1.深度学习的引入
组合低层特征,形成了更加抽象的高层特征。
表达式中的u,w参数需要在训练中通过反向传播多次迭代调整,使得整体的分类误差最小。
深度学习网络往往包含多个中间层(隐藏层),且网络结构要更复杂一些。
2.数据集及其拆分
Iris(鸢尾花)数据集
分类特征:花萼和花瓣的宽度和长度
数据集在数学上通常表示为{(x1,y1),(x2,y2),...,(xi,yi),...,(xm,ym)},其中xi为样本特征。由于样本(即一行)一般有多个特征,因而![]()
,而yi表示样本i的类别标签。
![](https://image.shishitao.com:8440/ZGF0YTppbWFnZS9wbmc7YmFzZTY0LGlWQk9SdzBLR2dvQUFBQU5TVWhFVWdBQUFKOEFBQUFYQ0FZQUFBQVYrSjFWQUFBUWlFbEVRVlJvZ2UxYWQxUlVaeGEvekF3b1JRRXJJSUlWSzVhSVNrekVHTUVTZFdPeTJWVE5Kc0gyeHhhTmJKcUptazJpTVhHamJuSk8zQmo3V2phSktZSkdWNHhkak5rVkZSUlVMRmhSYWRJVm1MZjNkMmUrNGMwd2cyaEFrM1AyZDg0N1BONzc1aXYzdSsxM3YrZW1NZWovK0QvdUFVejNlZ0szRFROZmhqdHZkKzNxTmFxb3FLREFvTUE2bnBobHpLdFhyMUxXbFN3cUxTMGxQejgvYXRldUhibDd1TmY5V0hVTnE3d3F5aXZJWkRSWlpNZlBsRzl5TTdyWk5kY3FOZHN6L1gxdHg1RSt6UlZtcmRZLy9LV2d0Z3FvUTA1MkR1M2J0NCtXTEZsQ2p6d3lraVpPbUhEYmZkd0tpVnNUYWVYS2xSUVlHRVFsSmNXVWxwWkdzYkd4OU16VHo5VDVXTDlVbk1vNFJZbUppYUswbFJXVlpEQWF5TTNOalc3Y3VFSHQyN2VuVVNOSDJXUmh3SXRmSFdxNWtiQklvTGlvbUw3ZjlEMGRQSGlRem1XZW83eThYS28wbSszYTFBVjI3dHhKbnA2ZU5IUG1UUHJvbzQvSTE5ZVB2dmppQ3lvcUxxcXpNZW9ORm5GUVlVRWhuVGw5aHZKeTgxbDV6SFR4d2tVNm4zbWVpZ3FMN05ycG9aZmhsczFiYU8yYXRlUnVjcWR2di91T1B2LzhjN25mdm4wN3JWdTdUcFJRd1lSL0duZzJxTTlsMVExNDBhZlBuQlp2NHVYcFJSRjlJNmhSNDBZMS9rUVpGc0xJMEtGRDVmN1lzV05rcmpTVDBXQ3dhMU1YaUpzV1J5WjNFM2w1ZThyL0xWdTBJRE43Z0YrTGdlZm01Rko4ZkFKdDI1WklJU0VoOVBoamo5T2V2WHRvMDhaTjlNUVRUOUQ0aWVQRjhGT09wRkk2NzBOSWFBaDE3OTZkdkgyOFpYL0tLOHJKMTgrWDNwb3hnNktpQnRMdVBidXBkKzllRkRzaGxqcDA3RUNYTDJmWjZacko1RjdMZk9RT1F0M1BoblhNOHB2bDlLOTEvNkl0VzdaUWFKdFFYbEJ2dW5uejVxM25abjJHQmJmd2JDSFdySmsxQ1FXT2Jlb0N2djYrdHZzZFAreWcvT3ZYYWVyVUtaYk4rUVZBZVNoWCtSdENaWmN1blNVeUxGKytndTY3N3o1Njdybm5LQ1VsaFg1Z3ovWGlTeTl4VGxndTc5UFMwK2c3OW14QnJZTG9sVmRlb1diTm0wbGZqN0hDZXJIM1R6NlVUT2ZQbjZlWW1CaDVQbURBQURJWWpIYmptZ3pPck5KeE02MkpweHZkWlF1MnpnSGU2dThmLzUyaW82UHA5ZGRmdDFtYVl6c0ZWd213OG5ZL0MwbzJOUmdqUEVOOFFqeE5uanlKK3ZicisvUEhyQXVZZFVybk1IZFJQSlpaMDJaTjVWcTllZzBGTVNFYlBQaGhrVmxSVVJIL0gwUkdrNEd2QmhRMUtJcWlCa2JSeGswYktTN3VMOVNvVVNQMmRtOUpoRkZrQmVsTkRudlM4UEJ3R2NNWjZiTGtmSTV4WERjeGVCMzhmeTlKeWQ0OWV5bWJDY1BRbUtGVlhzVEZ4aXZGS3lzcG81czNidHE5UTU2SDBLRFBPNXpsTURYQ09xNGxvYTc2c2JwUE81WkdHemNtME5peFkyV1RicFRlY05iTDNZZkJ4YjBWYW4rdjUxMm4xTlFVdXYvKys4bS9pWjk0cjVNblQ0cm5jdXdQYllLRGcybGI0alpiSDZxZjR5ZU9VMGpyMWt5K1hGY1ZERUt2S3l0azBNT0hEc3VGZTNIUkxNKzgzRHg1VmxKY1dxZkplVzJnTnZRNmg2OXlEck1OUFJ2YXZjZDhzaTVmb1FNL0hxQ01FeGsyNjBhNTRHem1XU2w3NkJXd2doVVA0ZG91Wk5jU1IxT08wbjkrK3ErUUZ3RGpYTGh3UWJ3Y2tuRjRCY3hsOW51enFXRkR5enkzL25zcnpmMWdyakR0ZXczc0grYUtDL2NBNUFkQ2dmMVY2enAxNmhTL0w1SFVCa0I0dlhidEdnVUdCTkxWckt2eVRHOTBKcE5SOGxwbHhPZ1RobithU1V0bkR1SE5tbHJDc2J4M3VLVE9aMmFQZ09Rd0tTbUpCZndmR2pWNk5QM3hUMzhRSzM3NzdiZnA4T0VqdEhqeFlucHc0QU4yQzhJa0VMYWRlVVhrUFB0LzNFOGU3aDdzY1NwZENzV2RjODRubjN4U2NnYzlzQWhzS0JBUUdFQisvbjRpQkQzT3NWVW1idDBxNDV3K2Zacm16djFBaFBiTjE5L1FyRm16YU1BREEyaisvUG5rMGNDRHJmTUgycjFyRngxUFAwNkZyQ3p6Rjh5bmtVejdPM2JvVUdNNkFjWGF2SG16RUIwazN1UEhqNmNYWG54UlVnR2tBRkRBRlN1V1U4OWVQZW5yOWVzbEZPM1lzWVBtekpsRDVad2ZvYnd3QVdXZHV3MWRhSVgzM2NUek9uNzhCRzNZc0VIS1RDQUJ5Y25KSXFkTGx5L1Jva1gvWUJMWFJ6eWRrUlVxUEx5SC9OYmZ2d2sxYjk2Y2M3NGZPUHBrMDRoSFJ0aENLSnhDYVVrcEJiUnNLV09wcUFQWkpDWHRvMGNmSFdQUkRSZlJSWlFQb2JjRGIwTC8vdjNwM1hmZm8zWHIxdElnRGhuTGx5Mm54bzBiMDVoSEg2WFc3RUlkb1pUREdSS1pNYTFldlpvcUt5dEZBZTNrd3BhaWFXWlJlaXdPTHQxUitmUUtqVUt0Z1hNUEtKZ2U3c3dzaDBRUGtienFoZCsvd0hTZUJjUUtpckVIUmtWUjU4NmR5Sk9aTWVERDRicHpseTcwNGJ4NUVuWnJtLzlCTm1GaFlSUVRIVU1aR1JtaVhCanZzODgrbzdDT0hTbVVXU0ZrQk1HcjUwYk9lOHhzY1BodEUxNGZycnNPUGFkaWdnVWpHRFpzR0dWeVJQajZtNjhwTWpLU2xpNWJTbTNhdHFHT1lid09acTRBakdqMjdOa1NNb0hJeVA0MDc4TjVCR0lhMGFlUFhlNVdXRmdvYVV6WHJsM2xmK3daUE9pbTd6ZFJTVWtKZTlVTEVrV0ZpRGxUUU03cE5EMVdMbCtwZGUzY1ZaczBjYksyY01GQ3JiUzRWT1BRcGQwb3V5SHZMMS9LMGpaOHQwRmpkNjNWaE55Y1BPMzh1ZlBTN3V5WlRMdnIzTmx6dHV2SzVTdGFSWG1sMHo1UUFNZTRxMWFzMGlMN1JXb0o4UWxPMjJWZnk5Ykd4NDdYQmo4MFdPYWR1SFdiaG5YeHdqWE5lZGQyWTdocTR6aXYxZjljclhVSzY2VEZ2aGdyc21HdnFCVVdGRHB0VzIyTStzQXQxdVpzSGxoRGNGQ3c5dnk0NTdXNTc4L1ZDcTRYYVBtNStjN25XT2t3aHNQL0Y4NWYwSDczMjk5cFUvNDB4U1lIOUFNZHlUaVpJZnRyNjdleStrV09DK0g4U1F2ckdLWTk4L1N6b2tDTzJMbGpwemJrNFNIYS9xVDl0N2Z5T3dBbXZ2NnI5ZHJRNktIYXpMZG1pcEFBdGRGNHJ4YjMvcHozTlc4dkgyM3hQeGJiSzRMMWRzM3FOZHFDK1Fzc0NxbnJ2NXBROU84YzdwUDJKV2wrdm42aTZFclllaVFmVEpaNVF2Q084NnNYM0lieUtaa2MvTzlCV1FPVVR5OEwxUitNSGNyRGVaczgwaHVudXRjN3JQZ044ZHBEZ3g3UzRxYkZpYk9wY2E0T2wrVnNWM2V1ZDR6elBNUnlVRzJ3SFR3VGwrcG1rRERiaDEwdlFrdFFZSkFUUDJvQkVzNzA0K21TUDZDNkRkZnNDSVJpaEQrUWlJaUlDR3JTMUhsbzZ0NDlYTno2Q1daY0Z5OWRGQmV1eWtQQzFIbmVTSVF6ejJaU2t5YisxSzFiTjVtbklrY0k4UWl4VjY1Y29XSk9wRDI5UEcxOXl6dUhmQlY1TEg2dkQvdnEvbWpxVWZuYnBrMWI4dkgySVNVN1RiTXk3TEl5eVovMFJlWGFWZ25VdUZXVEk3dlFLV2V1N3RXUDRsR05nSXh2VmE5RTN4Z0RhOEI4SWRQR2ZvM3RHM0VmeVBzLy9mUlRtajc5VFFydjBWM1dndmFrUzl1bG5FS1cwNUJEaHc1Ukk1OUdGRDBrbXVYZlZKN1g2cXlYeHpJcDVtS3VxSlNTeHI1OWU2a0w1MFpJTHBHbzJpclMzR3pQN3IyME1TR0JpY2VEMUs1OU85ZjljbzZ4ZHUwNldybHloWnp2ZVhoVXova3FtV0hqcEFGS3QyVEowbXBrUmlHc1UwZDY3UEhINk9XWFg2WTlQTCt1WGJyYUdDMVFWbHhHYTlhdW9aTFNFbXJKaVM4b1B2SVpKVFRrWGtWNWhmVHNNODlLYnFiUFdWVGVCeGxnUGhVOFYzVTZvWlJYRE1TcklXM2Z0cDJPcEtRSW9VbmpoUHBtK1UwaE1sS0dJa3RwSnpRMFZJaUd5dkdxbld5d0RLR2dXTDhheHpZWHErSWhSd0piaHR5aFdFVldGZ3BIQUlDcFlxNVFmdlN2WDQ5aXJDaEhpUUpZWlFCMmovNUF6bmJ1MmtsOSsvYWpBd2QrY2xxckRBZ0k0SHgvRVA5dGFkMU1Fc1d6S2FFT0lGc28