fanghgit/xmc

experimental code for XMC

Experimental Code for XMC

Description

This repo currently includes the following folders:

• data: folder to hold experimental data, currently only have EURLex
• util: contains code for preprocessing and evalution (some are from DISMEC)
• dismec: DISMEC's code
• linesearch: liblinear whose L2L2 is solved by Newton method with line search. linesearch in primary built on the code of [3]

Quick Start

If the data is download from XMC website, remove the first line.

Run the following code to remap features and labels.

$ python util/remap.py data/eurlex/train.txt data/eurlex/test.txt data/eurlex/train_remap.txt data/eurlex/test_remap.txt -r 1

Do if-idf transformation:

$ python util/transform.py data/eurlex/train_remap.txt data/eurlex/test_remap.txt data/eurlex/train_remap_tfidf.txt data/eurlex/test_remap_tfidf.txt

Go to linesearch_parallel folder and type make to build executable.

Create folder to hold models: mkdir linesearch_parallel/models

Parameter options:

  -s : type of solver, default to be L2L2 primal solver
      -0: L2R_LR
      -1: L2R_L2LOSS_SVC
      -2: L1R_L2LOSS_SVC
      -3: L1R_LR
      (default 1)
  -x : initialization
      -0: naive 0 initialization
      -1: all negative initialization
      -2: MST initialization
      (default 0)
  -B : bias
      (default 0)
  -P : number of threads
      (default 1)
  -e : stopping criterion
      |f'(w)|_2 <= min(0.0001, eps*min(pos,neg)/l ) *|f'(0)|_2
      (default 1.0)

Training with default initialization for EURLex:

$ ./linesearch_parallel/train -B 1 -P 1 -e 1.0 0.0001 data/eurlex/train_remap_tfidf.txt linesearch_parallel/models/eu1.model

Training with all negative initialization for EURLex:

$ ./linesearch_parallel/train -B 1 -P 1 -x 1 -e 1.0 0.0001 data/eurlex/train_remap_tfidf.txt linesearch_parallel/models/eu2.model

Training with MST initialization for EURLex:

$ ./linesearch_parallel/train -B 1 -P 1 -x 2 -e 1.0 0.0001 data/eurlex/train_remap_tfidf.txt linesearch_parallel/models/eu2.model

By using -x 1 or -x 2, training should be able to finish in around 200 sec.

Remark: -B 1 or -B 0.1 will not change prediction accuracy, but it will affect model size significantly, -B 0.1 will make our algorithm faster since the # of iterations relies on $$|x|_2$$, but -B 0.1 will increase the model size.

Stopping criterion

We stop when

|f'(w)|_2 <= eps*min(pos,neg)/l*|f'(w0)|_2

and

|f'(w)|_2 <= 0.0001*|f'(w0)|_2

where f initializations the primal function and w0 = zeros(n,1) and pos/neg are # ofpositive/negative data. pos/neg is very small for most classes, so we set eps = 1.0 or eps=0.1 for experiments.

DISMEC's code is changed to use the same stopping criterion.

Prediction: create new folder mkdir linesearch_parallel/output

$ ./linesearch_parallel/predict data/eurlex/test_remap_tfidf.txt linesearch_parallel/models/eu2.model linesearch_parallel/output/eu2.out

Training with DISMEC:

$ ./dismec/dismec/train -B 1 -s 2 -e 0.0001 data/eurlex/train_remap_tfidf.txt dismec/dismec/models/eu_dismec.model

$ ./dismec/dismec/predict data/eurlex/test_remap_tfidf_zeroed.txt dismec/dismec/models/eu_dismec.model dismec/dismec/output/eu_dismec.out

$ ./python util/evaluate.py data/eurlex/GS.txt dismec/dismec/output/eu_dismec.out

References:

[1] Rohit Babbar, Bernhard Shoelkopf, DiSMEC - Distributed Sparse Machines for Extreme Multi-label Classification, 2017

[2] R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and C.-J. Lin. LIBLINEAR: A library for large linear classification, 2008

[3] C.-Y. Hsia, Y. Zhu, and C.-J. Lin. A study on trust region update rules in Newton methods for large-scale linear classification, 2017