Myujth

I have some students data with 1500 rows and 30 columns. i have used GradientBoostingClassifier. All the data is categorical and in the range of 0-80 and 0-90 categories.

I need to build an prediction model to predict if the student will fail and pass. columns 'Status' is my target variable. below is the code i have used

from sklearn import preprocessing
le = preprocessing.LabelEncoder()
data=data.apply(le.fit_transform)

X=data.copy()
y=data['Status']
# Import train_test_split function
from sklearn.model_selection import train_test_split
# Split dataset into training set and test set
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=42)
#Import Gradient Boosting Classifier model
from sklearn.ensemble import GradientBoostingClassifier
gb = GradientBoostingClassifier()
gb.fit(X_train, y_train)

#Predict the response for test dataset
y_pred = gb.predict(X_test)

below is my model performance.

Accuracy: 0.9719317419707395
Precision: 0.9090272918124562
Recall: 0.5650282731622445

Please let me know what should I do to improve the model and how to handle the wide range of categorical data. Also when I am testing the model on different dataset, most of the time the categories are changing either due to spelling mistake or new addition

std_id std_name Dem secn_id location bucket Primary_subject status
144 amy SEP 5.3 P dev english pass
230 mani SEV 11.3 E Tech math fail
299 sam DE 5.1 nap prac science pass
568 samy SEP 1.1 P prac V1 pass
769 elle SEP 1.2 pe prac english pass
761 tanj SEP 1.3 N tech V2 pass
112 jon ERM 3.0 N prac phy fail
116 pal NAN 9.1 sc etc V1V2 pass
116 pal NAN 9.2 sc etc V1V3 fail
113 josh NAN 9.3 du etc. erp fail
100 sug EVV 9.1 sc NAN che pass
323 adi ERP 3.1 NAN fit math fail
323 adi ERP 3.2 NAN fit math fail

this is how my input data looks like.For missing value i have replaced it with string "NAN". There are duplicate records for student if they have changed any option.

Your model performance is decent.
To improve further

1. Tune the parameteres of GradientBoostingClassifier . You can set values for parameters like n_estimators, learning_rate etc and check the performance of your model.
   For this task I'll suggest GridSearchCV




2. Feature Engineering: You can create new features from the existing ones. As you have not provided data it's hard to suggest anything. You can check the feature importance by using Random Forest etc and use features with high importance.




3. You can try different algorithms like XGBoost, LightGBM or even neural network




4. You can use cross-validator like Stratified ShuffleSplit

Regarding your next problem.

Again it's hard to suggest anything without looking at any data.
To avoid spelling mistakes you can enforce users to select values from dropdown etc.
If that's not the case you can look at the difflib library which will find the closest match of your category.

First of all I would suggest to start with some data cleaning and data analysis.
The fact that your categories change due to mistakes need to be fixed in a preprocessing step. Here there are no too many shortcuts, you need to inspect and fix the data manually.

Check also the presence of missing value. If there are missing value you need to address also this issue. You can remove the samples (accepting the loss of information) or substitute the missing value with the average value for the specific feature. There exist other methods in the literature but as first step those two could do.

Please consider to check the number of samples that you have in each class. If the two classes are strongly unbalanced you could consider to look for solution that address "unbalanced data".

Classifier as Decision Trees/Random Forest are good option when handling categorical variables.

The use of a cross-validation to tune the hyper-parameter of the classifier could also improve the performance.

EDIT (after the add of the data table)

Probably you don't want to use the name of the students since that feature is not related to the success/ failure of the exam.