DP
why do you think this data is not good for neural networks?
nn can work well with small size of input too
nn can work well with small size of input too
Size: a a a
2019 July 22
LS
Status 1 is cancer
LS
Many more column : 152
LS
why do you think this data is not good for neural networks?
nn can work well with small size of input too
nn can work well with small size of input too
My research papers say neural networks need thousands or millions rows. Oresle overfit
LS
I have very small dataset of 1800 rows.
DP
I have very small dataset of 1800 rows.
ahhh, now i got what you mean by it.
DP
My research papers say neural networks need thousands or millions rows. Oresle overfit
thats right in general.
but there is more specific rule:
neural network will overfit if it has more neurons than 'rows of dataset'
use neural network with 100 neurons and it wont overfit.
but there is more specific rule:
neural network will overfit if it has more neurons than 'rows of dataset'
use neural network with 100 neurons and it wont overfit.
LS
I will have a very hard time defending neural networks on this data in my presentation
DP
thats right. nn is a black box
LS
My professor said already not to ven touch neural
DP
thats a good reason, okay
(banks have similar requirements, for example)
(banks have similar requirements, for example)
LS
I think my main problem is feature selection
DP
what sort of operation can your tree do?
can it use only (binary) 0\1 data? or can it use (float) 0.376 data too?
can it use only (binary) 0\1 data? or can it use (float) 0.376 data too?
LS
Random forest can use any data number /float/ category
DP
why did you decide to use one hot encoding like this
age 18 = no
age 25 = yes
age 30 = no
male = yes
female = no
instead of float inputs like this:
age: 0.5
gender: 1.0
age 18 = no
age 25 = yes
age 30 = no
male = yes
female = no
instead of float inputs like this:
age: 0.5
gender: 1.0
DP
this would decrease amount of features by a factor of 2...3
seems like worth trying as for me
seems like worth trying as for me
DP
also an option to cheat:
increase number of rules in your forest to 2000 or so
then your forest will 'overfit' on data too.
it will correctly answer on your dataset, but not on other.
might be enough to fool a professor
increase number of rules in your forest to 2000 or so
then your forest will 'overfit' on data too.
it will correctly answer on your dataset, but not on other.
might be enough to fool a professor
О
Its when your own neural got overfitted to pass an exam