so, you want your forest to solve a task that human cannot do, even given a long time for analysis?

I am not sure I understand

Кстати, хороший вопрос, самому стало интересно - как находить зависимости между фичами?

удалять одну и смотреть ухудшился ли результат модели - самый дорогой и точный метод

Я думаю это грубо, т.к. может быть парная зависимость, коррелирующая в точность предсказания

you want to guess a cancer status

lets say a doctor wants to do the same

how would a doctor do this job?

Yes deleting 1 by 1 is also knows as backward feature elimination technique

I tried that already gave me 39 features out of 150

Machine learning is all about finding pattern in data than humans cannot. Otherwise why use machine learning at all if humanly possible

можно сделать автоэнкодер

он конкретно эту задачу решает - восстановление потерянных данных имея большую часть данных

чем лучше восстанавливается фича если ее удалить - тем больше она предсказуема из других

это намного дешевле чем тренировать новую модкль на каждую удаленную фичу

Autoencoder is in neural networks right ? Unfortunately cannot use neural as data very less

im not saying its not possible

its just a question to determine how complex is your task

There were quite good count of materials when I googled feature selection

One I liked is: Correlation Matrix with Heatmap

I tried : hybrid feature selection like genetic algorithm/simulated annealing and wrapper feature selection like recursive feature elimination. But performance was best with accuracy of 69 only. Problem is most feature selection don't work good with mixed data

What do you mean - mixed?

Mixed : numeric+ categorical

so, to summarize

you use random forests
you give it data like this:
age 18 - no
age 25 - yes
age 30 - no
blood pressure 100 - no
blood pressure 120 - yes
blood pressure 140 - no
female - no
male - yes (as a category?)

at the end you want to receive a chance of cancer for this data entry.

and you have a data values of real patients who were suffering cancer or not, and you check the performance of your forest against this data of real patients

Can I share photo here of my data

i think yes