Incremental Random Forest
An implementation in C++ (with node.js and Python bindings) of a variant of Leo Breiman's Random Forests
The forest is maintained incrementally as samples are added or removed - rather than fully rebuilt from scratch every time - to save effort.
It is not a streaming implementation, all the samples are stored and will be reseen when required to recursively rebuild invalidated subtrees. The effort to update each individual tree can vary substantially but the overall effort to update the forest is averaged across the trees so tends not to vary so much.
IRF is licensed under the MIT license.
Features and limitations
- Sparse feature vectors
- Samples can be added, removed and changed
- Learning can be performed lazily or initiated explicitly
- The forest can be serialized to JSON for transmission/storage
- The forest needs to fit fully in RAM, performance suffers dramatically when swapping
- Currently only binary classification - 0 or 1. The classifier estimates the probability of belonging to class 1, as a float from 0 to 1
- Currently only binary features: y >= 0.5 is considered 1, otherwise 0
Node.js setup
npm install irf
Node.js usage
var irf = ; var f = 99; // create forest of 99 trees f; // add a sample identified as '1' with the given feature values, classified as 0f; // features are stored sparsely, when a value is not given it will be taken as 0f; // but 0s can also be given explicitly// ... var y = f; // classify feature vector // the forest will be lazily updated before classificationf; // but you can force an update at any time // you get a probability estimate from 0 to 1 for belong to class 1var c = Math; // round to nearest to get class (0 or 1) f; // remove a samplef; // and add it again with new values console; // serialize to json (for classification, not suitable for incremental update) f; var b = f; // serialize (complete) to buffervar f2 = b; // construct from buffer contents
Python setup
cd irf
python setup.py install
Python usage
import irf f = # create forest of 99 trees # add a sample identified as '1' with the given feature values, classified as 0 # features are stored sparsely, when a value is not given it will be taken as 0 # but 0s can also be given explicitly # ... y = ; print y, # classify feature vector, round to nearest to get class # save forest to file f = # load forest from file # remove a sample # and add it again with new values y = ; print y, # the forest will be lazily updated before classification # f.commit() # but you can force it for (sId, x, y) in : # iterate through samples in the forest, in lexicographic ID order print sId, x, y # and print them
C++ usage
to be written
Dependencies
System:
- STL
Included:
- MurmurHash3 (from smhasher)
External:
Tests
- simple.py - trivial made up data to illustrate how to use the API
- mushrooms.js, mushrooms.py - using the mushrooms dataset collected by LIBSVM from the UCI Machine Learning Repository