Wiki Wormhole

Description

This python library leverages graphs and artificial intelligence to find the path from a start wikipedia page to a destination wikipedia page.

Installation

Install wikiwormhole through pip package manager.

pip install wikiwormhole

Then run the setup script to accomplish the following:

• Generate config.yaml script for the pageviews api (How many pageviews does a wikipedia page have).
• Download pre-trained Word2Vec model using Gensim.
• Download corpus of stop words (insignificant words) to help embed titles using natural language toolkit (NLTK).

python <venv-name>/bin/setup_wormhole.py <config-path> <download-path>

Here config path is where the new config-yaml will be created. The download path is where the Word2Vec model weights and NLTK corpus will be downloaded on your system.

The final step in installation is providing a personal website and email in the config.yaml file. Once filled you'll be able to use the pageviews API.

Algorithm

The algorithm is split into two parts: popular and similar traversal.

Popular Traversal is an algorithm that attempts to find a more popular page connected to the original page. In the small-world phenomenon it's observed its much easier to connect two separate "hubs" rather than two obscure nodes. The purpose of this algorithm is to find those popular hubs.

Algorithm:

Explore(Node, Graph)
  for Link in Node do
    Graph.add(Node)

PopularTraversal(InitialPage)
  G <- Graph()
  Trace <- [InitialPage]
  Explore(InitialPage, Graph)
  for T iterations do
    ExploreNodes <- Select K random unexplored pages
    for Node in ExploreNodes do
      Explore(Node, Graph)
    Popular = find node with most connected edges not in Trace
    Trace.add(Popular)

  MostViews = (PopularPage, Views)
  for PopularPage in Trace do
    PageViews <- HowManyPageViews(PopularPage)
    if PageViews > MostViews then
      MostViews = (PopularPage, PageViews)

  return MostViews[0]

Similar Traversal is an algorithm that attempts to find the path of pages from the starting page to the destination page. The algorithm assumes that titles with similar words will be closer to each other in the graph than the contrary. This similarity calculation is powered by the infamous word embedder Word2Vec. The word embedder calculates the words position in a latent space and the similarity is then calculated for two separate words using a cosine similarity metric (emphasis on directional alignment rather than distance).

Algorithm:

SimilarTraversal(InitialPage, TargetPage)
  PQ <- PriorityQueue() - only stores top x nodes
  G <- Graph(InitialPage)
  PQ.add(InitialPage, Similarity(InitialPage, TargetPage))

  while TargetPage not in G do
    SimNode <- PQ.pop()
    for LinkNode in SimNode do
      G.add(SimNode, LinkNode)
      PQ.add(LinkNode, Similarity(LinkNode, TargetPage))

  return G.path(InitialPage, TargetPage)

Similarity(Title1, Title2)
  Title1 <- Embed(Title1)
  Title2 <- Embed(Title2)
  return CosineSimilarity(Title1, Title2)

Sample Usage

from wikiwormhole.traverse.popular import PopularTraverse
from wikiwormhole.traverse.similar import SimilarTraverse
from wikiwormhole.title2vec import Title2Vec
from tqdm import tqdm

data_dir = './data'
config_path = './config.yaml'

start_title = 'Apple'
end_title= 'Cadillac'

pop_rounds = 5

pop_start = PopularTraverse(start_title, config_path)
pop_end = PopularTraverse(end_title, config_path)

for _ in tqdm(range(pop_rounds)):
    pop_start.traverse()
    pop_end.traverse()

path_start = pop_start.most_popular_pathway()
path_end = pop_end.most_popular_pathway()
print("START: ", path_start)
print("END: ", path_end)

t2v = Title2Vec(data_dir)
sim_start, sim_end = path_start[-1], path_end[-1]
sim_trav = SimilarTraverse(sim_start, sim_end, t2v)

while not sim_trav.target_reached():
    sim_trav.traverse()

path_cnx = sim_trav.path_to_target()
print("RESULT: ", path_start[:-1] + path_cnx + path_end[:-1][::-1])