欢迎来到天天文库
浏览记录
ID:59121779
大小:180.50 KB
页数:22页
时间:2020-09-25
《基于实例的学习ppt课件.ppt》由会员上传分享,免费在线阅读,更多相关内容在教育资源-天天文库。
1、第8章基于实例的学习InstanceBasedLearning1InstanceBasedLearningk-NearestNeighborLocallyweightedregressionRadialbasisfunctionsCase-basedreasoningLazyandeagerlearning2NearestneighborKeyidea:juststorealltrainingexamplesNearestneighbor:Givenqueryinstancexq,firstlocatenearesttrainingexamplexn,thenestimate3k-Ne
2、arestneighbork-Nearestneighbor:Givenxq,Ifdiscrete-valuedtargetfunction:takevoteamongitsknearestnbrsIfreal-valuedtargetfunction:takemeanoffvaluesofk-nearestnbrs4WhenToConsiderNearestNeighborInstancesmaptopointsinLessthan20attributesperinstanceLotsoftrainingdata5Advantages&DisadvantagesAdvantagesT
3、rainingisveryfastLearncomplextargetfunctionsDon’tloseinformationDisadvantagesSlowatquerytimeEasilyfooledbyirrelevantattributes6VoronoiDiagram7BehaviorintheLimitConsiderp(x)definesprobabilitythatinstancexwillbelabeled1(positive)versus0(negative)NearestneighborAsnumberoftrainingexamples∞approache
4、sGibbsAlgorithmGibbs:withprobabilityp(x)predict1else08BehaviorintheLimitConsiderp(x)definesprobabilitythatinstancexwillbelabeled1(positive)versus0(negative)k-NearestneighborAsnumberoftrainingexamples∞andkgetslarge,approachesBayesoptimalBayesoptimal:ifp(x)>0.5thenpredict1,else0NoteGibbshasatmost
5、twicetheexpectederrorofBayesoptimal9Distance-WeightedkNNNearerneighborsaremoreheavilyweighted.Discrete-valuedReal-valuedwhereItmakessensetousealltrainingexamplesinsteadofjustk.[Shepard,1968]10CurseofDimensionalityImagineinstancesdescribedby20attributesbutonly2arerelevanttotargetfunction.Curseofd
6、imensionality:nearestnbriseasilymisleadwhenhigh-dimensionalXApproach:Stretchsomeaxesbyweight,andthencross-validation[MooreandLee,1944]11LocallyWeightedRegressionNotekNNformslocalapproximationtofforeachquerypointxq.Whynotformanexplicitapproximationforregionsurroundingxq?Fitlinearfunctiontokneares
7、tneighborsFitquadratic…Producespiecewiseapproximationtof12SeveralchoicesoferrortominimizeSquarederroroverknearestnbrsDistance-weightedsquarederroroverallnbrs…13RadialBasisFunctionNetworksGlobalapproximationtotargetfunctionin
此文档下载收益归作者所有