dpuenteramirez / semisupervised-learning-and-instance-selection-methods

TFG - Semisupervised learning and instance selection methods

Hi 👋, I'm Daniel

Currently I am Pentester and Cybersecurity consultant at Mnemo.

• 🔭 I’m currently working on 👾 Hangman, a game based on the Tauri framework.

• ⛳ My latests finished projects are Semisupervised learning and instance selection and UBUMLaaS

• 🌱 I’m currently learning Tauri (Rust & JS) and Pentesting

• 👨‍💻 All of my projects are available at WIP

• 📫 How to reach me [email protected]

• ⚡ Fun fact geek, passionate, HWPO

Connect with me:

Languages and Tools:

Support:

 

May the code be with us! Or not, just use Copilot :D

semisupervised-learning-and-instance-selection-methods's Issues

A method to compare with

There's a work quite old (2011) that probably we will have to implement.

Instance Selection in Semi-supervised Learning.pdf

Tri-Training change to use the same classificator

Write CNN - Report

Semi Supervised Algorithms

Algorithms to be implemented:

• Self Training -> It will no be necessary, it will be used Scikit-learning implementation, docs.
• #44
• #36
• #45
• #37
• #46
• #38
• #52

Write MSS - Report

ICF algorithm implementation

ENN algorithm implementation

Refactor Report to docs

Due to clarity, the folder will be refactored

Tri-training Implementation

Safe semi-supervised learning

In this paper, the authors analyse the impact of labelling instances and how this can negatively affect to the performance of the final classification.

This reading is for the future ;)

http://www.lamda.nju.edu.cn/liyf/paper/FCS19-SafeSSL.pdf

Support for for non numerical classes in datasets

RNN algorithm implementation

Add legend to all-in-one figures - test_unlabeled

DROP3 Implementation

DROP3 - With and without not correctly implemented

It is worthwhile to recode the whole algorithm

CNN algorithm implementation

Chnage S inicialization in CNN

Unify ACC & MSE graphs of tests with labels based on given percentages

Individual graphs had already been generated for each data set, in order to have a global representation, they should all be unified into one, respectively.

ICF results are 20% down of what is should be

Something in the implementation is wrong

Run test with scikit-learn self-training

After #17, this issue has the goal of using n/2 of the best accuracy possible to try to train the model

Co-Training Implementation

Validation of ENN, CNN, RNN, ICF and MSS algorithms in compare with Weka software

Reading of the theoretical concepts of the Alvar's Engineering Project, focusing on how some of the instance selection algorithms are and how they work.

Write RNN - Report

Reading Lui's 2002 and Jankowski's 2004 articles.

Write ICF - Report

Review of the documentation

Some comments on the memoria.pdf

memoriaAAG.pdf

ENN breaks when running from ICF in testing

When running test for algorithm validation, ENN works properly by itself. But when is called from ICF the code breaks when searching for neighbors of the last elements.

Might be because of how it considers the k neighbors. I will look into it on November 30, 2021.

DROP3 - Report

Noise filters for self-training SSL

Another interesting work, this is for the future, do not read it yet ;)

https://www.sciencedirect.com/science/article/pii/S0925231213011016

Self-Training - Report

SCRUM November - Annex

Add to the annexes the planning that has been followed in the first month of work.

• Introducción
• Sprint 1 - Nov 08, 2021 to Nov 19, 2021
• Sprint 2 - Nov 22, 2021 to Dec 03, 2021

ENN returns empty dataset when working with a small dataset which contains lots of different classes.

Traceback (most recent call last):
  File "testing_unlabeled.py", line 105, in <module>
    main()
  File "testing_unlabeled.py", line 51, in main
    results_dataset = np.array(__evaluate__(dataset=d1, precision=precision,
  File "testing_unlabeled.py", line 88, in __evaluate__
    acc, mse = __train_and_predict__(data_alg,
  File "testing_unlabeled.py", line 98, in __train_and_predict__
    tree.fit(data_alg['data'], data_alg['target'])
  File "/opt/anaconda3/lib/python3.8/site-packages/sklearn/tree/_classes.py", line 903, in fit
    super().fit(
  File "/opt/anaconda3/lib/python3.8/site-packages/sklearn/tree/_classes.py", line 157, in fit
    X, y = self._validate_data(X, y,
  File "/opt/anaconda3/lib/python3.8/site-packages/sklearn/base.py", line 430, in _validate_data
    X = check_array(X, **check_X_params)
  File "/opt/anaconda3/lib/python3.8/site-packages/sklearn/utils/validation.py", line 63, in inner_f
    return f(*args, **kwargs)
  File "/opt/anaconda3/lib/python3.8/site-packages/sklearn/utils/validation.py", line 694, in check_array
    raise ValueError(
ValueError: Expected 2D array, got 1D array instead:
array=[].
Reshape your data either using array.reshape(-1, 1) if your data has a single feature or array.reshape(1, -1) if it contains a single sample.

Find out accuracy level with cross-validation, with 1%, 5%, 10%, 20% and 50% of unlabeled data.

This issue whants to figure out where the classification tends to be optimal or asintotical, due to future work in self-training, cotraining, etc.

Democratic Co-Learning Implementation

Improve Instance Selection Algorithms documentation

Add docuemntation or improve the existing for:

• ENN
• CNN
• RNN
• ICF
• MSS

Report - Instance Selection Methods

• Introduction
• ENN
• CNN
• RNN
• ICF
• MSS

ENN - Remove samples dynamically from TS

DROP3 - Instance Selection Algorithms

• #42
• #43

Co-Training - Report

Improve minor aspects of the source code

I'm reading the source code of ENN and I have the following suggestions:

• Change "flower" by "instance", I know that you checked it with flowers (Iris) but there are more datasets! ;)
• Try to use the PEP8 standard . There are some online checkers.

MSS algorithm implementation

Test Instance Selection Algorithms using KNN against Weka results

Theoretical concepts - Report

Theoretical concepts.

• Minería de datos
• Aprendizaje supervisado
• Aprendizaje no supervisado
• Aprendizaje semisupervisado

Tri-Training - Report

Convert the semisupervised algorithms to classes

The idea is to have each algorithm form a class, which can be instantiated and used with its own fit and predict methods.

Algorithms:

• #60
• #61
• #62

Democratic Co-Learning - Report

Thorough reading of the scikit-learn API, as well as seeing how the transformers and estimators would be programmed.

Co-Training support dataset with two views

The algorithm is not implemented in the right way. It was a known issue, so this is correction of #36

ICF breaks when ENN returns only samples of one class.