Reading 3: Machine Learning

age No. 159Date//Reading: Machine Learning* MLThe goal of ML is to filter useful info from great quantities aof data by learning from known ex to find a pattern inthe data.The obj is to determine structer or generate forecasts withouthuman d interventionOne straightforward way to think of ML is "find thepattern apply the pattern"NoteML makes no assumption about distribution of underlying dataTarget Variable - This is dependent variable Lie they y variable)Features - Independent variable (i.e the X variables)Training data set - This is the Sample used to fit the modelHyper param eter - This is a model input specified by the researchML :- Machines display intelligent decision making ability throughactivities such as sensing, reasoning & understandingSupervised Machine LearningUse Labeled training data + organized sets of observedfeatures Cinputs , or Xs) & the associated target Cout put or 4) Y - TargetPage No. 160X - FeatuleDate//H is the process of training on algorithm to take a set ofinput X & find out a model that best relate them outputtheOutputCan be either a regression or classification problemHistorical Relaw-f the target variable is controus, then it is regression Eg Prob ofIf the target variable is ordinal or categorical (eg bond ratiesthen it is a classification. creditfatingFrau t detection ofUnsupervised Machine Learningcredit card transactionAlgorithms are trained without using labeled dataRelationship bet" features are inferred reflecting an underlyingStructure that was not provided.Kinds of problems that are appropriate for unsupervised MLinclude clustering and dimension reductionDeep LearningsDeep Learning is a ML algorithm based on neural networksDeep learning is used for complex tasks like classification,Speech recognition and natural language processing face recognition& reinforcement learning1Deep learning alograriths are used for complex tasks Such asimage recognition natural language processing and so on0Programs that learn from their own prediction error arecalled reinforced learning algorithms Page No. 161Date / /ML Algorithm Type VIMPoutputPatternVariableSupervisedUnsupervisedJ(Target variable)(No target variable)HContinuousRegressionDimensionality ReductionEx quaterlyLinear ; Penalized RegressionLASSOPrincipal components Analysis b)returnLogistic(P(A)classification & Regression Tree (CART)ClusteringRandom ForestK- - -MeanHierarchicalCategoricalClassificationDimensionality Reductionor ordinalLogit - Result I or2 True 01 felsePrincipal components (PCA)Odad thatSupport Vector Machine (SVM)Clusteringa0 06K- nearest Neighbor (KNN)K - MeanTime (TulseClassification & Regression TreeHierarchicalCORRT)Continuous orNeural NetworkNeural NetworkCategoricalDeep LearningDeep LearningReinforcement LearningReinforcement LearningfairBay Us Out VcBias - underfitiagoverfilling VarianceinsampleOutd sampleerrorerrol Page No. 162Date / /Dimension ReductionComplex databaseStep ISimple databaseStep 2Step 3clasilicationNumericalProblemPrediction problemSupervisedUnsupervisedClassificationclassificationData is LinearNon -lincar ComLinear NonLinearStep 4DataDataPenalizedCARTregressionRandom ForestKNNSUMCARTLinearComplex Non-lineurNeutral NetusthRandom ForestsNeural NetworkNeural Network600,00No of categoriesNo of Categories unknownknownK-MeansHierarchal clustering/ INTOverlittingOverlittingRiskN Data ComplexityGeneralizationreferstotheextenttowhichamachinelearningsmodelisabletomakeaccurateout-af-sample(vs in- sample) predictionsTrainy SampleOne big reason for a lack of generalization t is overfithingOverfitting has occurred when the ML alogriths has beenstyledtoocloselyagainstthetrainingdatawiththeresultunfamiliar datathat the algorithm may not John generally apply to new 1training sampleIn SamplePage No 163Validation Sampl }Out of SampleDate//Test sampleRandomnessismispereivedtobeapatternLargenumberoffeatures(ieindependentvariable)areincludedin the data sample0Overfit models donotgeneralizedwelltonewdata(ipout.of sampleSquaredwill be low)0 Decreasestheaccuracyofmodelforecastsonother(out-ofSampledatalostdiffbetinampleoutcoffsumplyOverfiting ofComplexityOut- of - Sample ErrorKOut-of-sample error can originate fromthreesource Bad with outofGood with insampletruptDue to over fitted models Model with poir ofitBias errorunder filting:Theextenttowhichthemodelfitstraining