vQThmUG9LR2pSaG1lMTViZ3pNVkZCVFErcSsrb3B6Y0hQRWVvMGFOSW5hZnRHclZLanB6OW94UWJDZ1Rqckt3V1VoV3dZWkdqaHJwc2xOWTRxUkpFMm5FOE9Ia1puQ1RVd1U5OEF5THdXVTBHS2xIeng1TysxR0x3QWFBT2RxRXdRQnJSZklQQmd5aFRKMDZsV2E4TlVNK0hnRExLaXNycFdiTm1vbUNvMEI5aEJuN2pKa3p4WnB0c1BhRk1zNzBONmVUcjY4dlRaczJUVFlQREIyZTV1TlBQaGJTQkxZN1ljSkVpUWpMbGkyalM1Y3V5VEVmRkFVZVpOZnVYYlJ3NFVKbTdrL1I4K1BHT1Uyd3MzT3lLUzR1VGhUNHoxUCtYTzM5VDZ3UWI3d3huWjUrNmlsaG8zQUFreWRObHZwa3dzWjRhZk14cndVbkN4Z0xCRWQ1R1pTVVlLQ3RXcldpMTE1N3pXSVlCS2VpMGJ5Ly9VMUtJaWxIVXBpcHZ5QW5Gb3NXTGFMTFdaZEZkZzBhTkpEaU1zNTBFZlhHakJsRHpWczByOW9IYTRRaGg2SUY5Z1JPQUI5UldBYWpXM3BnUFV5d0lyQ1R0TFIwbXNhVEgvMmIwVklMZ2pmQjRmamd3WVBGazBENWNHN1h1Vk5uV3cyb0dxeURReGp3akRWNXg5dkIyVE5ucER6U1NzZUlFZDZXTGwwcVJjNTMvdnFPc0RRdzNJVDRlRnJFWWFNVE0xMGM3V0RlS0JYQUk1cjFKUitkb0tCbysvZnZsNW9VUEJPVUQ0cUhNZ3RLRTFsWldhTFk4TTRlN0dudzdKTlBQcUdPekhiQklLRjhZSk9abVpuQzRtM2xCWWVOd0dhaGJPTUtPRnNGWTBmeEZvRFhSUm5KNWtHNXo1TW5UbExTL2lSUkVrQVpLSXdNSVJQeU1PaVlQUHJZdlBsN1NZRWdwK2lZYVBMejg2Y3Z2L3hTV0cydm5yMW8yUEJoMG5kcWFxcVVqOGF4OGJ6NjJxdFZFM09oVURCdUwyOHZPbnIwNkswVnoxRVc3TlZOMFBvcFU2YklFUW55SlFEZVljR0NoWFR1WEtia1hDRWhvZkljK1VKQllZRW9vSUpkZksrbnMxOTREQmlKUHd0Tm9WZXZYc1NNazlxMmF5ZWhHVUQ5cWlkNzBYSzJ4djc5KzlrK1BNalB6NWQxdEFsdDQ3Ui9lSCtFcGo0UmZVUU9DamlMbkRQbmZmRWF2WHRaREE2RzkrRUhINHJYNk5tanA1UXFnT0ppaS9WakhGZUFjaUVxOU9rVDRmUjk5L0R1NHRXZ3lFQkF5d0Qyc3N0bGZNdUVTRHo4Nk5HanFXK0UvYkVkenBGaE9BTWZIR2lYRzhKWXNBYkk3NEVITEtsTmp4N2g0bGd1WHJ3b09YeHc2MkI1anBLU3Y3OC9wdzlWY3RMdkw0d0FYbGlsSzBpbm9EKzV1WGt1MSt3SzhMZ21iQkNPZ1J5QmdpTXVRT1dGeDlLT2taZVhseFI5WFIxUzF3ZWFjaUtMUlpheWNCVVFGdlI1QnVEdDQwTXhRMlBzbmlGSFFzaHN6MTVZVHpiMGhvTHY4TEJwVDNHNFUrRUtsb204VE9XaUlnT3pKYVVZOEdEVlVaTWlDbnYzN2hVUEc4aXljUVVVbjVHL1BzNlhNN1FLYmlXWEFuTE5SMGFPa0h1bEJOM0N1OG5sQ0JTUnNZWnh6NCt6ZXc1RjFPZlRtQy9XR0hsL1pMVStRTXJ5OHZLa2ZnbFVJemtHKy9OeEtCQVUzdWNPUHB6QUhLcDZxdUdNRThKRlRwR2ZsODg1amljem5NT1NEOTZ0VDRWQWNGQ01SdFZjZjdUamJKNDJXSnRkWXV0R2NiaHQyN1pDS0VBTUhJR2M4NVZYWDVWUXJZQkVIbU9wWTBhRFNyaDFRNmhUR1B4TlRVbWxUaHdSNEEyY0hVT2lyMzdzamZHdG4xN0Jxa0VkUCtuV0FOaEN1ZGxKT3lMNXRPazlEcVBLaXpsckE0aU1IUHBSTWdWekJXbkNsME9BWXJYVjVtZjlpL2JRQ2FRZWR3TGpyQm16WnNucFVrMTZ4QU1aalVZcTVzUVNuNkdEb1lJUkkrRUdlYWhYbUMyNUJjTGlnUjkvbEZBQkF0T1U4N01hUDArM1RndVdqQTg2dmIyOTJYdDZ5S2MvRFJzMEZJWmF5WG1TbWNrRVRsZWtkTUk2Zzc3Vm10dzA2NmE3V1VtU29TcjN3alBrY0NCZmVQN3R0OTh5UWJnbW44Nmo5S0QvdkFvd2NCdVFBVVNPR3FIZkM3MW9WVTdsNW5CUEZ1VnBIUkpNUG8xOGJITnpiS1BhR2F3RTBCYXgzQ3h6UTJWanhZb1ZRcnFpb2diSjZZWGQ5NUptc2hISGt5Y3phUE9XelN6WEwrWDg5cVdYWXNtUGYzZTdIejBaWjczTnluY3JXQVdDMmhDU2VKejdZYUgxcm5qV3NjRzJlM09PaC9HUk40RWc0TWhOaFVpYndKM0FxNkVYYjB4cjJYVGt0TGlYZVZ1RmJ2dTJ6N3BHL1pwY3JzLzYyR2lxK2o2dGJadDJjdHdIRDRzTmdiSFdLZHhjM0pObEhkWGVPWm02YXVkc1hTaUpKU2Nmb2xNY0plQzlXN2NPWVVPeUdwQWlrZ1kzWWY3cDZlbnl5ZFR3NGNOcDdMaXgxS0psaTl0V1BKbUg1bGk4K1lXaVdoRVd1QlhEdWhjZndQNUtBZm5tWkdkTFZRRGx0UTVoSGVwOXpQOEI4YWQ3UTcvTE9UQUFBQUFBU1VWT1JLNUNZSUk9.jpg?w=700&webp=1)
