SVMlightの特徴量の重み

svmligh.read-features-weight.coffee

#!/usr/bin/env coffee

# 特徴量の重みを見る
# [LIBSVMの特徴量の重みを見る - LIBSVMのモデルの読み方 - 唯物是真 @Scaled_Wurm](http://sucrose.hatenablog.com/entry/2013/07/10/223952)
#
# SVM^{light} の場合
#

fs = require 'fs'
model = process.argv[2]

unless model
  console.warn 'coffee ./% <model-file-by-svmlight>'
  process.exit 0

fs.readFile model, 'utf8', (err, data) ->
  throw err if err
  ls = data.trim().split '\n'
  len = ls.length
  k = 0 # SV line
  for i in [0 ... 40] # 高々40行目までにあるはず
    if ls[i].indexOf('is a SV') isnt -1
      k = i + 1
      break

  weight = {}
  add = (id, x) ->
    if weight[id]
      weight[id] += x
    else
      weight[id] = x

  for i in [k ... len]
    xs = ls[k].split ' '
    coef = +xs[0]
    m = xs.length
    for j in [1 ... m]
      continue if xs[j] is '#'
      [id, val] = xs[j].split ':' # id:value
      add id, coef * (+val)

  items = for id of weight
    { id: id | 0, w: Math.abs weight[id] }
  items.sort (a, b) -> b.w - a.w
  console.dir items

# vim: set ft=coffee:

参考; LIBSVMの特徴量の重みを見る - LIBSVMのモデルの読み方 - 唯物是真 @Scaled_Wurm