posted by user: Hassanat || 490 views || tracked by 1 users: [display]

MLCID 2023 : Machine Learning from Class Imbalanced Data

FacebookTwitterLinkedInGoogle

Link: https://www.mutah.edu.jo/Lists/NewAtMutah/Disp_Form.aspx?ID=25
 
When Jul 9, 2023 - Jul 12, 2023
Where Tunis, Tunisia
Submission Deadline Apr 15, 2023
Notification Due May 22, 2023
Final Version Due May 30, 2023
Categories    machine learning   class imbalance   resampling
 

Call For Papers

This workshop will be organized by Prof. Hasanat as part of the 28th IEEE Symposium on Computers and Communications (ISCC),
https://2023.ieee-iscc.org/ .
and it will include presentations, tutorials, and panel discussions to facilitate knowledge sharing and collaboration among participants. This workshop is peer-reviewed and published as part of the symposium proceedings and needs registration fees as per the ISCC.

Problem and scope of the workshop:

Class imbalance occurs when training a dataset with examples from one class vastly outnumbering those from the other. The majority class is commonly referred to as such, while the minority class is commonly referred to as such. A single dataset may contain more than one majority class and more than one minority class. Machine learning models trained on unequal training sets exhibit a prediction bias, which is connected to poor performance in the minority class(es), The bias could range from a minor imbalance to a major imbalance depending on the dataset used.



This problem has grown and becomes a substantial challenge since the minority class is frequently crucial, as it reflects favorable instances that are rare in nature or expensive to collect. This is true when considering contexts such as Big Data analytics, Biometrics, gene profiling, credit card fraud detection, content-based image retrieval, disease detection, Natural Language Processing, network security, image recognition, Anomaly Detection, etc.

There are several approaches to solving class imbalance problem before starting machine learning, such as:

· More samples from the minority class(es) should be acquired from the knowledge domain.

· Changing the loss function to give the failing minority class a higher cost.

· Oversampling the minority class.

· Undersampling the majority class.

· Any combination of previous approaches.

Each of the aforementioned approaches has advantages and disadvantages. Oversampling, on the other hand, is the most commonly utilized approach, as seen by the plethora of oversampling methods developed in the last two decades. However, this does not necessarily imply that the oversampling approach is desirable.

Oversampling methods, according to several recent studies, increase the number of minority-class cases by creating new ones out of out of nothing based only on their similarity to one or more of the minority's examples. This is problematic because such methods increase the probability of the learning process becoming overfitted.

Overfitted synthetic datasets generate impressive machine-learning results on paper, but this is not necessarily the case in practice. Another more serious issue with oversampling is that the generated examples may exist in the real world as members of a different class, regardless of how similar they are to the minority's examples, as demonstrated by a recent study calling for an end to oversampling class imbalanced data.

The main issue with oversampling methods is that these methods assume that all synthesized data belong to the minority class without providing any guarantee, even though some of these generated examples may actually belong to the majority class, making a learned model dangerous in real-world applications, especially crucial applications like security, autonomous driving, aviation safety, and medical applications, where even a single unrealistic false synthetic example can cause fetal harm.

Workshop goal:

This workshop is about creative ideas for class imbalance learning that are based on the aforementioned methodologies, with the exception of the oversampling approach, because the problem is still significant and requires stronger and more practical solutions.

We anticipate that at least ten accepted papers will be presented on the workshop's designated date and that the workshop's chair, Prof. Ahmad Hassanat, will give a talk about the problem and current solutions, as well as lead lively discussions of each accepted paper in order to create a brainstorming discussion based on the provided solutions and their viability.

Related Resources

Ei/Scopus-AACIP 2024   2024 2nd Asia Conference on Algorithms, Computing and Image Processing (AACIP 2024)-EI Compendex
DataMod 2024   12th International Symposium DataMod 2024: From Data to Models and Back
ICMLA 2024   23rd International Conference on Machine Learning and Applications
IEEE-Ei/Scopus-SGGEA 2024   2024 Asia Conference on Smart Grid, Green Energy and Applications (SGGEA 2024) -EI Compendex
DSIT 2024   2024 7th International Conference on Data Science and Information Technology (DSIT 2024)
AIM@EPIA 2024   Artificial Intelligence in Medicine
AMLDS 2025   2025 International Conference on Advanced Machine Learning and Data Science
CCBDIOT 2024   2024 3rd International Conference on Computing, Big Data and Internet of Things (CCBDIOT 2024)
ICONDATA 2024   6th International Conference on Data Science and Applications
ICDM 2024   IEEE International Conference on Data Mining