User Modeling in Folksonomies: Relational Clustering and Tag Weighting

Reference

Takuya Kitazawa and Masahide Sugiyama User Modeling in Folksonomies: Relational Clustering and Tag Weighting. In Proc. of WIMS'15, July 2015. [slide]

The Infinite Relational Model

Our paper deeply refers the Infinite Relational Model (IRM) proposed by C. Kemp et al. Source code developed by the author is available at http://www.charleskemp.com/code/irm.html

Experiments in this study consider dataset described below, with applying this source code. IRM-based relational clustering iteratively works as:

Unfortunately, the code may not work (segmentation fault occurs) under incompatible environments. We guess that the program has some minor issues around memory management. For reference, our environment is:

Vine Linux 6.3 (32bit)
Intel(R) Core(TM) i7-2600 CPU @ 3.40GHz
4GB RAM

Dataset

Dataset used in this paper is available at: https://drive.google.com/open?id=1sSXm24tHHSm1TV8XNDyBLnDFKoKH5k0z

Open the .zip file, and you can find:

.
├── 5-fold
│   ├── output
│   ├── test
│   └── train
├── IRM_output
│   ├── page_assignments
│   ├── status
│   └── user_assignments
├── bookmarks.graph
└── pages.json

bookmarks.graph is an input to IRM-based relational clustering, and outputs are in a directory named IRM_output. For more detail about how to understand the outputs, please refer README file of the IRM code.

pages.json includes complete information of 7,000 web pages as follows. Note that, since the authors focused on Hatena Bookmark, a Japanese social bookmarking service, most of data is written in Japanese.

[
  {
    "title": "${web page title}",
    "url": "${url}",
    "tags": {
      "${tag #1}": ${frequency of tag #1},
      "${tag #2}": ${frequency of tag #2},
      "${tag #3}": ${frequency of tag #3},
      ...
    },
    "bookmark": ${number of bookmarks}
  },
  {
    ...
  }
]

Importantly, ordering in pages.json corresponds to indices of the third column in bookmarks.graph.

Code

We disclose a python script weighting.py to generate the same results as Fig. 7 and 8 in our paper. This script reads the IRM outputs and page information, and compute overall tag weights based on the equations shown in section 4.1.

$ python weighting.py

The program has eta(9, ell), strength between each web page cluster and 9-th user cluster, as a list. These values are pre-computed based on the IRM outputs.

Also, recommend.py will make recommendations and evaluate based on F-measure and running time, as discussed in Section 5.

$ python recommend.py

This code reads outputs from each cross validation step, and evaluates by using test data. Finally, average F-measure and running time will appear. Note that it will run for quite a long time, especially during user-based collaborative filtering (CF). The implementation of CF is closed because it was written by the other programmer.