yalesong / hcrf-light Goto Github PK

Dear Yale,

Looks like this light version HCRF does not support window size for dependency range. Currently just consider consecutive nodes, right? Or is there a way to specify such window size in your current code? Thanks a lot.

Hi Song, I know it's been some time since this paper is published. I'm trying to run the code for OCCRF model for a few days but I couldn't figure out the correct configuration. FYI, I could run the hcrf model in README.md and I hope to have some guidance on running the occrf model.

Here's what I'm trying:

./hCRF-light/distribute/bin/hcrf-light -m occrf \
-Fd ./hCRF-light/data/toy/dataAnomalyTrain.csv -Fl ./hCRF-light/data/toy/labelsAnomalyTrain.csv \
-FD ./hCRF-light/data/toy/dataAnomalyTest.csv -FL ./hCRF-light/data/toy/labelsAnomalyTest.csv

And here's what I'm getting:

L-BFGS optimization terminated with status code = -1000
fx = 0.0151543

Reading testing set... 
Starting testing ...
Press a key to continue...

Here's the cat stats.txt. Looks like it's predicting all data points in test set to be normal (label 0).

I noticed that the status code is -1000, and I've checked the lbfgs library and this error code links to LBFGSERR_ROUNDING_ERROR. I'm not sure why this is happening, I'm guessing that the gradient became too small during optimization.

Then, I try to solve it by changing to use nrbm optimizer by adding -o nrbm, and here's the cat stats.txt result. It's predicting all rows in test set as anomalies:

(1) I'll be very pleased if you would guide me on the configuration of occrf for the toy anomaly dataset / other dataset. 😄

(2) I also have another question in mind: Shouldn't occrf be predicting just one instead of multiple labels per sequence? According to the publised paper, there is only 1 anomaly score per employee. But in this code, it's considered a continuousModel, and uses fl & fL (1 label / timestep) instead of fq & fQ (1 label / sequence)...

Thank you! 😃

When I ran this library with a dataset of 4200 files. Dimension of seqs data(93 columns and rows [not fixed]) It gave me accuracy to be somewhere around 0.0015 which is very low. I gave the split valid parameters to be same as split test. Can anyone explain the reason for this low accuracy?
Do I need to change seeds or any other parameter?

I followed the instructions in compile.sh and got no error.
After make and make distribute, I tried to run the command
./distribute/bin/hcrf-light and got the following error :
./distribute/bin/hcrf-light: error while loading shared libraries: liblbfgs-1.8.so: cannot open shared object file: No such file or directory
The library liblbfgs-1.8.so indeed exists in ./lib/liblbfgs/lib/.libs/. How can I solve this problem ?

I am using Debian 8.7.

When I run the file 'test_oc.m', I get the error "Undefined function or variable 'matHCRF'." How could I solve this problem?