Permalink
Fetching contributors…
Cannot retrieve contributors at this time
145 lines (100 sloc) 5.53 KB
output
html_document
keep_md
true

FeatureHashing

Linux: Travis-ci Status Win : Build status OS X: Travis-ci Status

Coverage Status CRAN_Status_Badge rstudio mirror downloads

Implement feature hashing with R

library(knitcitations)
bib <- read.bibtex("README.bib")
# citep(bib[[1]])

Introduction

Feature hashing, also called as the hashing trick, is a method to transform features to vector. Without looking the indices up in an associative array, it applies a hash function to the features and uses their hash values as indices directly.

The package FeatureHashing implements the method in r citep(bib[["DBLP:conf/icml/WeinbergerDLSA09"]]) to transform a data.frame to sparse matrix. The package provides a formula interface similar to model.matrix in R and Matrix::sparse.model.matrix in the package Matrix. Splitting of concatenated data, check the help of test.tag for explanation of concatenated data, during the construction of the model matrix.

Installation

To install the stable version from Cran, run this command:

install.packages("FeatureHashing")

For up-to-date version, please install from github. Windows user will need to install RTools first.

devtools::install_github('wush978/FeatureHashing')

When should we use Feature Hashing?

Feature hashing is useful when the user does not easy to know the dimension of the feature vector. For example, the bag-of-word representation in document classification problem requires scanning entire dataset to know how many words we have, i.e. the dimension of the feature vector.

In general, feature hashing is useful in the following environment:

  • Streaming Environment
  • Distirbuted Environment

Because it is expensive or impossible to know the real dimension of the feature vector.

Getting Started

The following scripts show how to use the FeatureHashing to construct Matrix::dgCMatrix and train a model in other packages which supports Matrix::dgCMatrix as input.

The dataset is a sample from iPinYou dataset which is described in r citep(bib[["zhang2014real"]]).

Logistic Regression with glmnet

# The following script assumes that the data.frame
# of the training dataset and testing dataset are 
# assigned to variable `ipinyou.train` and `ipinyou.test`
# respectively

library(FeatureHashing)

# Checking version.
stopifnot(packageVersion("FeatureHashing") >= package_version("0.9"))

data(ipinyou)
f <- ~ IP + Region + City + AdExchange + Domain +
  URL + AdSlotId + AdSlotWidth + AdSlotHeight +
  AdSlotVisibility + AdSlotFormat + CreativeID +
  Adid + split(UserTag, delim = ",")
# if the version of FeatureHashing is 0.8, please use the following command:
# m.train <- as(hashed.model.matrix(f, ipinyou.train, 2^16, transpose = FALSE), "dgCMatrix")
m.train <- hashed.model.matrix(f, ipinyou.train, 2^16)
m.test <- hashed.model.matrix(f, ipinyou.test, 2^16)

# logistic regression with glmnet

library(glmnet)

cv.g.lr <- cv.glmnet(m.train, ipinyou.train$IsClick,
  family = "binomial")#, type.measure = "auc")
p.lr <- predict(cv.g.lr, m.test, s="lambda.min")
auc(ipinyou.test$IsClick, p.lr)

Gradient Boosted Decision Tree with xgboost

Following the script above,

# GBDT with xgboost

library(xgboost)

cv.g.gdbt <- xgboost(m.train, ipinyou.train$IsClick, max.depth=7, eta=0.1,
  nround = 100, objective = "binary:logistic", verbose = ifelse(interactive(), 1, 0))
p.lm <- predict(cv.g.gdbt, m.test)
glmnet::auc(ipinyou.test$IsClick, p.lm)

Per-Coordinate FTRL-Proximal with $L_1$ and $L_2$ Regularization for Logistic Regression

The following scripts use an implementation of the FTRL-Proximal for Logistic Regresion, which is published in r citep(bib[["DBLP:conf/kdd/McMahanHSYEGNPDGCLWHBK13"]]), to predict the probability (1-step prediction) and update the model simultaneously.

source(system.file("ftprl.R", package = "FeatureHashing"))

m.train <- hashed.model.matrix(f, ipinyou.train, 2^16, transpose = TRUE)
ftprl <- initialize.ftprl(0.1, 1, 0.1, 0.1, 2^16)
ftprl <- update.ftprl(ftprl, m.train, ipinyou.train$IsClick, predict = TRUE)
auc(ipinyou.train$IsClick, attr(ftprl, "predict"))

If we use the same algorithm to predict the click through rate of the 3rd season of iPinYou, the overall AUC will be 0.77 which is comparable to the overall AUC of the 3rd season 0.76 reported in r citep(bib[["zhang2014real"]]).

Supported Data Structure

  • character and factor
  • numeric and integer
  • array, i.e. concatenated strings such as c("a,b", "a,b,c", "a,c", "")

Reference

bibliography(style="text")