2012-12-28

Authors in a Markov matrix Part 2 (3) Experimental results: Which author do people find most inspiring?


In this article, I will explain about our implementation and how the adjacency matrix looks like.

Implementation

We implemented the following programs:

  • Link_Vector_Extractor: Generate the author vector from the data.
  • Graph_Extractor: Generate the adjacency matrix from the data and the author vector.
  • Page_Rank: Compute page rank.
  • Remapper: Re-map the author vector according to the page rank result.

Our experiment is done with the computer CPU: Intel(R) Core(TM) DUO CPU P8400, 2 Cores, OS: 64bit Linux 3.2.0.32, Kubuntu 12.04. We used the following software: Python 2.7.3, Beautiful Soup 4.0.2, matlab R2006a, octave 3.2.4.

Adjacency matrix

Figures 2, 3, 4 show the adjacency matrices. In these Figures, blue points represent the connection.
Figure 2: Adjacency matrices. Top to bottom: German authors in de. wikipedia.org, en.wikipedia.org, ja.wikipedia.org.

Figure 3: Adjacency matrices. Top to bottom: English authors in de. wikipedia.org, en.wikipedia.org, ja.wikipedia.org.
Figure 4: Adjacency matrices. Top to bottom: Japanese authors in de. wikipedia.org, en.wikipedia.org, ja.wikipedia.org.
In Figure 2, German author, en.wikipedia.org has a regular pattern.  We haven't check what exactly causes this, however, we think this is the same problem of the template bias (will be discussed later) (Note 1). The German author in en.wikipedia.org has another peculiarity that we can see the higher number of links compare to other Wikipedia data shown in Table 2. We count the number of links is how many links are connected to the author vector entry, not all the links in the page since these are the connection in the adjacency matrix. For example, we didn't count the links to other language pages, self links, author links that were not on the author vector. Therefore, if the page has links to non-author person, we didn't count these links.
Table 2: Matrix size, non zero elements, and the average number of links between authors. Wiki ``en'' means en.wikipedia.org.

We removed rank sink nodes according to the PageRank algorithm [1]. We removed links related rank sink node pages, therefore we also removed so called dangling links (2.7 in [2]), since the links refereeing single sink rank node are the dangling links.  Additionally, we removed pages that has only outgoing links. In the original PageRank paper [2], nodes that have only outgoing links are not mentioned, though this doesn't make sense in the PageRank concept. However, it is easy to imagine the reason why the paper didn't mentioned about them. Since the original paper is for the web, it is hard to determine a node is not linked from any pages. To determine this, all the pages are considered. But, this is practically impossible on the web, therefore it is natural that the original paper didn't consider this outgoing link only nodes. However, our author vector has a limited size and we know all valid pages, therefore, we can remove these nodes. The link normalization value of PageRank calculation depends on whether considering these links or not. This means, it may affect the absolute value of the PageRank. However we are not interested in the absolute value of the PageRank. We are rather interested in the relative value, which author's influence is larger. The PageRank paper also mentioned this normalization has not large effect since they are also interested in the relative influence.

The matrix size is reduced by this operation, the result size is shown in Table 3.
Table 3: Adjacency matrices: original size, reduced size, and its rank.

Finally, I will show you who is the most important author in the next article (in a Wikipedia sense).

Note 1: When I wrote this article in my blog, my friend, Joerg M. pointed out this possibility (2012-12-29). Thanks.

No comments: