while / ftrlproximal Goto Github PK

Running valgrind on a test script:

R -d valgrind --vanilla < tests/memcheck/norm2d.r

identifies a memory leak

==5301== 
==5301== HEAP SUMMARY:
==5301==     in use at exit: 102,506,744 bytes in 58,522 blocks
==5301==   total heap usage: 102,898 allocs, 44,376 frees, 189,628,726 bytes allocated
==5301== 
==5301== LEAK SUMMARY:
==5301==    definitely lost: 24,240 bytes in 1,010 blocks
==5301==    indirectly lost: 48,480 bytes in 3,030 blocks
==5301==      possibly lost: 0 bytes in 0 blocks
==5301==    still reachable: 102,434,024 bytes in 54,482 blocks
==5301==         suppressed: 0 bytes in 0 blocks
==5301== Rerun with --leak-check=full to see details of leaked memory
==5301== 
==5301== For counts of detected and suppressed errors, rerun with: -v

Add support for ff tables that does not fit in memory. One of the benefits with FTRLProximal is that you do not need to have all data in memory so it should support training models on on disk data. ff is a good contender for this support.

It would be really nice w a built in cv function to validate parameters similar to the glmnet package has.

It's awesome to find such online learning algorithms package on R-Cran.
I try to analyze data from Kaggle for practice, and the returning prediction is either 0 or 1.
Even I set predict(, type='response'), the prediction is fixed as 0/1.
If I remember it correctly, the core of FTRL-Proximal is still logistic regression for application on CTR.
I notice one possible cause may be that when training the ftrl model by using ftrlprox() function, it requires the dependent variable should be factor type.
Is it possible to get the prediction being probabilities (0~1, like logistic regression) rather than binary type?