ground truth:翻译为地面实况。机器学习领域一般用于表示真实值、标准答案等,表示通过直接观察收集到的真实结果。
gold standard:金标准,医学上一般指诊断疾病公认的最可靠的方法。
机器学习领域更倾向于使用ground truth,如果用gold standard则表示可以很好地代表ground truth。
数据集与有监督学习:
有监督学习中数据通常分成训练集、测试集两部分。
训练集( training set)用来训练模型,即被用来学习得到系统的参数取值。
测试集( testing set)用于最终报告模型的评价结果,因此在训练阶段测试集中的样本应该是不可见的。
对训练集做进一步划分为训练集和验证集 validation set)。验证集与测试集类似,也是用于评估模型的性能。区别是验证集主要用于模型选择和调整超参数,因而一般不用于报告最终结果。
训练集测试集拆分:
- 留出法( Hold-out Method)数据拆分步骤
1.将数据随机分为两组,一组做为训练集,一组做为测试集
2.利用训练集训练分类器,然后利用测试集评估模型,记录最后的分类准确率为此分类器的性能指标
- K折交叉验证
过程:
1.数据集被分成K份(K通常取5或者10)
2.不重复地每次取其中一份做测试集,用其他K1份做训练集训练,这样会得到K个评价模型
3.将上述步骤2中的K次评价的性能均值作为最后评价结果
K折交叉验证的上述做法有助于提高评估结果的稳定性
- 分层抽样策略(Stratified k-fold)
将数据集划分成k份,特点在于,划分的k份中,每一份内各个类别数据的比例和原始数据集中各个类别的比例相同。
K折交叉验证的应用-用网格搜索来调超参数
什么是超参数?指在学习过程之前需要设置其值的一些变量,而不是通过训练得到的参数数据。如深度学习中的学习速率等就是超参数。
什么是网格搜索?
- 假设模型中有2个超参数:A和B。A的可能取值为{a1,a2,a3},B的可能取值为连续的,如在区间[0-1]。由于B值为连续,通常进行离散化,如变为{0,0.25,0.5,0.75,1.0}
- 如果使用网格搜索,就是尝试各种可能的(A,B)对值,找到能使的模型取得最高性能的(A,B)值对。
![](https://image.shishitao.com:8440/ZGF0YTppbWFnZS9wbmc7YmFzZTY0LGlWQk9SdzBLR2dvQUFBQU5TVWhFVWdBQUFTY0FBQUJnQ0FZQUFBQmZjTFFoQUFBUCtFbEVRVlI0bk8yZEMxUlZaUmJITjFjZWlqd1Vnb1JRMFVZS0pQTlZ5c1BSMVFvZlNUYU9vZGFhV28wcDB6RFdpS1N1c2t5c3FVd05UWE9OUWpWak0rUDR5SnlpY3RReVgxd3NRUlFmcVJOaUlKaUdDQ2p5WnM3KzhGd2U5OXdMWFBLZWI5UCtyWFhXUGZlY2krN3ZuTy83ZisrOUhlb1ZnR0VZUmpJTWVodkFNQXlqQllzVHd6QlN3dUxFTUl5VXNEZ3hEQ01sTEU0TXcwZ0ppeFBETUZMQzRzUXdqSlN3T0RFTUl5VXNUZ3pUZ3VORkJlSmc5TVZSYndNWStxUVY1SUN4TUFjR2V2dkIrTUNCZXB0ak15aEk4Vjl2Z1JOWENzWDNnVjUra0RRbUJrSzkvWFcyekhZMm5jbUEvTEppR0JjWVFpNGREckpzWDRuZGZBWlMwaStDd1FGZzR4UEJFRFBZeDNTdnBQSUdyTWpjRFNuSEQ0cnZBVzQ5SVdIWWd6QXRhSmhlNW5hSTVPd0RrS3lrSmYrYWttbjZoc0RLMFRIZzZkSk5iN1BhRGI2WFIxUFhtd296NHVIY0ZiWkd4NUlyQ0hsS0FSN3g3NlZtMXpFOWg2WXZJUGQrVUdqeDNaUldWWml1b2RqaXU2R1NGbW02ZFg3dXpsRDM5cS9oM1NrRFlOcUdVNmJyYWdGUWhRbkJRaDIvZDR1b0ZhZ3hSNm1aWDBsUEZXbEEvbnYrcENnVVdEaW84WW94dFprd0lWZ1ladXo4VUNlTGJHZEZ4bTdONjVnZVRDYzE4QjAwRlNZRTN4V2x0T2pTclRQTTNRZmRuUTF3dmFvT1Bwc1ZDaE9DdlNCeFFxQzRWMUZUQi80ZXpxYmZvaWkxTEFBcUtGQzlsVlpVdUg5L2U1amRZYkQ3cy9tc3VhQmlKcHFqcE9VanBWYWp4STd6SnpTdm8vQml4WUh2aGdvbnJJd3g0VDE4ZDFUSVU1Ni9Xdm0xcE9HZHhkalhJQnZSUlp5cWw0K0NMa3IvN1oxOUYyRDEvZ3RDbk9LM2Z3K3JsTzhQQnZXQWJUTkNUTDl0clhXRTNUMWpJUkZ4S3JTY3dYSE14bEx0TFNQWTBtdFpNemZGcUJUbWZIYzY0bFJpSlMxNHoyamwzY21HdGNGOGErOU1OblFScDNIcnN1R3JzMWZGZVdRL0QvR1o5SnM3eGJIemRER01YSmtsdW5pSXA5TG56N2Z5YjRYNzlSZmpUeVRJMkcwMWs1TkpCelIwdDdWYWdTcFRnNGFSYWRFaVdIRllhbTMwZHU5Sjd0MEViMGpVdkJkQXFEVnI5ekduVDQ0WFFlNlZDaUUrNzAwUE1ycy85cTdtRDI5cUs0UGVyZDJYQ1p3eHNjVFVBWFRTZ2VDZ0tnN21hNEdEeUpTRUNSbHZJUzBJcFR5RzRMdVpHUnFoZVc5SldMU2RyYkVkdTR2VC9YM2NJYWVvUW1rZEhZRTN2OHdUMTBwdTFJaHhLSitYMDhUbkk2SGVwdDlqeHNCWkJpMlNSc2VJV28wS09JT2xKVUpZbUNuVnpDcG9NOXJla3NTd2gzV3dwbU5nUGd2ek14ZFV6SHNVWjRXWEtPOGdZV2hqbnNMM2hPV0YwbElQM1pZU0ZKZFhRMDlYcDFhdklkaE0zWEgrcEZpdmdXTlEwMjVtSkdxMXM4cU8zQk53b3FoUWRDV3d4c2FDUVdWNnR5VTQ5b1Rwd2FVUitGNG9ycWRwQ3VZdjQ4M0I3ekFsZjFFVXBwYmdXQ2JGeWsrYWRVNXRaVXJxZW5LeldwYWdtbW0wNkV6dlJaMlpvMXI1dFlScVBwTm1uUlBETUV4VFdKd1locEVTRmllR1lhU0V4WWxoR0NsaGNXSVlSa3BZbkJpR2tSSVdKNFpocElURmlXRVlLV0Z4WWhoR1NsaWNHSWFSa2c2SlUyWm1KZ3dkT2hRKy92ampuOHNlaG1FWVFZZkVhY21TSmVEcjZ3dUppZHErWTlySzFSczFjTyt5RE9HUjRMNjNNenYwYnpIMkIvZWlxUnVBS1lNYnpOR044b3hkRzhTQjV4VGRKemNGTnpMakJuT0swV1E2dFBIWFlEREE0Y09IWWZqdzRWQlhWMmV6RWY4NS9oTTRHaHhnWW9nMytMNXNoQmtqYm9jM281dHZ1c1JNc2xsNTBQaXdkLzMyT2JLNytCSFZ5d0ttSnpFc211d3UvczRXNENCcTJ5b3pUNUVjNEVBL2JHNDViZCsrWFh4aXR3NVp1WEpsbS83dWpkMC9pQlpTMTNuN3dYWCtBWEh0a2REYmhEQWhRVDdkb0x5cXVkQmhEWVpCQU5BbEwzb3JSQzkvRklNYklHZzMyby8rejlFcjV0aHQ3NGdkL1ZqUXFhSGwzeDBMQTRaWG9rYnk4UU9hTG16eFd0UGdHbFRBUUFaYUFRNndERkhCWm5IQ0xsMXNiSU9Makprelo0cnZiZUhwRWIyRUY4eUtaYVBFOTJNRjEwejN5aXBxSUMyM0ZPWS8wTnQwRFF1emxqdFlMTnpVbXR4b0w5cmRFaFFwaWdVQUM3UVdXQWlvZFNQUXY1WWxyUGwrbHhGODlwYmNRVzhtVktuYjdFTThLeXNMTm03Y0NLZFBuNGFFaEFSSVNVbUJrcElTOFBUMHRQcDNsNjVWdzkxdkhoYmpURWpUVnBMdklpTXNtOVFmQW5xNG1LNVpHOGZBZTdQdWliUTFDWGJIV2xxd29GUHl1WVBqVE5hYzVhZGtIeVRscGRSYVJZZjNLQVdmc0NhbW5UN0F3ZXJWcThWbmNIQndzK3ZZZWxxeFlvWFZ2MzNzdzFQdzE1Z0JNSFd3ajNETHE5THYxVU1RRitFUENXTUNtdjNlV2tBQUhMZWg5TEJieXpTVUNrQkpsZlZ1S0FjNDBJOXhTc3NKaHd1MDBIS3JMQ3MyaVJPSzBEUFBQQU5yMTY0MVhZdVBqNGVrcEtSV3hTazZ4QnVtYnpnRmNWdlB3c0JlM2NXMStaL2t3UG5pU2tqYWUwRWNpQnA5Wlh6ZmdSYWpmR0FCb09SR05RQmR3Rm9RS0RXS01TVjI1SjYwV3FBcGdUSDJqRmJ1VVFJbkl6QS9hYjBiTEU5VWFQZHNYWGw1T2JpNXVVRnViaTcwNmRQSGRMMm9xQWg4Zkh6ZzJMRmpFQm9hYXZYZnFLMnJGM0hyMm9LbFdSUnFNdzhxVVIrdDBnd1N1blZpTEttV0JvTGQxQm03ektQN1loQ0hsV05vQkc1VTRkazYrU0RoUXh3Zk5NNEFxWVVhZ3h1OEgvVUVtWWZjRkxHV1p1OFdFWVljd2N6L2Z0U1Q1SVJKQlFVSzA2TVdBZ3hKdElSZzlCV2s1ZElJTE16dmozMlNYQ3V3S1NMS3ROSmluM2xQQkxubEhTVEVxU21keVpFK1ZjZnpXblNtOThJQkR1U0E5OVl4RENNbExFNE13MGdKaXhQRE1GTEM0c1F3akpTd09ERU1JeVVzVGd6RFNBbUxFOE13VXNMaXhEQ01sTEE0TVF3akpTeE9ETU5JQ1lzVHd6QlNZck96dVorYjBvb2FDRitWQlpOQ3ZlSDFpZjMwTm9mNWhZSWJtWk52ZWlVdHJZcUE4WUYwWEl4ME5xUVFwNHVsVmVDL09CMm1ETHBOYjFPWWRxSUduamhSVkFDTGpKOEtIMXZVZHIrcjRBYlpwajYyMGZjV1pTOEx5ZGtIaE5DaVh5ZjA3SUcrenlpSnJkMjlFbUNBZzRXZjU0SnpGd2N3T0RoQStWdU5iblpqL25ZU0J2aDBNMnM1NFlORnQ2K3EwemwwcElVUG11Sk9hM1RMZ1FWQWRkUTJybThJekFxTkpMa0RYc3RuRUpJME9vYVVFMEFFZmRWcitYZEgwRDBQcFVLTllFV2g1WmVlVWxyc1B1WmtMY0NCRmxnell3Rm82ZzBUQ3pVVzhEa0VvM3hnV2xKdTFtWUkrblY2OUxQMUpqY2RsTUNXaHBhYjVGZVVna0VOYTQ3L2s0a0ZuOEFLMEZMQURFcXVvTzNlcmJNVzRFQUxTd1VBUWNFS1Uxb2NWR3BwckoyMXZHQWk2TER0bThjVzJOa2kyOEVDb0RyTWF3bStMNnc0S0RscDYwd0JEcXlseFZMK2t4RzdpNU9sQUFlVzJISGVlaFJaekRSVXd0MVl5elRZa2tLSGJWUW9iU1hPM2tCdlArV2dNL2FFQStINUZ1NTVPbmNWM2xmcFFLOFZyb1hkeFVrcndBSE8xSGt0YkJTcXQ3N0tnNW9WRFFFT01GTllxcUdSV2FFUlpNSkQ0UURsSyttcG12ZlFYUzgxVDVKMy8zMnh4Vll0dG1ncERZeWprRnBxVmVBOVNtT0NXREZZQ2hLS3JvZXBZSGR4ZWlPNkg3ejJVS0JaZ0FOVmpGcUNnM2VXeEFrTDlGUWlYVG9FMDJKSm5IQlFuQnFXSXVQZ2hBVWxZVUl3SDFtSzhrTnQ0Z1Y5NnllR1BXdzJ3SS9sSllsUTRJa3VpeFhzL1ovaUxGMWJ3VXlPM2FHV3RSbythSXdrMGNmZDYrYzI3NWFCbVNiQXZhZVoyR0xyTURFc0dybzZPdWxrbVcyZzNWL25uWUhMTnhvbk5mQzkvSFBDRFBCMWRkZlJzdmFENDJPWUhveGZwN1k0VUdSWGpvNkJZYmYzYWVXdjVRUExqZG9WOVZEeUhZN0x2bmovZUZLVkJwa0FCK3A2R2dRTCtQaStJU1NqcnlBNG1Jd0JRZk9WTkkwTERDR1ZZYlRBbVVaY0U0VGRDU3JUMUl6OGtCRW5obUYrV2ZEZU9vYnBCQ1R0elFmRDNIMFc3dzk2NjdDNEg1MTgzSTVXZFF3V0o0WWh6dkMzTXlIdFhLbkYrNE9YWjhEVUlUNWk4Zk94d3V1UWtrNWpyWk1VZSt0c0JmZmtSYTdPZ3B5aUNvaUw4SU0xVXdib2JaSW1mOXh5RnRZWkM2RjNEeGY0Tm40SStMbzdtKzdoWXRUUmE0NUN0cEpwaGdXNHdiZHpoNHJyUFY0OGFGcWcyczNKQUNWdlJPaGlPNEsxN3ZHTDVmQlFzQmVrem1vZWF2NTMvL2dPTm1kZE5uMnZXajdLM3VhMWk5YnlqT3Y4QTgyK3Z6UzJEOXpqMXgybWZOQTRpZkZsM0NBWTFkL1RMdmEyaGNNMzg0eWxsdE94Z3V1UTlYekRyUGFUdzMxaFkrWWxtRG1Td0pLQ2VzSTRKdXl0MzVqNW96aDNpTjliYnp4WG9yTkY1bXpQdmx6djgxS2FPRit4SjYvZWI1SFI3SDdxaVovRU9mNXV3YWZmaTNOTXo5WHlhdE9oRi9jdU8xei82czVjY2Q0N01iMCsyVmpRN0g1MGNuYjlwaU9YZExlenJiU1daeTZWVlpvT3ZIOHc1NnI0L2FTVWJPblRpUGEyZHYzZEF4ZnFnMTcveGw0bWRRalMzYnBhcFdFeGZZaXZPSDk0b0Jkc2FsS0R5OEs2dEVLWU0vb09jVDUzVEFCY0xLdHFkditSME50Z1lvaTNPQS95NmRac08wOXRmVDE0ZG5NVWgxNWdyZnRTVkY5eHJ0YTZXdWh0WjF0cExjLzR1RG1MdzVoYkpqYW5oL2RyYkNFNUdoeElwTEUxZWpWcHVjc01XWEhLdmRKODlhdS9od3ZrWDYzVXlSckxuRlBzOVBkb1BUT1VWZFJBV200cHpIK2d0L2plVXlrRTl5Y2RFVTMxOEZWSGJyV1piY0xmVTNuR0pjM0YxZDJsQ3l6NE5BZGM1dTBIanhmazNpRGJuanp6cDQvT3dxckp2eExucms1ZElDUHZHZ1FrcG92M2tYWEIrbVoxR1VuUGJSaVQydmxkc1ZJWjBsZ2JTTFlhOEhZMU4xM0dHc0hiMVFtdWxOZTAranZmUlVaWU5xay9CUFJ3RWQrTC9oSnV1b2NGb3VSR2pSUzFkc3RuL0s4bmdrM25nYThlZ3JVSEN5QXVRczUxVzIzTk02ZCtMSWNMaWdqL0lieGhYQVlkSU9LQnZQalpPWGgyMi85Zy83T0RiNjJ4N1NEaW5TdzRkTDVCZkJ3VDlzRUhqOTBsTnRqUCt5UkhESUluVHd1Q2NPVTMzdDBkd2NsZ2dPMVAwMWlMcG45dXR4SDNyZzJtWDZ1c0JUZWw5dDUxcGxpcDZlN1UyU3B6Zm4ybkozeCs2b3JvMG1Ya2xZR0xvL25xK0g1S29jWUNuYUQ4eGhKNkNoUFd1aU1EUFZxdGRhOVgxWUt2bTd5cjNOdWFaM0RmNSt6SVJvRzlVVjBMM1pUV0U0SnA5T2t1VnhvUFBxY3RsR3ArUWpkRmVGVFcxQ241ajA1bmlmUWl6RVZmNU1KcnUzNEFWMmNERFBMckRtbC9IcUszU1diVTF0V0QwL1A3UlRldFdHbjlmQkViQ3VQdTloS3RJYXpWNWl1MTIvS3ZtKytIdjdCNEJOeXgrQkQ0S2QzQnd0SXFtRG15RjZ5ZkdxU0wvZThkdWdpek5wMHgxYm9GaVNOaDUrbGlHTDh1VzlpUDZlaWhwQTFuSGUvMTd3NUhucGQ3cjZOV252bjl4dE9pUzMzNmhmdWd1bFlwd1BNT2lMU3BERm1lSVZwVGpsMGNvS0s2RHNxWFJvSXpvVUpPRmRMaVJJa3I1ZFhnNWRwWTQ5WXBvbVV3V045amVQbGFsUmljbFFGcnRTNTZsZkRvU3JZUjNpWXcvZmk2bkxxd0tOa0xGaWVHWWFTRXF3R0dZYVNFeFlsaEdDbGhjV0lZUmtwWW5CaUdrUklXSjRaaHBJVEZpV2tYZS9ic2djaklTSEZNbmp3WjFxNWRxN2RKVENlRnhZbHBGMGVQSG9XMHREUjQvUEhIWWRDZ1FUQjc5bXlJamFVVk5ZYWhRZWRlT2NmY011TGk0c1JuU1VrSnJGbXpCdGF2cHhOemo2RUJ0NXlZRHJGNzkyNllNR0dDM21Zd25SQVdKOFltREFhRE9Db3JLMkhPbkRsNm04TjBRbGljR0p1b3E2dURpb29LZU9xcHB5QXFLZ3JLeXNyME5vbnBaTEE0TVRiajdPd01DeGN1Rk9lcHFkcVJqQm5HVmxpY21BNnhkZXRXOFRscWxOeUJEUmg2OEd3ZFl4TTQzcVN5ZE9sU0NBaXc3Q2lQWVd5QlhhWXdEQ01sM0sxakdFWktXSndZaHBFU0ZpZUdZYVNFeFlsaEdDbGhjV0lZUmtwWW5CaUdrUklXSjRaaHBPVC9oUC81eklKbmJiMEFBQUFBU1VWT1JLNUNZSUk9.jpg?w=700&webp=1)
网格搜索与K折交叉验证结合调整超参数的具体步骤:
- 确定评价指标(准确率等)
- 对于超参数取值的每种组合,在训练集上使用交叉验证的方法求得其K次评价的性能均值
- 最后,比较哪种超参数取值组合的性能最好,从而得到最优超参数的取值组合。
3.分类及其性能度量
分类问题是有监督学习的一个核心问题。分类解决的是要预测样本属于哪个或者哪些预定义的类别。此时输出变量通常取有限个离散值。
分类的机器学习的两大阶段:
- 从训练数据中学习得到一个分类决策函数或分类模型,称为分类器( classifier);
- 利用学习得到的分类器对新的输入样本进行类别预测。
两类分类问题与多类分类问题:
多类分类问题也可以转化为两类分类问题解决,如采用一对其余One-Vs-Rest的方法:将其中一个类标记为正类,然后将剩余的其它类都标记成负类。
分类性能度量
假设只有两类样本,即正例(positive和负例 negative)。通常以关注的类为正类,其他类为负类。
表中AB模式:第二个符号表示预测的类别,第一个表示预测结果对了(True)还是错了(False)
- 分类准确率( accuracy):分类器正确分类的样本数与总样本数之比
![](https://image.shishitao.com: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%3D.jpg?w=700&webp=1)
- 精确率( precision):反映了模型判定的正例中真正正例的比重。在垃圾短信分类器中,是指预测出的垃圾短信中真正垃圾短信的比例。
![](https://image.shishitao.com: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%3D%3D.jpg?w=700&webp=1)
- 召回率( recall):反映了总正例中被模型正确判定正例的比重。医学领域也叫做灵敏度( sensitivity)。在垃圾短信分类器中,指所有真的垃圾短信被分类器正确找出来的比例。
![](https://image.shishitao.com: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%3D%3D.jpg?w=700&webp=1)
P-R曲线:
如何绘制?要得到PR曲线,需要一系列 Precision和Recall的值。这些系列值是通过阈值来形成的。对于每个测试样本,分类器一般都会给了“Score”值,表示该样本多大概率上属于正例。
步骤:
- 从高到低将“ Score"值排序并依此作为阈值 threshold;
- 对于每个阈值,“ Score"值大于或等于这个 threshold的测试样本被认为正例,其它为负例。从而形成一组预测数据。
Area( Area Under Curve,或者简称AUC):有助于弥补P、R的单点值局限性,可以反映全局性能。
Area的定义(p-r曲线下的面积):![](https://image.shishitao.com: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%3D%3D.jpg?w=700&webp=1)
![](https://image.shishitao.com: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%3D%3D.jpg?w=700&webp=1)
F值:
