机器学习:4.3 模型验证 Model Validation

@[toc]
Generalization Error

  • Approximated by the error on a holdout test dataset, which has never been seen by the model and can be only used once

    • Your midterm exam score

    • The final price of a pending house sale

    • Dataset used in private leaderboard in Kaggle

  • Validation dataset:

    • Often a subset of the dataset, not used for model training

    • Can be used multiple times for hyper param tuning

    • “test error” usually refers to error on “validation” dataset

Hold Out Validation

  • Split your data into “train” and “valid” sets (often calls “test”)
  • Train your model on the train set, use the error on the valid set to approximate the generalization error
  • Often randomly select n\% examples as the valid set
    • Typical choices n = 50, 40, 30, 20, 10

Split non I.I.D. data

  • Random data splitting may lead to underestimate of generalization error

  • Sequential data

    • e.g. house sales, stock prices

    • Valid set should not overlap with train set in time

  • Examples are highly clustered

    • e.g. photos of the same person, clips of the same video

    • Split clusters instead of examples

  • Highly imbalanced label classes

    • sample more from minor classes

K-fold Cross Validation

  • Useful when not sufficient data

  • Algorithm:

    • Partition the training data into K parts

    • For i = 1, …, K

      • Use the i-th part as the validation set, the rest for training
    • Report the validation error averaged over K rounds

  • Popular choices: K = 5 or 10

Common Mistakes

  • If your ML model performance is too good to be true, very likely
    there is a bug, and contaminated valid set is the #1 reason.
  • Valid set has duplicated examples from train set
    • Often happens when integrating multiple datasets
      • Scrape images from search engine to evaluate models trained on ImageNet
  • Information leaking from train set to valid set
    • Often happens for non I.I.D data
      • use future to predict past, see a person’s face before
  • Excessive use of valid set for hyper param tuning is cheating

Summary

  • The test data is used once to evaluate your model

  • One can hold out a validation set from the training data to estimate the test data

    • You can use valid set multiple times for model selections and hyper
      param tuning

    • Validation data should be close to the test data

    • Improper valid set is a common mistake that lead to over estimate of the model performance

©著作权归作者所有,转载或内容合作请联系作者
  • 序言:七十年代末,一起剥皮案震惊了整个滨河市,随后出现的几起案子,更是在滨河造成了极大的恐慌,老刑警刘岩,带你破解...
    沈念sama阅读 199,393评论 5 467
  • 序言:滨河连续发生了三起死亡事件,死亡现场离奇诡异,居然都是意外死亡,警方通过查阅死者的电脑和手机,发现死者居然都...
    沈念sama阅读 83,790评论 2 376
  • 文/潘晓璐 我一进店门,熙熙楼的掌柜王于贵愁眉苦脸地迎上来,“玉大人,你说我怎么就摊上这事。” “怎么了?”我有些...
    开封第一讲书人阅读 146,391评论 0 330
  • 文/不坏的土叔 我叫张陵,是天一观的道长。 经常有香客问我,道长,这世上最难降的妖魔是什么? 我笑而不...
    开封第一讲书人阅读 53,703评论 1 270
  • 正文 为了忘掉前任,我火速办了婚礼,结果婚礼上,老公的妹妹穿的比我还像新娘。我一直安慰自己,他们只是感情好,可当我...
    茶点故事阅读 62,613评论 5 359
  • 文/花漫 我一把揭开白布。 她就那样静静地躺着,像睡着了一般。 火红的嫁衣衬着肌肤如雪。 梳的纹丝不乱的头发上,一...
    开封第一讲书人阅读 48,003评论 1 275
  • 那天,我揣着相机与录音,去河边找鬼。 笑死,一个胖子当着我的面吹牛,可吹牛的内容都是我干的。 我是一名探鬼主播,决...
    沈念sama阅读 37,507评论 3 390
  • 文/苍兰香墨 我猛地睁开眼,长吁一口气:“原来是场噩梦啊……” “哼!你这毒妇竟也来了?” 一声冷哼从身侧响起,我...
    开封第一讲书人阅读 36,158评论 0 254
  • 序言:老挝万荣一对情侣失踪,失踪者是张志新(化名)和其女友刘颖,没想到半个月后,有当地人在树林里发现了一具尸体,经...
    沈念sama阅读 40,300评论 1 294
  • 正文 独居荒郊野岭守林人离奇死亡,尸身上长有42处带血的脓包…… 初始之章·张勋 以下内容为张勋视角 年9月15日...
    茶点故事阅读 35,256评论 2 317
  • 正文 我和宋清朗相恋三年,在试婚纱的时候发现自己被绿了。 大学时的朋友给我发了我未婚夫和他白月光在一起吃饭的照片。...
    茶点故事阅读 37,274评论 1 328
  • 序言:一个原本活蹦乱跳的男人离奇死亡,死状恐怖,灵堂内的尸体忽然破棺而出,到底是诈尸还是另有隐情,我是刑警宁泽,带...
    沈念sama阅读 32,984评论 3 316
  • 正文 年R本政府宣布,位于F岛的核电站,受9级特大地震影响,放射性物质发生泄漏。R本人自食恶果不足惜,却给世界环境...
    茶点故事阅读 38,569评论 3 303
  • 文/蒙蒙 一、第九天 我趴在偏房一处隐蔽的房顶上张望。 院中可真热闹,春花似锦、人声如沸。这庄子的主人今日做“春日...
    开封第一讲书人阅读 29,662评论 0 19
  • 文/苍兰香墨 我抬头看了看天上的太阳。三九已至,却和暖如春,着一层夹袄步出监牢的瞬间,已是汗流浃背。 一阵脚步声响...
    开封第一讲书人阅读 30,899评论 1 255
  • 我被黑心中介骗来泰国打工, 没想到刚下飞机就差点儿被人妖公主榨干…… 1. 我叫王不留,地道东北人。 一个月前我还...
    沈念sama阅读 42,268评论 2 345
  • 正文 我出身青楼,却偏偏与公主长得像,于是被迫代替她去往敌国和亲。 传闻我的和亲对象是个残疾皇子,可洞房花烛夜当晚...
    茶点故事阅读 41,840评论 2 339

推荐阅读更多精彩内容