todo

使用具有增量学习的监控式学习方法。包括几个不同的分类算法。

Bugs and small fixes
~~~~~~~~~~~~~~~~~~~~
selectd:
- Fix folder names -> class bindings:
  When adding new folders, they must never get the number of an unnamed class.
  Add command: C> "get:folder:5\n" S> "spam\n"
  Maybe allow classification results with folder names instead of numbers.

select_mailstat:
- Hardcoded configuration; should be configurable with command line options.

pifiled:
- Program name and options should be configurable with command line options.
- Limited buffer size.

cygwin:
- Signals apparently don't work very well, need some other way to quit
  selectd and save classifier databases.
  Probably need some gui (taskbar icon) to do this in a good way.

Misc:
- Potential buffer overruns in protocol_*.c.
- Error checking when saving and loading files.
- When failing a memory allocation, make sure the database is saved
  (at least tried to be) before quitting. (via a signal or maybe atexit)
- Add options to tokenizer to select minlen, maxlen, lower, unicode/byte etc.


High priority features
~~~~~~~~~~~~~~~~~~~~~~
Better clients.

Better classifiers.

Improved feature selection abilities.


Middle priority features
~~~~~~~~~~~~~~~~~~~~~~~~
Tokenizer which ignores html-tags.

ECOC with random codes so that they can grow.

Collect simple performance statistics in selectd.


Low priority features
~~~~~~~~~~~~~~~~~~~~~
Make thread classifier that uses thread information from headers.

Ability to take timings of training and classification.

Make unlearning work better.

Leave-one-out mode in select_test.