predictive modeling: predict premium subscriber for a leading international music website
TRANSCRIPT
MSBA 6420 Rapid WinnersKaushik Nuvvula, Pankaj Singhal, Wenqiuli Zhang, John Tong, Rohith D
Rapid Approaches
Agenda
Approach
Different techniques used
Techniques that worked
Techniques that did not work
Best Model
Future Scope and Learnings
Approach
Neural Network
Voting (SVM, Neural)
Boosting (Neural)
Bagging (Neural)
Sampling, NormalizationData Pre-processing, Weights, Bagging, Voting, SamplingData Pre-processing, Sampling, Generate AttributesSelect by weights, Data Pre-processing, SamplingData Pre-processing, Sampling
Top 5 Models
Model Techniques Used
Stacking (k-NN, Neural)
F-measure and cost: Top 5 Models
Techniques that worked
Data Processing
Data Preprocessing
Generate Attributes
Attribute Selection Techniques
Attributes Selection
Optimize Parameters
Techniques
PCA, SMOTE
Voting
NormalizationBagging, Boosting,
Stacking
Filter Examples Sampling
6
SMOTE: Resampling Approach
• SMOTE -Synthetic Minority Oversampling combines Informed Oversampling of the minority class with Random Under-sampling of the majority class.
• For each minority Sample– Find its k-nearest minority neighbors– Randomly select j of these neighbors– Randomly generate synthetic samples along the lines joining the minority sample and its j
selected neighbors
*SMOTE currently yields best results as far as re-sampling and modifying probabilistic estimate techniques (Chawla, 2003).
Deep Dive: SMOTE Sampling
: Minority sample
: Synthetic sample
What happens if there is a nearby majority sample?
: Majority sample
Techniques that did not work
• Meta-cost• Forward Selection• Logistic Regression
Best Model
Neural Network
Class 0: Above 0.1
Class 0: Between 0.03 – 0.1
F- Measure and Misclassification Cost Improvement
Scope - ImprovementsFi
lter E
xam
ples
Metric Change Improvement
Average Friend Age
17 to 31 Positive
Tenure > 4 Positive
Songs Listened
> 1 Negative
Age > 8 and <70 Negative
Key Learnings: Warnings
• Remove Oversampling - Bias in the data• Generate Calculated Attributes• complex f-measure• Try to train your models on relatively higher
variability capturing records – Using Filter Examples
Appendix
True 0 True 1
Pred. 0 24259 335
Pred. 1 1302 109
True 0 True 1
Pred. 0 23442 320
Pred. 1 1289 400
F- Measure
Misclassification Cost