How to increase the accuracy of the classifiers?

haziqros_97

Hye everyone!

I really need your help. I am doing my thesis research using RM. My supervisor asked me to use Turbo Prep. The issue I faced now is the accuracy of the classifiers is below 70%. My supervisor wanted me to get above 80%, if possible. Can you guys give me some tips on how to increase the accuracy?

Thanks in advance.

lionelderkrikor

@haziqros_97,

You can find here a process with a performance strictly greater to 70 % using : 
 - Smote upsampling
 - feature engineering

Feel free to increase the time limit associated to the Automatic Feature Engineering operator (currently se to 180 seconds).

I think that to reach 80 % accuracy will be extremely difficult, maybe impossible...
Keep in mind that the accuracy is not the only criteria (the "Kaggle syndrom") to choose a model, you have to take into account  : 
 - the simplicity of the model
 - the interpretability of the model
 - the capacity to "generalize" of the model...

Hope this helps,

Regards,

Lionel

lionelderkrikor

HI @haziqros_97,

wide subject !!!

You can  : 

 - preprocess /clean your data : impute missing value, replace rares values etc.
 - perform feature engineering (feature selection/feature generation)
 - try different models (NB, neural networks, Decision tree)
 - optimize the parameters of the best model (using Optimize Parameters operator)

and ideally perform a Cross Validation to evaluate the accuracy of your models. It is considered as a best practice, representative of the performance of your model(s) on future unseen data.

I hope it helps  !

Regards,

Lionel

PS : Note that all the 4 steps described above can be automatically performed by AutoModel inside RapidMiner...

haziqros_97

Hi mate, thanks a lot for answering my question. I try it first. Later I'll get back to you.

haziqros_97

Sorry mate. I have tried as you said but the accuracy is still the same. Here I attach my original data before I cleaned it. Hopefully, someone can help me to do this and get the accuracy above 80%. 
Please help me, guys. 

Thanks in advance.

lionelderkrikor

@haziqros_97

What is the label in your dataset ?

Lionel

haziqros_97

SPM performance for English and Mathematics.

lionelderkrikor

@haziqros_97,

You can find here a process with a performance strictly greater to 70 % using : 
 - Smote upsampling
 - feature engineering

Feel free to increase the time limit associated to the Automatic Feature Engineering operator (currently se to 180 seconds).

I think that to reach 80 % accuracy will be extremely difficult, maybe impossible...
Keep in mind that the accuracy is not the only criteria (the "Kaggle syndrom") to choose a model, you have to take into account  : 
 - the simplicity of the model
 - the interpretability of the model
 - the capacity to "generalize" of the model...

Hope this helps,

Regards,

Lionel

lionelderkrikor

@haziqros_97,

And an other process using the same preprocessing steps described above, but using the deep-learning library....

Hope this helps,

Regards,

Lionel

haziqros_97

Thanks lad for answering. Wishing you best of luck.