Sophomore SlumpwareOverviewSlide 3Feature dataSlide 5Data labellingData preprocessingThe Neural NetworkResultsFuture ImprovementsSophomore SlumpwareSophomore SlumpwarePredicting Album Sales with Predicting Album Sales with Artificial Neural NetworksArtificial Neural NetworksMatthew Wirtala ECE 539Matthew Wirtala ECE 539OverviewOverviewRecord sales have decreased ~30% over Record sales have decreased ~30% over the past 4 yearsthe past 4 yearsNo consensus on why this isNo consensus on why this isFile-sharing?File-sharing?Inferior albums being released?Inferior albums being released?OverviewOverviewPerhaps album sales can be predicted Perhaps album sales can be predicted with an MLP networkwith an MLP networkMay show what factors determine how well an May show what factors determine how well an album will sellalbum will sellIndicate which albums deserve a better Indicate which albums deserve a better marketing pushmarketing pushFeature dataFeature dataCritical acclaimCritical acclaimReview scores gathered from 4 sourcesReview scores gathered from 4 sourceswww.pitchforkmedia.comwww.pitchforkmedia.comwww.allmusic.comwww.allmusic.comwww.metacritic.comwww.metacritic.comRolling StoneRolling StoneFeature dataFeature dataHype levelHype levelAmount of press coverage will lead to higher Amount of press coverage will lead to higher public awareness and possibly higher album public awareness and possibly higher album salessalesPrevious album salesPrevious album salesServe as barometer of how established an Serve as barometer of how established an artist may be. artist may be.Data labellingData labellingToo difficult to predict exact album salesToo difficult to predict exact album salesData labelled as one of three classesData labelled as one of three classesAlbums that sell fewer than 500,000 copiesAlbums that sell fewer than 500,000 copiesGold albums (500,000 – 1,000,000 copies)Gold albums (500,000 – 1,000,000 copies)Platinum albums ( > 1,000,000 copies sold)Platinum albums ( > 1,000,000 copies sold)Data preprocessingData preprocessingData gathered for 60 albumsData gathered for 60 albums20 from each class20 from each classSome from same artist falling into separate Some from same artist falling into separate classesclassesData randomized and split into three Data randomized and split into three partitionspartitionsFeature vectors normalized to -5 - +5Feature vectors normalized to -5 - +5The Neural NetworkThe Neural NetworkUtilized Professor Hu’s standard bp.m Utilized Professor Hu’s standard bp.m algorithmalgorithmTrialed many different configurationsTrialed many different configurationsOptimal configurationOptimal configuration2 hidden layers2 hidden layers7 neurons in first layer, 8 in second7 neurons in first layer, 8 in secondLearning rate = 0.267, momentum = 0.007Learning rate = 0.267, momentum = 0.007Tested with 3-way cross validationTested with 3-way cross validationResultsResultsHighest classification rate 60%Highest classification rate 60%Correctly classified class 1 and 2 albums with Correctly classified class 1 and 2 albums with 80-90% accuracy80-90% accuracyCould not separate class 2 albumsCould not separate class 2 albumsClass 2 featured albums with vectors similar to Class 2 featured albums with vectors similar to those of classes 1 and 3those of classes 1 and 3Sample confusion matrix:Sample confusion matrix: 4 0 24 0 2 2 0 52 0 5 1 0 61 0 6Future ImprovementsFuture ImprovementsFurther analysis of feature vectors to Further analysis of feature vectors to determine possible differences in class 2 determine possible differences in class 2 albumsalbumsPossible reduction of labelling to two Possible reduction of labelling to two classes (combine Gold and Platinum)classes (combine Gold and Platinum)Classification does show that predictions Classification does show that predictions can be made based on the features can be made based on the features considered in this studyconsidered in this
View Full Document