GenSVM: A Generalized Multiclass Support Vector Machine
Authors: Gerrit J.J. van den Burg, Patrick J.F. Groenen
JMLR 2016 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Rigorous numerical experiments compare linear Gen SVM with seven existing multiclass SVMs on both small and large data sets. These comparisons show that the proposed method is competitive with existing methods in both predictive accuracy and training time, and that it significantly outperforms several existing methods on these criteria. |
| Researcher Affiliation | Academia | Econometric Institute Erasmus University Rotterdam P.O. Box 1738 3000 DR Rotterdam The Netherlands |
| Pseudocode | Yes | Algorithm 1: Gen SVM Algorithm |
| Open Source Code | Yes | The source code for this C library is available under the open source GNU GPL license, through an online repository. A thorough description of the implementation is available in the package documentation. |
| Open Datasets | Yes | For the experiments 13 data sets were selected from the UCI repository (Bache and Lichman, 2013). |
| Dataset Splits | Yes | To compare the classification methods properly, it is desirable to remove any bias that could occur when using cross validation (Cawley and Talbot, 2010). Therefore, nested cross validation is used (Stone, 1974), as illustrated in Figure 5. In nested CV, a data set is randomly split in a number of chunks. Each of these chunks is kept apart from the remaining chunks once, while the remaining chunks are combined to form a single data set. A grid search is then applied to this combined data set to find the optimal hyperparameters with which to predict the test chunk. This process is then repeated for each of the chunks. The predictions of the test chunk will be unbiased since it was not included in the grid search. For this reason, it is argued that this approach is preferred over approaches that simply report maximum accuracy rates obtained during the grid search. |
| Hardware Specification | Yes | To run all computations in a reasonable amount of time, the computations were performed on the Dutch National LISA Compute Cluster. A master-worker program was developed using the message passing interface in Python (Dalc ın et al., 2005). This allows for efficient use of multiple nodes by successively sending out tasks to worker threads from a single master thread. Since the total training time of a classifier is also of interest, it was ensured that all computations were done on the exact same core type.4 Furthermore, training time was measured from within the C programs, to ensure that only the time needed for the cross validation routine was measured. The total computation time needed to obtain the presented results was about 152 days, using the LISA Cluster this was done in five and a half days wall-clock time. 4. The specific type of core used is the Intel Xeon E5-2650 v2, with 16 threads at a clock speed of 2.6 GHz. At most 14 threads were used simultaneously, reserving one for the master thread and one for system processes. |
| Software Dependencies | Yes | The implementation used in the experiments described in Section 6 is written in C, using the ATLAS (Whaley and Dongarra, 1998) and LAPACK (Anderson et al., 1999) libraries. The source code for this C library is available under the open source GNU GPL license, through an online repository. A thorough description of the implementation is available in the package documentation. |
| Experiment Setup | Yes | The regularization parameter was varied on a grid with λ {2 18, 2 16, . . . , 218}. For Gen SVM the grid search was extended with the parameters κ { 0.9, 0.5, 5.0} and p {1.0, 1.5, 2.0}. The stopping parameter for the Gen SVM majorization algorithm was set at ϵ = 10 6 during the grid search in the training phase and at ϵ = 10 8 for the final model in the testing phase. |