The given dataset contains details about organic chemical compounds including their chemical features, isomeric conformation, names and the classes in which they are classified. The compounds are classified as either ‘Musk’ or ‘Non-Musk’ compounds.
Initial info
No. of sample : 6598
No. of Feature : 169
No. of numerical feature : 166
No of classes : 2 (labels arre 0 & 1)
samples per Class : 5581 & 1017 for class 0 & 1 respectively
class Distribution : Highly Imbalance
Missing Values : None
Real world constraints
Minimise binary class error
Probability estimates
Average time constraint(i.e few Minutes)
Mapping Real world problem to ML problem
Type of problem : Binary Classification
Performance Metrics : 1:) AUC and 2:) Confusion Matrix
ML Constraint
TPR must be high
class probability needed
Before performing any kinda univarite or multivariate analysis as soon as I found that dataset is highly imbalance my intution (based on past experience) said me to go for ensembles directly. As one of obvious choice I decided to go for xgboost because it's a great implementation with great ideas in hyper-parameter like row_sampling, column-samling by level.
Before Tunning all the hyperparameter decided to tune only number of base estimators and max depth by setting all the remaining paramerers to default values.
Preprocessing
There are only 2 non-numerical feature so simply dropped it. Definately they could have conained important inormation but thought to give it a try if things will not work i'll try by converitn these features using response coding
There is one column name ID seems like identifier so simply removed
Seperated out the class column (it is label field)
As our base learners will be decision trees so no need to perform standardization
Splited the dataset into 80:20 ratio for training and validation, Fixed seed value for reproducible result. splited using stratify so class disb can be in same ratio.
Training/ Hyperparameter tunning
Used 3 fold cross-Validation
values for estimators : [150,250,350,500]
values for max_depth : [1,2,3,4,5]
Images for tunning
The above images are snapshot of 3-d plot plotted with the help of plotly it's an interactive plot so STRONGLY RECOMMEND to download the notebook and try visualizing the plot.
This 3-D PLOT SHOWs THE TRAIN AND VALIDATION AUC FOR EACH OF 20 HYPERPARAMETER PAIRS BUT NOT POSSIBLE TO RENDER HERE HENCE VISUALIZE IT IN NOTEBOOK.
Took 13 minute for trainnig
Best Hyper-pparam found
no_of_estimator : 350
Max_depth : 4
Training with Best hyper parameter
Train auc is 1 So it might be the case of over fitting (& it's also very easy to have so with DTS)
But the validation auc is also 0.999 so values are very close and definately it's not the case of Overfitting
Calculated the thresold value which will yeild the max AUC
Testing
For the test set got the probability values
Converted the into class labels using previously calculated thresold
Result
Confusion Matrix
Test AUC
Test Precision
Test Recall
Test F1 Score
0.999342
0.939252
0.990147
0.964028
Took approx 1.5 minute to train
Calss accuracies are 98.83% & 99.01% for class 0 & 1 respectively.
Conclusion
We achieved very high accuracy on test data 99%
The data seems genuine without of ouliers & preety strainght Forward
I don't think now there is need of trying further models or DL techniques
Training Time could differ due to system configuration
Bonus
Set n_jobs or n_thread param of xgboost for maximizing the utiliztion of resources for training.
Thankyou for visiting ...
Always listen to your inner voice, it will never let you fall.
.png)
.png)
.png)
