here, we utilize a first principles generated dataset of the electronic and dielectric properties of a chemical space of polymers to test different kinds of regression algorithms used by the machine learning community today. We explore several possibilities for the hyper-parameters that go into such learning schemes, and establish optimal strategies and parameters for high-fidelity polymer dielectrics property prediction models. Published by Elsevier B.V.