ROC(受试者工作特征曲线,receiver operating characteristic curve)描绘了分类器在 tp rate(真正正例占总正例的比率,反映命中概率,纵轴)和fp rate(错误的正例占反例的比率,反映误诊率、假阳性率、虚惊概率,橫轴)间的trade-off。ROC曲线绘制和P-R曲线类似。
ROC- AUC( Area Under Curve)定义为ROC曲线下的面积
AUC值提供了分类器的一个整体数值。通常AUC越大,分类器更好。取值范围为[0,1]
分类性能可视化
混淆矩阵的可视化:可用热图(heatmap)直观展现类别的混淆情况
分类报告:显示每个类的分类性能,包括每个类标签的精确率、召回率、F1值等。
4.回归问题及其性能分析
回归分析( regression analysis)是确定两种或两种以上变量间相互依赖的定量关系的一种统计分析方法,和分类问题不同,回归通常输出为一个实数数值。而分类的输出通常为若干指定的类别标签。
回归性能度量方法:
- 平均绝对误差MAE
MAE( Mean absolute error)是绝对误差损失( absolute error loss)的期望值。
- 均方差MSE
MSE( Mean squared error)该指标对应于平方误差损失( squared errorloss)的期望值。
均方根差RMSE:是MSE的平方根
- logistic回归损失(二类)
简称 Log loss,或交叉熵损失( cross-entropy loss),常用于评价逻辑回归LR和神经网络
对于二类分类问题:
1.假设某样本的真实标签为y(取值为0或1),概率估计为p=pr(y=1),
2.每个样本的 log loss是对分类器给定真实标签的负log似然估计(negative log-likelihood)
![](https://image.shishitao.com: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%3D%3D.jpg?w=700&webp=1)
eg.下面y_pred的每个元素有两个值,分别表示该样本属于不同类标签的概率,两数之和为1。
-
logistic回归损失(多类)
对于多类问题( multiclass problem),可将样本的真实标签( true label)编码成1-of-K(K为类别总数)的二元指示矩阵Y:
举例:假设K=3,即三个类
假设模型对测试样本的概率估计结果为P,则在测试集(假设测试样本总数为N)上的交叉熵损失表示如下:
![](https://image.shishitao.com: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%3D.jpg?w=700&webp=1)
其中yi,k表示第个样本的第k个标签的真实值(即ground truth,具体含义为第i个样本是否属于第k个标签),注意由于表示为“1-of-K模式因此每个样本只有其中一个标签值为1,其余均为0。pi,k表示模型对第i个样本的第k个标签的预测值。
举例:6个样本,三个类
回归评价的真实标签(即ground truth)如何获得?
MAE,RMSE(MSE)常用于评分预测评价,eg很多提供推荐服务的网站都有一个让用户给物品打分的功能预测用广对物品评分的行为称为评分预测。
5.一致性的评价方法
一致性评价是指对两个或多个相关的变量进行分析,从而衡量其相关性的密切程度。
假设两评委( rater)对5部电影的评分如下,则二者的一致如何?
rater1=[0.5,1.6,25,25,24]
rater2=[1.5,26,35,3.5,34]
-
皮尔森相关系数法
应用背景:用来衡量两个用户之间兴趣的一致性
用来衡量预测值与真实值之间的相关性
既适用于离散的、也适用于连续变量的相关分析
X和Y之间的皮尔森相关系数计算公式:![]()
![](https://image.shishitao.com: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%3D%3D.jpg?w=700&webp=1)
![](https://image.shishitao.com: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%3D%3D.jpg?w=700&webp=1)
其中,cov(X,Y)表示X和Y之间的协方差( Covariance),![]()
是X的均方差,![]()
是X的均值,E表示数学期望
![](https://image.shishitao.com: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%3D%3D.jpg?w=700&webp=1)
![](https://image.shishitao.com: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.jpg?w=700&webp=1)
取值区间为[1,1]。-1:完全的负相关,+1:表示完全的正相关,0:没有线性相关。
-
Cohen\'s kappa相关系数
与 皮尔森相关系数的区别:Cohens kappa相关系数通常用于离散的分类的一致性评价。
其通常被认为比两人之间的简单一致百分比更强壮,因为 Cohenskappa考虑到了二人之间的随机一致的可能性
如果评价者多于2人时,可以考虑使用Fleiss\' kappa
Cohen\'s kappa的计算方法:
kappa score是一个介于-1到+1之间的数。
![](https://image.shishitao.com:8440/aHR0cHM6Ly9pbWcyMDIwLmNuYmxvZ3MuY29tL2Jsb2cvODE3MTI4LzIwMjAwNS84MTcxMjgtMjAyMDA1MzAxNzI2MzIxNzUtODQxMTcxMDIwLnBuZw%3D%3D.png?w=700&webp=1)