jdnc / ml-project Goto Github PK

Captures some of the future directions for this project, mostly enhancements

• Impact of transfer learning :
  Show that the transfer learning approach improves the prediction performance as compared to current neurosynth baseline
• Impact of label correlations:
  Test performance improvement for e.g. label-pair classifiers vs. single label classifiers.
• Hierarchical classification approach :
  Split prediction into two steps 1. prediction of topic loadings (e.g using regression) 2. prediction of words conditional on the topic loadings. There is precedent in the literature for such an approach

Strangely, there is no write to the scratch data folder for storing the confusion matrix array from naive bayes, though this seemed to be working fine two days ago

Need to look further into this - program seems to be hanging?

Refactor code in most branches. Too redundant - most code remains the same across files with small changes. Allow these to be passed as configuration options indeed.

The decision tree code throws index out-of-range error when evaluated using scoring functions from utils.py.

E.g.:
File "/scratch/02863/mparikh/utils.py", line 253, in score_results
s = Sp[n]
IndexError: index 22 is out of bounds for axis 0 with size 22

Need to add tests corresponding to each new module. Also some functions have changed, need to review their docs.

Currently starting with the 25 terms used in the original paper - but it would be good to pick out which terms are the most useful as labels from the entire set in the multi-label classifier.

cool concept suggested by Sanmi for dealing with correlated labels in multi-label case

http://en.wikipedia.org/wiki/Classifier_chains

openfMRI and neurosynth masks do not match - try the technique suggested by sanmi
(Use mask 'a' to mask mask 'b' and then proceed further)

Seems to be finding just nine common studies after the code was tweaked today morning. Reviewing latest commits

Since big data files are removed, this no longer works for experiment.py filter_studies_voxels. Must be fixed