Cleanup for release

README.md

@@ -0,0 +1,6 @@
+# EMI Music Hackathon entry
+
+This is code for the 3rd place entry in the [EMI Music Hackathon][hackathon]. It's really not something you should aspire to. Described in more detail [here][blogpost].
+
+[hackathon]: http://www.kaggle.com/c/MusicHackathon
+[blogpost]: http://mewo2.com/kaggle/2012/07/29/emi-music-hackathon---how-i-did-it/

blend.R

@@ -6,12 +6,15 @@ train <- read.csv('data/train.csv');
 
 ratings <- train$Rating;
 
-# preds <- c('lm', 'svd', 'rf', 'demo', 'gbm', 'knn', 'rfbya', 'svdslow', 'lmbya2', 'lmbyt2', 'rfq1');
-preds <- c('svd', 'rf', 'rfbya', 'svdslow', 'lmbya2', 'lmbyt2', 'rfq1');
+preds <- c('lm', 'svd', 'rf', 'demo', 'gbm', 'knn', 'rfbya', 'svdslow', 'lmbya', 'lmbyt');
+# preds <- c('svd', 'rf', 'rfbya', 'svdslow', 'lmbya2', 'lmbyt2', 'rfq1');
 
 trains <- sapply(preds, function (name) read.csv(paste('predictions/', name, '.csv.cross', sep=''))$x)/100;
 tests <- sapply(preds, function (name) read.csv(paste('predictions/', name, '.csv', sep=''))$x)/100;
 
+for (i in 1:length(preds)) {
+  cat(preds[i], rmse(trains[,i] * 100, ratings), '\n');
+}
 mix <- nnet(trains, ratings, size=5, decay=0.1, maxit=500, linout=T, reltol=0, abstol=0, skip=T);
 # mix <- lm(ratings ~ ., data=as.data.frame(trains))
 print(summary(mix));

funk.R

@@ -43,7 +43,7 @@ funksvd <- function (users, tracks, ratings, nfeats=16) {
   # return(list(baseline=baseline,ufeats=ufeats, tfeats=tfeats));
 }
 
-funkpred <- function(train, test, ratings, nfeats=32) {
+svdpred <- function(train, test, ratings, nfeats=16) {
   sv <- funksvd(train$User, train$Track, ratings, nfeats);
   pred <- sapply(1:nrow(test), 
     function (i) {

funk.c

@@ -8,7 +8,7 @@ void funksvd(int* pn, int* users, int* tracks, double* ratings, int* pnfeats, in
   int ntrack = *pntrack;
   int min_loops = 30;
   double min_improve = 1e-6;
-  double k = 5;
+  double k = 4;
   double lrate = 1e-4;
 
   int i = 0;

predict.R

@@ -16,20 +16,14 @@ testfeats <- cleaned[[2]];
 ratings <- cleaned[[3]];
 
 rfpred <- function (trainfeats, testfeats, ratings) {
-  rf <- randomForest(trainfeats, ratings, do.trace=T, sampsize=1000, ntree=100);
+  rf <- randomForest(trainfeats, ratings, do.trace=T, sampsize=50000, ntree=100);
   pred <- predict(rf, testfeats);
   cv <- rf$predicted;
   return(list(pred=pred, cv=cv));
 }
 
-svdpred <- function (trainfeats, testfeats, ratings, n=64) {
-  library(irlba);
-  mu <- mean(ratings);
-  spmat <- sparseMatrix(i=trainfeats$User + 1, j=trainfeats$Track + 1, x=ratings - mu);
-  sv <- irlba(spmat, nu=n, nv=n);
-  rec <- sv$u %*% (sv$d * t(sv$v));
-  pred <- sapply(1:nrow(testfeats), function (i) rec[testfeats$User[i] + 1, testfeats$Track[i] + 1]) + mu;
-  return(list(pred=pred));
+svdslowpred <- function (trainfeats, testfeats, ratings) {
+  svdpred(trainfeats, testfeats, ratings, n=64);
 }
 
 svmpred <- function (trainfeats, testfeats, ratings) {
@@ -171,7 +165,7 @@ rfqpred <- function (train, test, ratings) {
   return(list(pred=pred, cv=cv));
 }
 
-nndemopred <- function (train, test, ratings) {
+demopred <- function (train, test, ratings) {
   mu <- mean(ratings);
   ages <- c(train$AGE, test$AGE)
   ages <- cut(ages, quantile(ages, c(0, 0.2, 0.4, 0.6, 0.8, 1.0)));
@@ -229,19 +223,28 @@ nnpred <- function (train, test, ratings, k=16) {
   return(list(pred=pred));
 }
 source('funk.R');
-# s <- sample(nrow(testfeats), 100000)
-# pred <- cross.val(remove.global(lmbytrackpred), 4, trainfeats, testfeats, ratings);
-pred <- remove.global(rfqpred)(trainfeats, testfeats, ratings);
 
-cat('Estimated RMSE: ', rmse(ratings, pred$cv), '\n');
+save.pred <- function (name, pred) {
+  cat('Estimated RMSE: ', rmse(ratings, pred$cv), '\n');
 
-argv <- commandArgs(T);
-if (length(argv) == 0) {
-  filename <- 'predictions/scratch.csv';
-} else {
-  filename <- argv[1];
+  filename <- paste('predictions/', name, '.csv', sep='');
+
+  write.csv(pred$pred, filename, row.names=F, quote=F);
+  write.csv(pred$cv, paste(filename, '.cross', sep=''), row.names=F, quote=F);
 }
 
-write.csv(pred$pred, filename, row.names=F, quote=F);
-write.csv(pred$cv, paste(filename, '.cross', sep=''), row.names=F, quote=F);
+save.pred('lmbyt', cross.val(remove.global(lmbytrackpred), 10, trainfeats, testfeats, ratings));
+save.pred('lmbya', cross.val(remove.global(lmbyartistpred), 10, trainfeats, testfeats, ratings));
+save.pred('rfbya', cross.val(remove.global(rfbyartistpred), 10, trainfeats, testfeats, ratings));
+
+save.pred('rf', rfpred(trainfeats, testfeats, ratings));
+save.pred('gbm', cross.val(remove.global(gbmpred), 10, trainfeats, testfeats, ratings));
+save.pred('lm', cross.val(remove.global(lmpred), 10, trainfeats, testfeats, ratings));
+
+save.pred('svd', cross.val(remove.global(svdpred), 10, trainfeats, testfeats, ratings));
+save.pred('svdslow', cross.val(remove.global(svdslowpred), 10, trainfeats, testfeats, ratings));
+
+save.pred('knn', cross.val(remove.global(nnpred), 10, trainfeats, testfeats, ratings));
+save.pred('demo', cross.val(remove.global(demopred), 10, trainfeats, testfeats, ratings));
+
 cat('Done\n')