This code is part of my PhD research at PPG-CC/DC/UFSCar in colaboration with Katholieke Universiteit Leuven Campus Kulak Kortrijk in Belgium. This code test the best hybrid partition chosen - with any of the criteria methods - using CLUS or Random Forests.
@misc{Gatto2023, author = {Gatto, E. C.}, title = {Standard Version of Hybrid Partitions for Multi-Label Classification}, year = {2023}, publisher = {GitHub}, journal = {GitHub repository}, howpublished = {\url{https://github.com/cissagatto/Standard-HPML}}}
This code source is composed of the project R to be used in RStudio IDE and also the following scripts R:
- libraries.R
- utils.R
- misc.R
- test-clus-maf1.R
- test-clus-mif1.R
- test-clus-silho.R
- test-rf-maf1.R (not implemented)
- test-rf-mif1.R (not implemented)
- test-rf-silho.R
- run-clus.R
- run-rf.R
- start.R
- jobs.R
- config-files.R
A file called datasets-original.csv must be in the root project directory. This file is used to read information about the datasets and they are used in the code. We have 90 multilabel datasets in this .csv file. If you want to use another dataset, please, add the following information about the dataset in the file:
| Parameter | Status | Description |
|---|---|---|
| Id | mandatory | Integer number to identify the dataset |
| Name | mandatory | Dataset name (please follow the benchmark) |
| Domain | optional | Dataset domain |
| Instances | mandatory | Total number of dataset instances |
| Attributes | mandatory | Total number of dataset attributes |
| Labels | mandatory | Total number of labels in the label space |
| Inputs | mandatory | Total number of dataset input attributes |
| Cardinality | optional | ** |
| Density | optional | ** |
| Labelsets | optional | ** |
| Single | optional | ** |
| Max.freq | optional | ** |
| Mean.IR | optional | ** |
| Scumble | optional | ** |
| TCS | optional | ** |
| AttStart | mandatory | Column number where the attribute space begins * 1 |
| AttEnd | mandatory | Column number where the attribute space ends |
| LabelStart | mandatory | Column number where the label space begins |
| LabelEnd | mandatory | Column number where the label space ends |
| Distinct | optional | ** 2 |
| xn | mandatory | Value for Dimension X of the Kohonen map |
| yn | mandatory | Value for Dimension Y of the Kohonen map |
| gridn | mandatory | X times Y value. Kohonen's map must be square |
| max.neigbors | mandatory | The maximum number of neighbors is given by LABELS -1 |
1 - Because it is the first column the number is always 1.
2 - Click here to get explanation about each property.
To run this experiment you need the X-Fold Cross-Validation files and they must be compacted in tar.gz format. You can download these files, with 10-folds, ready for multilabel dataset by clicking here. For a new dataset, in addition to including it in the datasets-original.csv file, you must also run this code here. In the repository in question you will find all the instructions needed to generate the files in the format required for this experiment. The tar.gz file can be placed on any directory on your computer or server. The absolute path of the file should be passed as a parameter in the configuration file that will be read by start.R script. The dataset folds will be loaded from there.
You will need the previously best chosen partitions by one of the following codes:
-
https://github.com/cissagatto/Best-Hybrid-Partition (beeing implemented)
You must use here the results generated from the OUTPUT directory in that source code. They must be compressed into a TAR.GZ file and placed in a directory on your computer. The absolute path of this directory must be passed as a parameter in the configuration file. Please see the configuration file example in this source code (config-files directory).
You need to have installed all the Java, Python and R packages required to execute this code on your machine or server. This code does not provide any type of automatic package installation!
You can use the Conda Environment that I created to perform this experiment. Try to use the command below to extract the environment to your computer:
conda env create -file AmbienteTeste.yaml
See more information about Conda environments here
You can also run this code using the AppTainer container that I'm using to run this code in a SLURM cluster. Please, check this tutorial (in portuguese) to see how to do that.
To run this code you will need a configuration file saved in csv format and with the following information:
| Config | Value |
|---|---|
| Dataset_Path | Absolute path to the directory where the dataset's tar.gz is stored |
| Temporary_Path | Absolute path to the directory where temporary processing will be performed* |
| Partitions_Path | Absolute path to the directory where the best partitions are |
| Implementation | Must be "clus" or "rf" |
| Dendrogram | The linkage metric that were used to built the dendrogram: |
| - single, complete, average, ward.D, ward.D2 or mcQuitty | |
| Similarity | Must be "jaccard", "rogers" or another similarity measure |
| Criteria | Must be "maf1" to test the best partition chosen with Macro-F1, |
| "mif1" to test the best partition chosen with Micro-F1, | |
| or "silho" to test the best partition chosen with Silhouette Coefficient | |
| Dataset_Name | Dataset name according to datasets-original.csv file |
| Number_Dataset | Dataset number according to datasets-original.csv file |
| Number_Folds | Number of folds used in cross validation |
| Number_Cores | Number of cores for parallel processing |
- Use directorys like /dev/shm, tmp or scratch here.
You can save configuration files wherever you want. The absolute path will be passed as a command line argument.
This code was develop in RStudio Version 2022.07.2+576 "Spotted Wakerobin" Release (e7373ef832b49b2a9b88162cfe7eac5f22c40b34, 2022-09-06) for Ubuntu Bionic Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) QtWebEngine/5.12.8 Chrome/69.0.3497.128 Safari/537.36
This code may or may not be executed in parallel, however, it is highly recommended that you run it in parallel. The number of cores can be configured via the command line (number_cores). If number_cores = 1 the code will run sequentially. In our experiments, we used 10 cores. For reproducibility, we recommend that you also use ten cores. This code was tested with the birds dataset in the following machine:
System:
Kernel: 5.4.0-136-generic x86_64 bits: 64 compiler: gcc v: 9.4.0. Desktop: Cinnamon 5.2.7 wm: muffin dm: LightDM. Distro: Linux Mint 20.3 Una. Base: Ubuntu 20.04 focal
Machine:
Type: Laptop System: LENOVO product: 82CG v: IdeaPad Gaming 3 15IMH05 serial: Chassis: type: 10 v: IdeaPad Gaming 3 15IMH05 serial: Mobo: LENOVO model: LNVNB161216 v: SDK0R33126 WIN serial: UEFI: LENOVO v: EGCN33WW date: 12/24/2020
CPU:
Topology: 6-Core model: Intel Core i7-10750H bits: 64 type: MT MCP arch: N/A | L2 cache: 12.0 MiB | flags: avx avx2 lm nx pae sse sse2 sse3 sse4_1 sse4_2 ssse3 vmx bogomips: 62399 | Speed: 4287 MHz min/max: 800/5000 MHz Core speeds (MHz): 1: 4264 2: 4240 3: 4254 | 4: 4240 5: 4273 6: 4275 7: 4267 8: 4223 9: 4275 10: 4226 11: 4264 12:4282
Then the experiment was executed in a cluster at UFSCar.
The results stored in the OUTPUT directory.
To run the code, open the terminal, enter the ~/Standard-HPML/R directory, and type
Rscript tbhp.R [absolute_path_to_config_file]
Example:
Rscript tbhp.R "~/Standard-HPML/config-files/rf/jaccard/ward.D2/silho/srf-GpositiveGO.csv"
- Generated Hybrid Partitions
- Best Hybrid Partitions
- Test