dataVariance error :Overfitting +Theextenttowhichamodels resultschangeOut of Sample errrorsinresponsetothe test&validationsampledataBase error : Originatesromrandomnessinthe dataResidualerrorsduetorandomnois'RegimeChargeOut - sample error = bias error +Varianceerror+baseerrorRandomCostAccuracyLearning CurveRateSize of tes F sampleAlearningCurveplotstheaccuracyrate(i.e1-errorrate )vsthesize of the training sampleatHigh bias ErrorHigh Variance ErrorRobust ModelDesiredaccuracyratoHighRathnotfthengunnysample si eout of sampleLowIn sample Page No. 164Date//ofPreventing Overlitting in Supervised MLK-fold cross-validation : A method for alleviating the holdout SampleAccutacyproblem (when the training set is reduced too muchDetails of the process:- Reduce Sampling ErrorThe data are shuffled randomlyThen divided into k equal sub-SamplesOne Sample, the Kth T is used as a validation sample.K-1 Samples are used as training SamplesNotem ;Variance error increases with model Complexity while bias errordecrease with complexityAn Data Scientists often express this as a trad off bet cost &part of successful model generalization.optimal lvl of complexity minimizes the total error & is a KeyComplexingMethod to reduce overfitting(tea)Similar to Adj R2Complexity ReductionIn Complexity reduction a penalty is imposed to exclude features thatare not meanifully contributing to out-of- sample predictionaccuracy reducing a large # of feature to managable SetThis penalty value increases with the no or independent variablesfeatures) used by the modelKAPenalty Term =LASSOZIbklK=1 Page No. 165Date//Cross ValidationFor a model to learn sufficiently researchers must ensure thatthe training data set is both large & representativeof the populationThe validation Sample similarly should be large & representativeto property test the model.A sampling tech known as cross validation est out-of-sampleerror rates directly from the validation sampleK Fold Cross ValidationIn a K-fold cross validation the sample is sandomly dividedequally into K part.The training Sample comprise (K-1) partsWith one part left for validationError is thenmeasured for the model in each of the partsThis process is repeated k times, & the avg in-sample &out of sampleerrorrates are compiled.Supervised Machine Learning Algorithms.Supervised ML ex :-- Penalized regression- Support vector machineDetermine the problem for- K - nearest neighborwhich they are best suited- Classification & regression tree- Ensemble Learning- Random forest LASSO =W(4i-yi)2+1KNi =1(bk)Page No. 166K=lDate//SSE + Penalty factorSupervised ML are trained using labeled data & can bedivided into two groups.Regression for a Continous target variables which includesPenalized regressionLASSO - eliminate least imp features T performs feature selationClassification for a categorical or ordinal target variable, which includeSupport Vector machine (SVM)k-Nearnest - neighbour (KNN)Classification & regression free (CARt) algorithsPenalized RegressionIt useful for reducing a large no. of features into amanageable set can help avoid the overfitting problem.Ex: Least absolute Shrinkage and selection Operator CLASSO )If 1 =0 penalty OPenalty based on sum of absoute regression coefficient valuesPennalty increase with no. of features includes A >0 :Penalty term = A W HbklK=1Afeature needs to add. to model fit to offset the penaltyfrom it being included few predictive variableProduces persimanious model ( few predictor variable)Ensures that the & variable is only included if sum ofsquared residuals declines by more than the increase inpanalty term. Page No. 167Date//nK^( 4i-+1>bkt=1K=1Support Vector Machine (SVM)SVM is a supervised algorithm used for classification,regression & Outlier detectionlinearTwo factorBest way to sepclassifiervisivalizetwo set of dataludeAttempts to find optimal perplan ethat separate thein two sets of data by max amt.Uses:- Bond classification or sentiment classification of textual dataTwo- Suitable for small - to medium size but highlyComplexmanyfeaturahigh dimensional data sets- Is used to predict compte Company failures-Can be used to classify text from documents intouseful categoriesEx:- Classifying debbis as likely or not likely to defaultortsorting text as +ve vs - veK- NearestNeighbor (KNN)Create cutom inder , Predict bankruptcy, Bond rating predict stock