[Numpy-discussion] A first proposal for dataset organization

[David Cournapeau]

Hi there,

    A few months ago, we started to discuss about various issues about 
dataset for numpy/scipy. In the context of my Summer Of Code for machine 
learning tools in python, I had the possibility to tackle concretely the 
issue. Before announcing a first alpha version of my work, I would like 
to gather comments, critics about the following proposal for dataset 
organization.

The following proposal is also available in svn:

http://projects.scipy.org/scipy/scikits/browser/trunk/learn/scikits/learn/datasets/DATASET_PROPOSAL.txt


Dataset for scipy: design proposal
==================================

One of the thing numpy/scipy is missing now is a set of datasets, 
available for
demo, courses, etc. For example, R has a set of dataset available at the 
core.

The expected usage of the datasets are the following:

        - machine learning: eg the data contain also class information 
(discrete or continuous)
        - descriptive statistics
        - others ?

That is, a dataset is not only data, but also some meta-data. The goal 
of this
proposal is to propose common practices for organizing the data, in a 
way which
is both straightforward, and does not prevent specific usage of the data.

Organization
------------

A preliminary set of datasets is available at the following address:

http://projects.scipy.org/scipy/scikits/browser/trunk/learn/scikits/learn/datasets

Each dataset is a directory and defines a python package (e.g. has the
__init__.py file). Each package is expected to define the function load, 
returning
the corresponding data. For example, to access datasets data1, you 
should be able to do:

 >>> from datasets.data1 import load
 >>> d = load() # -> d contains the data.

load can do whatever it wants: fetching data from a file (python script, csv
file, etc...), from the internet, etc... Some special variables must be 
defined
for each package, containing a python string:

    - COPYRIGHT: copyright informations
    - SOURCE: where the data are coming from
    - DESCHOSRT: short description
    - DESCLONG: long description
    - NOTE: some notes on the datasets.

Format of the data
------------------

Here, I suggest a common practice for the returned value by the load 
function.
Instead of using classes to provide meta-data, I propose to use a 
dictionnary
of arrays, with some values mandatory. The key goals are:

        - for people who just want the data, there is no extra burden ("just
          give me the data !" MOTO).
        - for people who need more, they can easily extract what they 
need from
          the returned values. More high level abstractions can be built 
easily
          from this model.
        - all possible dataset should fit into this model.
        - In particular, I want to be able to be able to convert our 
dataset to
          Orange Dataset representation (or other machine learning 
tool), and
          vice-versa.

For the datasets to be useful in the learn scikits, which is the project 
which
initiated this datasets package, the data returned by load has to be a dict
with the following conventions:

    - 'data': this value should be a record array containing the actual 
data.
    - 'label': this value should be a rank 1 array of integers, contains the
      label index for each sample, that is label[i] should be the label 
index
      of data[i]. If it contains float values, it is used for regression 
instead.
    - 'class': a record array such as class[i] is the class name. In other
      words, this makes the correspondance label name > label index.

As an example, I use the famouse IRIS dataset: the dataset contains 3 
classes
of flowers, and for each flower, 4 measures (called attributes in machine
learning vocabulary) are available (sepal width and length, petal width and
length). In this case, the values returned by load would be:

        - 'data': a record array containing all the flowers' 
measurements. For
          descriptive statistics, that's all you may need. You can 
easily find
          the attributes from the dtype (a function to find the 
attributes is
          also available: it returns a list of the attributes).
        - 'labels': an array of integers (for class information) or 
float (for
          regression). each class is encoded as an integer, and labels[i]
          returns this integer for the sample i.
        - 'class': a record array, which returns the integer code for each
          class. For example, class['Iris-versicolor'] will return the 
integer
          used in label, and all samples i such as label[i] ==
          class['Iris-versicolor'] are of the class 'Iris-versicolor'.

This contains enough information to get all useful information through
introspection and simple functions. I already implemented a small module 
to do
basic things such as:

        - selecting only a subset of all samples.
        - selecting only a subset of the attributes (only sepal length and
          width, for example).
        - selecting only the samples of a given class.
        - small summary of the dataset.

This is implemented in less than 100 lines, which tends to show that the 
above
design is not too simplistic.

Remaining problems:
-------------------

I see mainly two big problems:

        - if the dataset is big and cannot fit into memory, what kind of 
API do
          we want to avoid loading all the data in memory ? Can we use 
memory
          mapped arrays ?
        - Missing data: I thought about subclassing both record arrays and
          masked arrays classes, but I don't know if this is feasable, 
or even
          makes sense. I have the feeling that some Data mining software use
          Nan (for example, weka seems to use float internally), but this
          prevents them from representing integer data.

Current implementation
----------------------

An implementation following the above design is available in
scikits.learn.datasets. If you installed scikits.learn, you can execute the
file learn/utils/attrselect.py, which shows the information you can easily
extract for now from this model.

Also, once the above problems are solved, an arff converter will be 
available:
arff is the format used by WEKA, and many datasets are available at this
format:

http://weka.sourceforge.net/wekadoc/index.php/en:ARFF_%283.5.4%29
http://www.cs.waikato.ac.nz/ml/weka/index_datasets.html

Note
----

Although the datasets package emerged from the learn package, I try to 
keep it
independant from everything else, that is once we agree on the remaining
problems and where the package should go, it can easily be put elsewhere
without too much trouble.

cheers,

David