data mining notes.docx

data mining notes.docx

《data mining notes.docx》由会员分享，可在线阅读，更多相关《data mining notes.docx（13页珍藏版）》请在冰豆网上搜索。

dataminingnotes

Intro

Problemscategories

Clustering

Classification

Regression

Dimensionreduction

DataVisualisationandInterpretation 

Descriptivestatistics

Mean

Median

Dispersion

Statisticaldistributions

Themedianisamorerobustestimatorofthecentraltendency

ThedifferenceQ3-Q1istheinterquartilerange（orIQR）it'samorerobustdispersionmeasure

NormalDistribution

SkewedDistribution

Visualisation

Boxplot

Scatterplot

Boxplot:

Theredmarkshowsthemean

Theboxgoesfromthelowerquartiletotheupperquartile

Theboxisthuscentredonthemedian

Thewhiskersaretheminimumandmaximumvalues

Outliersvaluesareshownasbluecrosses

Outliersarevalueswhicharebeyond1.5*IQRfromthequartiles

Distances

HammingDistance

Levenshteindistance

TheLevenshteindistanceistheminimumnumberofeditsneededtotransformonestringintotheother

1insertionofacharacter

2deletionofacharacter

3substitutionofacharacter

DamerauLevenshteindistance

TheDamerauLevenshteindistanceislikethe

Levenshteindistance,withonemoreeditoperation

1insertionofacharacter

2deletionofacharacter

3substitutionofacharacter

4transpositionof2adjacentcharacters

JaroDistance

K-Means （P166）

Clusteringquality

Internal:

External:

Initialization:

random

kmeans++:

distantplots

Perceptron（P67）

Multi-LayerPerceptron（P76）

AMulti-LayerPerceptronismadeofcneurons,connectedtoanoutputneuron

_Eachinnerneuronactsasanindependenthyperplanes

_Thetopneuroncombinestheindependenthyperplanes

_Wecannowclassifydatabycombininghyperplanes

Flaws

Datanormalization

MLPwon'tworkifyoudon'tnormalizeyourdata

_TheoutputoftheMLPisinalimitedrange（say[-1;1]）

_Iftheinputsareoutofrange,MLPlooseinformation

Over-fitting

Pastanumberofaneurons

_Verylittleimprovementoftheerror

_Mostlylearningnoiseofthedata

_over-fitting

Trainingerror:

errorforthetrainingpoints

Testingerror:

errorforpointsnotusedduringtraining

MaximizaionStep:

Initialization

SimpleinitializationforKcomponentsandNsamples

1RandomcreateKclustersofsamples,samesize

2Initial

forCiisthesamplemeanoftheclusteri

3Initial

forCiisthesamplestandarddeviationoftheclusteri

4InitialwforCiis1/k

Kmeansinitialization

multivariateGaussianmix

DecisionTree（p40）