DAMICORE

DAMICORE is an easy-to-use clustering and classification tool.

DAMICORE is a pipeline of algorithms which is agnostic to the type of data thanks to NCD, a compressor-based metric which views any piece of data as simply a string of bits. This is particularly well-suited for heterogenous datasets, datasets with difficult characteristics extraction and text datasets.

 Data    Metric      Distance matrix    Simplification Phylogenetic tree
-----.           .---------------------.              .-----------------.
 x_1 |           |     x_1 x_2 ... x_n |              |   x_1-.  x_n    |
 x_2 |  .-----.  | x_1 d11 d12     d1n |    .----.    |        \  |     |
 x_3 |->| NCD |->| x_2 d21 d22     d2n |--->| NJ |--->|         O-O---- |-->
 ... |  '-----'  | ...                 |    '----'    |        /        |
 x_n |           | x_n dn1 dn2     dnn |              | x_2---'         |
-----'           '---------------------'              '-----------------'

        Community detection        Clusters
-----.                        .-----------------.
     |                        |  ___       ___  |
     |    .-------------.     | |x_1|     |   | |
tree |--->| Fast Newman |---->| |x_2| ... |C_m| |
     |    '-------------'     | |x_n|     |   | |
     |                        | '---'     '---' |
-----'                        '-----------------'

Dependencies

Installation

• python 2.7
• igraph library (visit http://igraph.org/c for more informations)
• python-igraph (visit http://igraph.org/python for more informations)

Execution

DAMICORE relies on compressors to calculate the distance between a pair of objects (files).

• gzip (available in almost all *nix systems)
• bzip2 (available in almost all *nix systems)
• ppmd (available via packages or at http://ctxmodel.net/files/PPMd)

Recommended

• FigTree (available in http://tree.bio.ed.ac.uk/software/figtree/)

Usage

The simplest way to use DAMICORE is to provide as argument a directory containing all files to be compared:

 $ ./damicore.py examples/texts

This outputs a list of pairs of file name and corresponding cluster index. For now we are lacking a good tool to visualize this clustering, but there are other tools that might help. We can output intermediate steps into different files for analysis:

 $ ./damicore.py examples/texts --ncd-output texts.phylip --format phylip \
 --tree-output texts.newick --graph-image texts.png --output texts.clusters

This outputs the NCD matrix using PHYLIP format, the neighbor-joining tree in Newick format (readable by FigTree), an image with a graph visualization and the final clusters into another file.

By default, the script uses PPMd as compressor. If you don't have it installed, try using gzip or bzip2:

 $ ./damicore.py examples/texts --compressor gzip

For more information on available options, see --help.

Contact

If you believe you have found a bug, or would like to ask for a feature, please inform me at [email protected].

Known TODOs

Lots of things to do! Among them:

• Implement UPGMA as tree joining strategy
• Use other community detection algorithm (Girvan-Newman, for example)
• Create/find a cladogram layout for graph visualization
• Include compressors commonly available for Windows (WinRAR, 7zip, etc.)
• Implement classification tool - receives a training dir and a test dir to classify ** Also, automate the sampling process and implement k-fold cross-validation
• Implement verbosity level
• Implement other simplifications strategies that produce a sparse graph from the distance matrix instead of a tree (kNN, for example)

License

This software is licensed under the GPLv2.

Texts examples dataset collected by Francisco José Monaco ([email protected])