Machine Learning/Kaggle Social Network Contest/load data

R

igraph

The full dataset loaded pretty fast using the R package igraph. With the full data set loaded R is using less than 900MB of RAM.

Grab the package with:

install.packages("igraph")

Load the data using:

data <-as.matrix(read.csv("social_train.csv", header = FALSE));
dg <- graph.edgelist(data, directed=TRUE)

Note that the resulting graph contains an additional vertex with id zero. If you delete this vertex the id names will not be preserved, and so it is a good idea to just leave it in there. The vertex zero has no edges.

Python

How to load the network into networkx

There is a network analysis package for Python called networkx. This package can be installed using easy_install.

The network can be loaded using the read_edgelist function in networkx or by manually adding edges

NOTE: John found that it took up about 5.5GB of memory to load the entire network. We may need to process it in chunks - or maybe decompose it into smaller sub networks.

Method 1

import networkx as nx
DG = nx.read_edgelist('social_train.csv', create_using=nx.DiGraph(), nodetype=int, delimiter=',')

Method 2

import networkx as nx
import csv
import time

t0 = time.clock()
DG = nx.DiGraph()

netcsv = csv.reader(open('social_train.csv', 'rb'), delimiter=',')

for row in netcsv:
    tmp1 = int(row[0])
    tmp2 = int(row[1])
    DG.add_edge(tmp1, tmp2)


print "Loaded in ", str(time.clock() - t0), "s"

Below is the time to load different numbers of row using the two methods on a 2.8Ghz Quad core machine with 3GB RAM. The second method seems quicker. Note that these are just based on single loads and are intended to be a guide rather than a rigorous analysis of the methods!

Rows	1M	2M	3M
Method 1	20s	53s	103s
Method 2	15s	41s	86s

Ruby

Note on CSV Libraries

If you happen to be using Ruby (like Jared) for loading data in and out of CSV files, you should definitely try [[ http://fastercsv.rubyforge.org/ | FasterCSV ]](require 'faster_csv') instead of the stock CSV (require 'csv'). For example, when loading the adjacency list it was literally ten times faster using FasterCSV than using the normal CSV.

Loading Adjacency Lists

require 'rubygems'
require 'faster_csv'
def load_adj_list_faster(filename)
  adj_list_hash={}
  FasterCSV.foreach(filename, :quote_char => '"', :col_sep =>',', :row_sep =>:auto) do |row|
    node_id=row.shift
    list_of_adj=row
    adj_list_hash[node_id] = list_of_adj
  end
  return adj_list_hash
end

adj_list_lookup = load_adj_list_faster('adj_list.out.csv')
rev_adj_list_lookup = load_adj_list_faster('reverse_adj_list.out.csv')