A Brief Introduction to Berserker

A Brief Introduction to Berserker

Nov 1, 2015. | By: Jake

If you compete in or follow the machine learning competitions on kaggle, then by now you’re probably familiar with the concept of ensembling. Almost without exception, the winners of each competition are combining a variety of estimators into a single, more powerful model. It doesn’t matter what kind of data is given, or whether regression or classification is required, there isn’t a single machine learning algorithm that can compete with an ensemble. Even if a shiny new algorithm were published tomorrow that was objectively better than random forests, gradient boosting, and the like - you could just add it to you ensemble, making it even better.

There exists a variety of ensembling techniques, which mostly stem two methods known as “stacking” and “blending”. I may go over these in depth in a future post, but for now all you need to know is that several diverse models separately make predictions, which are then combined into a single prediction. The basic idea is that while all models have imperfections, most of the models are correct for a given prediction so each individual error has negligible effect.

The Case for Berserker

I think its fair to say that most of us don’t rewrite the random forest algorithm in C code every time we want to use it. We have things like scikit-learn so we don’t have to continuously reinvent the wheel. Yet for some reason we don’t have a generalized, reusable tools for creating ensembles, despite the fact that everyone and their mother is using them. That is why I created Berserker.

You can get all of the details in the readme, but here are a few key features:

  • A familiar scikit-learn api/syntax
  • Prediction memoization
  • Generate models algorithmically

A Simple Example

Finally, I’ll leave you with the source code and output for a demo using the popular Boston housing prices dataset. With only a few lines of code, you can create an ensemble (of ensembles) which outperforms the vanilla random forest and GBT in scikit-learn.

from berserker.ensemble import Ensemble
from berserker.layers import Layer
from berserker.nodes import Node

model = Ensemble(X_trn, y_trn, mean_squared_error)

# base estimator pool
model.add_node(RandomForestRegressor(50), name='50 Tree Random Forest')
model.add_node(GradientBoostingRegressor(n_estimators=250), name='250 Gradient Boosted Trees')

# meta-estimator
model.add_node(LinearRegression(), name='Lin Reg Meta Estimator')

preds = model.predict(X_tst)
Level 1 Estimators (12 features)     Validation Error
50 Tree RF                            16.1368
Gradient Boosted Trees                18.4357

Level 2 Estimators (14 features)      Validation Error
Lin Reg Meta Estimator                15.5071

I urge you to try it out. This is my first attempt at writing a library, so I openly welcome any criticism.


Subscribe to this blog via RSS.


Recent Posts

Popular Tags