FAQ

Frequently asked questions

• Is icsiboost fully compatible with boostexter?

No. The lack of documentation of boostexter make some things very obscure and difficult to guess. You will probably always get results different numerically because of implementation details.

• Given the same model (.shyp file), why do icsiboost and boostexter give different output?

They have a different behavior regarding the interpretation of the n-gram generation options. By default, icsiboost only generates unigrams even if the model contains n-grams. The default is different for boostexter.

• Does icsiboost work on the (fill in) platform?

icsiboost was only tested on linux and mac osx. It might work on other platforms (such as windows/cygwin) with some minor modifications. It does not work on solaris but could be ported with non-trivial effort (regarding the use of NaN and mmap).

• Why is icsiboost slower than boostexter?

The multi-threaded code prevents some compiler optimizations when run as a single thread. The implementation itself is not particularly optimized for speed. The code for loading files is quite generic and therefore slow.

• Why increasing the number of parallel threads does not decrease training time?

Only the exhaustive search for good weak learners runs in parallel, at the column level. This means that if you have only one feature column, you won't get any speed improvement. If each iteration is already fast, the overhead of synchronization may undermine any gain from multi-threading. And remember that loading data, computing error rates and updating example weights are not run in parallel.

• My model seems to output only the majority class.

Adaboost is not very good at finding decision thresholds when the class distribution is skewed. You can try to use the optimal_threshold.pl script to find a better decision threshold for binary problems.

• How do I get probabilities from icsiboost?

See OutputProbabilities.

• Can you implement (fill in)?

Unlikely unless it's directly in the interest of my boss.

• Can I contribute to the source code?

Yes, please contribute. Send us a clear set of patches, we'll definitely review them.