{ "id": 12329, "title": "Under-performing Random Forest code. What's wrong?", "url": "/competitions/digit-recognizer/discussion/12329", "author_name": "davider", "post_date": "2015-01-25T19:18:36.957000", "votes": 0, "comment_count": 3, "views": null, "content": "

Hello,

\n

instead using the RF Benchmark I tried to use cforest_unbiased.

\n

The performance is bad: 0.69186 but I don't understand what's wrong. Any suggestions?

\n

set.seed(666)

\n

train <- read.csv("data/train.csv", header=TRUE)
test <- read.csv("data/test.csv", header=TRUE)

\n

fit <- cforest(label ~ ., data=train, controls=cforest_unbiased(ntree=1000, mtry=3))

\n

Prediction <- predict(fit, test, OOB=TRUE, type = "response")

\n

submit <- data.frame(ImageId = labels(Prediction), Label = Prediction)

\n

write.csv(submit, file = "digit-rf.csv", row.names = FALSE)

", "messages": [ { "id": 63225, "postDate": "2015-01-27T04:19:42.020Z", "content": "

From the package description here, mtry is the "number of input variables randomly sampled as candidates at each node for random forest like algorithms."

\n

The random part of random forest is choosing variables randomly at each levels to capture variances. In this case, setting mtry=3, you are asking to only use 3 variables out of 784 values at each level. That's too few.

\n

You can try the package 'randomForest'. It might be easier to understand.

\n

If you want to use other algorithms, I wrote a blog about all the resources I find here.

", "votes": 1 }, { "id": 63246, "postDate": "2015-01-27T19:23:55.487Z", "content": "

The default value is square root. I got around 95% with it. I do not know if it is the best. You can always tune this parameter using cross validation.

" }, { "id": 63226, "postDate": "2015-01-27T06:42:46.810Z", "content": "

Thank you Qiwei, now it's clear. I found this interesting publication:

\n

http://www.math.u-psud.fr/~genuer/genuer-poggi-tuleau.rf-insights.pdf

\n

For this classification problem do you advise to use the square root of number of inputs (mtry=28)?

" }, { "id": 63152, "postDate": "2015-01-25T19:18:36.957Z", "content": "

Hello,

\n

instead using the RF Benchmark I tried to use cforest_unbiased.

\n

The performance is bad: 0.69186 but I don't understand what's wrong. Any suggestions?

\n

set.seed(666)

\n

train <- read.csv("data/train.csv", header=TRUE)
test <- read.csv("data/test.csv", header=TRUE)

\n

fit <- cforest(label ~ ., data=train, controls=cforest_unbiased(ntree=1000, mtry=3))

\n

Prediction <- predict(fit, test, OOB=TRUE, type = "response")

\n

submit <- data.frame(ImageId = labels(Prediction), Label = Prediction)

\n

write.csv(submit, file = "digit-rf.csv", row.names = FALSE)

" } ], "comments": [ { "id": 63225, "author_name": "", "author_url": "", "post_date": "2015-01-27T04:19:42.020000", "content": "

From the package description here, mtry is the "number of input variables randomly sampled as candidates at each node for random forest like algorithms."

\n

The random part of random forest is choosing variables randomly at each levels to capture variances. In this case, setting mtry=3, you are asking to only use 3 variables out of 784 values at each level. That's too few.

\n

You can try the package 'randomForest'. It might be easier to understand.

\n

If you want to use other algorithms, I wrote a blog about all the resources I find here.

", "votes": 1, "replies": [] }, { "id": 63246, "author_name": "", "author_url": "", "post_date": "2015-01-27T19:23:55.487000", "content": "

The default value is square root. I got around 95% with it. I do not know if it is the best. You can always tune this parameter using cross validation.

", "votes": 0, "replies": [] }, { "id": 63226, "author_name": "davider", "author_url": "", "post_date": "2015-01-27T06:42:46.810000", "content": "

Thank you Qiwei, now it's clear. I found this interesting publication:

\n

http://www.math.u-psud.fr/~genuer/genuer-poggi-tuleau.rf-insights.pdf

\n

For this classification problem do you advise to use the square root of number of inputs (mtry=28)?

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "63225": "", "63246": "", "63226": "", "63152": "" } }