Share
Embed
Transcript
7 slides
-
How you win ML competitions: “ you take other people’s
work and ensemble them together” -‐ Vitaly Kuznetsov NIPS 2014 -‐ Bargava Subramanian (@bargava) -‐ Nischal HP (@nischalhp) -
Increased Accuracy Robustness Efficiency Parallelization Wider
search of solution space Reduces over fitting -
Model human readability is not great Time/Effort
required to build complex ensemble models might not be directly proportional to the accuracy -
{ Building Base Models Model Aggregation Different
Training Sets Different Algorithms Different Parameter Setups Algorithm Randomization Feature Sampling Voting / Averaging Weighted Voting Using as attributes Stacking Bagging -
Input Data Model 1 Model 2 Model
3 Model 4 Combine the models using some logic Final Output -
Random Forest Gradient Boosting Logistic Regression
Ensemble Output 1 0 1 1 0 0 1 0 1 1 0 1 1 1 1 1 0 0 0 0 70% 70% 70% 90% -
Input Data Model 1 Model 2
Model 3 Model 4 Combine the models using some logic Final Output pre processing feature extraction model [ [ - Pipeline identify the models Assign weights to models [ [ - Hyperopt finding randomized hyper parameters for models [ [ - RandomizedSearchCV [ [ - joblib ( running models in parallel ) sklearn keras xgboost [ [ - libraries used to build base models