ltflores / csc-869-mlog Goto Github PK

Change SentenceBasedTextDirectoryLoader and TextDirectoryLoader to ignore .svn 
directory and contained files

Play around with different settings of the StringToWordVector:

- Different tokenizer
- Different stemmer
- Different stopword list or no stopword list
- Different min. word freq.
- Different num. of words to keep
- Different pruning
...

Document how each of those things influence the results.

Best would be to have some normal people => anybody personally knows people who 
knows twitter and would be willing to share if they are rep./democrat?

Other possibility: Check out organizations whose members are strongly 
affiliated with either republican or democrat (e.g. NRA, ...)

Implement different weighing schemes and try out the ones implemented in 
StringToWordVector. E.g.:

- In-party-frequency / Total-frequency
- Higher/Lower weight of special attributes like #hashtag or @user
...

Exclude attributes added by the SentenceBasedTextDirectoryLoader.java like 
democrats_03152011/JamesMcDermott or republicans_03152011/MarioDiaz-Balart from 
the classification. Optionally just disable the generation of them as they're 
not used at the moment anyways

As Lorenzo already did with J48 add new classifier to the app (SVM, kNN, ...) 
and document the results

From Orens email:

one quick question, in the 10-fold cross validation, did we make sure
that there are no shared people between two sections? i mean, if we
just divide into 10 sections according to tweets, then we may have
tweet1 and tweet2 of the same congressmanA in two different sections.
in this case, we may get a good result in the cross validation simply
because the classifier can find similarity between tweets of
congressmanA (e.g. if tweet1 is in the verification section, and
tweet2 is in one of the 9 training sections, it may simply learn that
tweet1 and tweet2 are similar in language and we'll get a good
misleading score...).

This needs to be fixed by custom code to distribute.

Some classifier have an unacceptable runtime on the whole data set for fast 
development (e.g. J48 = 12h on Lorenzos machine) => Implement a command-line 
argument which cuts down the input size to speed up development