priceK- negrest neighbor (KNN) is a Supervised learning techmost often used for the classification.New observations are classified by finding the " b nearest(most similar) bet a new observation and its K - nearestpieces of data in the current data set.if K=5 the algrorithm will look for the 5 nearest neighborsInvest applications of KNN include assigning bond ratingspredicting bankruptcy predicting stock prices, etc Page No. 168Date//Is used to classify a new observation by finding similaritiesbet the new observation & existing datalimited predictive power section of tree is trimmed (prured)Classification and Regression Tree (CART)CART is another supervised machine learning tech.CART is used to forecast either a categorical target variableresulting in a classification tree, or a Continuous target variablegenerating a regression treeBinary CART algorithms repeatedly divide the data until theterminal nodes are reached Ceg Yes / No)Applications statements of CART include detecting fraudulent financiaand selecting Stocks bondsWe can put Constraint (Regulation parameter)CART Can be applied to :Predict a categorical target variable to product a classification treePredict continuous variablea target to produce a regression treeTo binary classification & regression Visual explanationEnsemble Learnings More accurate & more stableThe Practice of Combining many predictions from many model andavg the result to redence noiseAny individual model will have a certain effor rate & will makeWe can reduce noise Law of large Nom noisy predictions But by avy predictions from many modelEnsemble Learning generally produce predictions that aremore stable and accurate than those of any single nodel Page No. 16416th Jan 23 MunDate//s1] Random Forest Classfier overfittingSignal to noise RatioAssembly of a large no. of decisions trees trained viaabagging methodIn bagging the original training data set is usedtogenerate n new training data sets or "bags" of data- Each new bag is generated by random sampling withreplacement Bagging helps to improve the stability ofpredictions and protects against overfetting the model.Investment applications includes factor based asset allocation&prediction models for the success of an IPOAdvantage-Protect against overfiting on the trapping data ImProve predictionDrawback :-Individualtreescannot be interpreted with relatives easeUses :-In factor - based invest strategies for asset allocation &invest" selection Ex: PO success rat-- In predicting whether an IPO will be succeful given attributesIIVoting ClassifierMajority voting classifiers will assign the predicted label withthe greatestno.ofvotestoanewddepointThe greater individualmodelwhicharetrain ed thethigher the accuracy of aggregated prediction up to a certainpoint &Page No170Date/IIIBootstiap Aggregating (Bagging)TheorginaltrainingdataSetisusedtogeneratennewtraining data Sets or bags of dataAdvantagesofbaggingincludemodelstabilityand=pretedagainst overfitting themodelUnsupervisedMachineLearningAlgoUnsupervisedmachinePrincipallearningsalgoincludeK- meansComponentsanalysisclusteringHierarchical clusteringEachoftheseansupervisedMLalgorithmsissuitedKind of problemfordiffPCASummarize orDoes not use labelleddatareduce highlyAlgorithmsincludecorrelated featuresDimensionreducedbasedonof data into aHierarchicaldustering StepPStandafew main, unconsilatedis analysis2Var-CoComposite variablenuctor & Eig entingdate%Principal Components Analysis (PCA)Xmintion Reduction fech (featureI)Too manUnsupervised ML Algo PCAhighly correlatedLinear cm of fraumeingReducingcomposite variablefeatures into a fow main T uncorrededTransformsthefeature.Covariancematrixinordertoreducehighly corelat ad features into a Smaller no. of uncorrelatedComposite volu 6Page No. 170Date//IIIBootstiap Aggregations (Begging)The orginal training data Set is used to generate n newAdvantages training data of Sets bagging or include bags of model data stability and protectionagainst overfitting the modelUnsupervised Machine Learning AlgoUnsupervised machine learnings algo includePrincipal Components analysisK-means clusteringHierarchical clusteringEach of these ansupervised ML algorithms is suited for diffkind of problemPCASummarize orDoes not use labelled datareduce highlyAlgorithms includeDimension reduced based on principal Components analysisHierarchicalClustering StepStandardize3Grnvetor & Eig house2Var-Cover matrix4Soole%Principal Components Analysis (PCA)Dimintion Reduction fech (featureb)4 Too many featuresUnsupervised ML Algo PCAoverfitingLinear combination of featureReducing highly correlated features into a few main , uncocreak dcomposite variableTransforms the feature Covariance matrix in order to reducehighly corelated features into a smaller no. of uncorrelatedComposite volu SumPage No. 171Date / /Key Concept as 0.1 time Hime } This o' thatthe betteran Eigen vector : New mutually uncorrelated composite Variablesthat are linear Combinations of the origin al featuresProportion of total Variance in the Initial data thatisexplained by each eigenvect orDraw back :- Principal Components Cannot be easily labled or directlyinterpreted by the analyst Black box approachClusteringCoresionwithin group we want conesionSeparationInter group we want SeparationCluster Contains a subsed of observations from thedatasetwhich are similar Obseravation are for way & dissimal toInvest uses of clustering each otherFor grouping companies based on financial statement items orfinancial ratiosImproving folio diversificationPopular dustering approaches includesK-mean clusteringHierachical dustingHuman judgement plays a fole in dustivingK-Means ClusteringHyperparameter - parameter which value has to be selected before handInvolves partitioning the data into a fixed no- K, ofnon-overlapping clusters, K & the no of clusters representsa model parameter.K-means is an older Kind Eninof unsupervised ML algorithmK-means partitions observations into a fixed no. GO (k) of tzooltonon-overlapping clustersront I on notfoEach cluster is described by a controid Page No. 172Date//Each observation is assigned to the cluster whose centroidit is closest to. will change its controrideHierarchical Clustering : Agglomerative & DivisiveHierarchical clustering is an interative unsupervised algo usedto form a hierarchy of clustersBottom-up (agglonerative) clustering beigns with each observationas it own cluster. (small cluster) identifying 5 aTop - Down (divisive) clustering starts with all observations belonginto one single cluster. Clarge cluster )The algorithms create rounds of cluster of increasing ordecreasing Size until a final clustering is reached,=Creationof intermediate rounds of clustering of increasein laggomerative ) of decreasing (anglomeratives) size until G finalclustering is reached.Dend programs Visualization techiqnormallDendrograms highests the hierarchical relationship amongtheclusters. Short dendritiesrepresent a shorter distance betClusterNeural NetworksGrag similarityNeural Networks consist of nodes Connected by linksThree layer S :Input layer - learningHleast 3 HiddenlayerHidden layer s -worksoccusoften mole than 20 output layer - Output Node activationUse to set peaeer of ItStlengthof the inputPage No. 173Date//Learnings happens in the hidden layer nodes- Each has a summation Concultion operator & activiation functionNeural networkswhere it will 8- has been effective in tasks thatfeature -linearities and Complex interactions any variableNeural Network CAKA Artifical neural network ANNS) arehighly flexibile and are commonly used for :Classification learningRegression supervised learningReinforcement learning Computer learn by tiail 4 error*Deep Learnings Net (DLNs)DLNs Comprise of a min of 3 hidden layers but oftenmore than 20 hidden layer Many hidden layersAre the backbone of the artificial intelligence revolutionDLNs take a set of inputs, which are then passed throughseveral layers of non -linear mathematical functions (neurons)At the final F layer / a category is assigned based onhighest probused for complex activities including pattern image & speechrecognitionUsesPattern recognition problemsCredit Card fraud detectionVision & control problem S in autonomus CarsNatural language processingpricing option Page No. 174Date / /Predicting Corporate fundamental factors & price- tech fatesMax rewardReinforcement Learning CRL) consideration for environmentAn IRL algorithm attempts to max rewards over theme withinthe constaints of its environment.Reinformert learning does not have direct labeled data foreach observation , nor instantaneous feedbackAn RL algorithm learns by testing new actions & reuvingits previous experiencesLearnings occurs via millions of trial and error repetitions.