Political machine learning

After being rather disappointed with the recent election, coming to grips with Jim Inhofe taking control again of the Senate Environment Committee, I've been wondering what I can do for future elections.

In the wake of the elections, NPR featured a former Bush (W) administration official gladly claiming that America had spoken clearly that they didn't like that the average family income had gone down by 2600$ per year since 2007. This infuriated me. It left everyone to draw their own conclusions, but one was obvious: it is Obama and the Democrats' fault for the decrease. Well, let's think for a moment - what happened in 2007? Some kind of economic crash? Was his reference point before or after the crash?

This kind of news is exceedingly misleading, and I am infuriated by how often this kind of thing influences people. We only ever get news with selective context, never the whole story. Anyone wanting the whole story needs to find it themselves, and that's really hard. There are some reliable journalistic sources, but people generally go to the loudest source, or to the one they trust most (often one politically aligned with them).

Enter my newest idea: can you come up with a machine learning system that is capable of aggregating facts, classifying news articles based on their usage of facts, and generally providing a tool for enlightening the public?

I've come up with a few ideas of what this system might do:

Monitor news sources; aggregate them into meta-pages on topics, sort of like Wikipedia summaries, with links about where content originated to read more.
Expand contexts for all numeric figures stated. People always choose whatever contexts push their agenda. This system should be able to restate facts/statistics in many ways (at least whatever ways the aggregate news source collects), and also be able to make simple inferences about related events (such as average annual salary going down since 2007 <-> recession).
Provide something like a browser extension to overlay sites with ratings/warnings about slanted facts. (perhaps something like this: http://link.springer.com/chapter/10.1007%2F978-3-540-77485-3_11)
Provide lobbying/contribution information to examine which interests are involved in events/decisions (see the already excellent browser extension, Greenhouse: http://allaregreen.us/)

Whether anyone would read such a thing is another question, but the burden on people to collect facts outside of sound bites is currently too great. The opportunity for voting based on reason rather than emotion is limited as such.

Next: how to get more people to vote.

CMake and HDF5 revisited

Thanks to JC of Kitware at SciPy 2014, the situation with CMake and HDF5 on windows has greatly improved.

First, there's been a new release of HDF5 (1.8.13). This release removes a buggy CMake library finder (targets) file. The improperly generated file in HDF5 1.8.12 was thehdf5-targets-release.cmake file. Go get the new release now: http://www.hdfgroup.org/HDF5/release/obtain5.html

Secondly, there is a bug in CMake's builtin script for finding HDF5. On Windows, it finds a DLL before the Lib file, and won't link properly (http://public.kitware.com/Bug/view.php?id=14111). Thanks to JC, I understand that the way forward is not to patch that file, but ratherto move towards removing CMake's builtin scripts entirely, and push the responsibility of finding these libraries out to the library authors. Several library authors have already done this (including HDF5, thankfully). The included FindSomething.cmake scripts should be treated as a legacy, fallback approach. CMake has a great option for forcing the use of the library's scripts for defining libraries and headers. When calling the find_package function, you need to specify the NO_MODULE argument, like so:

FIND_PACKAGE(HDF5 "1.8.5" REQUIRED NO_MODULE)

This bypasses much of CMake's internal FindHDF5 script, and instead goes looking for hdf5-config.cmake in several locations, ideally finding it in the place we installed the HDF5 libraries. The new version of the HDF5 libraries now properly tells CMake where to find everything, and life is peachy.

In summary, use NO_MODULE when the library you're using provides a -config.cmake file. It will likely be better (more current) than CMake's internal Findsomething.cmake file. However, be careful! If not all platforms actually provide this file, then your build will break where this file does not exist. Notably, Ubuntu does not (currently) ship the hdf5-config.cmake files with the libhdf5-dev package. So, you'll need a code section like this (based onhttps://github.com/commontk/CTK/blob/master/Utilities/CMake/FindDCMTK.cmake):

# back up any user-defined HDF5_DIR setting set(_SAVED_HDF5_DIR ${HDF5_DIR}) FIND_PACKAGE(HDF5 "1.8.5" NO_MODULE) # First look for hdf5-config.cmake in defined locations. This file is generated by HDF5 team and is likely more up to date than FindHDF5.cmake included in CMake. IF(NOT HDF5_FOUND) # didn't find hdf5-config.cmake. Need to fall back on CMake's built-in FindHDF5.cmake logic. # required says we fail if it isn't found. # Restore the value reset by the previous call to 'find_package(HDF5 "1.8.5" NO_MODULE)' set(HDF5_DIR ${_SAVED_HDF5_DIR} CACHE PATH "HDF5 install dir" FORCE) FIND_PACKAGE(HDF5 "1.8.5" REQUIRED) ENDIF()

More information:
details on find_package: http://www.cmake.org/cmake/help/v3.0/command/find_package.html?highlight=find_package
Creating your own package descriptions to make includes and libraries easy for downstream users:

Thanks again to JC for helping me figure this stuff out - now I need to get QSTEM using it!

Political machine learning

Post Footer

Older · View Archive (62)

CMake and HDF5 revisited

Newer

Gitting rid of merge commits from git pull