文件潜在树模型第二部分应用科技计算机科学与工程系大纲第一部.pdf

LatentTreeModels

PartII:Applications

NevinL.Zhang

DepartmentofComputerScienceandEngineering

TheUniversityofScienceandTechnology

Outline

PartI:ConceptandAlgorithms

Theoreticalissues

Basiclearningalgorithms

Scalingup

Learninghierarchicalmodels

PartII:Applications

Hierarchicaltopicdetection

Itemmendationwithimplicitfeedback

Others

HierarchicalLatentTreeysis(HLTA)

Eachwordisabinaryvariable

0–absencefromdoc,1–presenceindoc

Eachisabinaryvectorovervocabulary

HLTAhasbeenextendedtohandlewordcountsbyChal(2017).

Topics

Eachlatentvariablepartitionsdocsinto2clusters

clustersinterpretedastopics

Z14=0:backgroundtopic

Z14=1:“-card-driver”

Eachlatentvariablegivesopic

TopicHierarchy

Latentvariablesathighlevels

“long-range”wordco-occurrences,moregeneraltopics

Latentvariablesatlowlevels

“short-range”wordco-occurrences,morespecifictopics.

TheNewYorkTimesDataset

FromUCI

300,000articlesfrom1987-2007

10,000wordsselectedusingTF/IDF

HLTAtook7hoursonadesktopmachine

ModelLearnedfrom300,000NewYorkTimesArticles

()

PartofTopicTreeaboutSports

(Entiretree:)

Top10TopicsaboutCountries

0.09beijinghong_kong

0.07japanesejapantokyoyen

0.06soccerworld_cupseoulsouth_korea

0.0