Add vw_ensemble backend

[osma]

Fixes #235

This PR adds a new Vowpal Wabbit based learning ensemble backend. It is an alternative for PAV. The implementation is based on ideas from the Annif-fusion experiments.

This is a draft PR. Some TODOs:

• discounting mechanism (increase trust in the model with more training data)
• refactor and reduce duplication w.r.t. vw_multi backend (maybe extract a common base class for VW based backends)
• coordinate with PR Issue277 auto generate api documentation #280: need to make sure documentation is autogenerated for the new backend

[codecov]

Codecov Report

Merging #284 into master will increase coverage by 0.04%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master     #284      +/-   ##
==========================================
+ Coverage   99.32%   99.36%   +0.04%     
==========================================
  Files          52       55       +3     
  Lines        2651     2842     +191     
==========================================
+ Hits         2633     2824     +191     
  Misses         18       18

Impacted Files	Coverage Δ
annif/backend/mixins.py	93.75% <ø> (-0.37%)	⬇️
tests/test_project.py	100% <ø> (ø)	⬆️
annif/backend/tfidf.py	100% <ø> (ø)	⬆️
annif/backend/ensemble.py	100% <100%> (ø)	⬆️
tests/test_backend_vw_ensemble.py	100% <100%> (ø)
annif/backend/__init__.py	100% <100%> (ø)	⬆️
annif/backend/vw_ensemble.py	100% <100%> (ø)
annif/backend/vw_base.py	100% <100%> (ø)
annif/backend/vw_multi.py	100% <100%> (ø)	⬆️
annif/backend/backend.py	96.55% <100%> (+0.12%)	⬆️
... and 3 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 09c4084...70d3973. Read the comment docs.

[osma]

The Travis build breaks on Python 3.5 (the only Python version where VW is available). Apparently there's a problem with the unit test expecting the tfidf-fi backend to be initialized, when it's not.

[osma]

Still TODO:

• discounting mechanism
• refactor _create_model - the code is practically identical in vw_ensemble and vw_multi, so the method should be moved to vw_base, but need to separate out the default options for each type as they are a bit different
• auto-generation of API docs for the new modules
• remaining quality issues and ensuring test coverage

[osma]

This pull request introduces 7 alerts when merging 9f25308 into f9b1294 - view on LGTM.com

new alerts:

• 4 for Unused import
• 3 for Modification of parameter with default

---

This project has automated code review enabled, but doesn't use the LGTM GitHub App. Migrate over by installing the app. Read about the benefits of migrating to GitHub Apps in the blog.

---

Comment posted by LGTM.com

[lgtm-com]

This pull request introduces 6 alerts when merging 34dc5a4 into f9b1294 - view on LGTM.com

new alerts:

• 3 for Unused import
• 3 for Modification of parameter with default

[osma]

Feature-wise everything should be done now. Test coverage is 100%. API docs are updated. The remaining quality issues can be ignored (most are false positives).

[osma]

I've done some testing on real data sets and although the results are not quite as good as I hoped quality-wise, the backend seems to be working well now after the latest fixes.