Saved in:
| Main Authors: | , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2502.18891 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866910974659788800 |
|---|---|
| author | Zhong, Ziyuan Zhou, Junyang |
| author_facet | Zhong, Ziyuan Zhou, Junyang |
| contents | In this study, we propose an innovative dynamic classification algorithm aimed at achieving zero missed detections and minimal false positives,acritical in safety-critical domains (e.g., medical diagnostics) where undetected cases risk severe outcomes. The algorithm partitions data in a self-supervised learning-generated way, which allows the model to learn from the training set to understand the data distribution and thereby divides training set and test set into N different subareas. The training and test subsets in the same subarea will have nearly the same boundary. For each subarea, there will be the same type of model, such as linear or random forest model, to predict the results of that subareas. In addition, the algorithm uses subareas boundary to refine predictions results and filter out substandard results without requiring additional models. This approach allows each model to operate within a smaller data range and remove the inaccurate prediction results, thereby improving overall accuracy. Experimental results show that, with minimal data partitioning errors, the algorithm achieves exceptional performance with zero missed detections and minimal false positives, outperforming existing ensembles like XGBoost or LGBM model. Even with larger classification errors, it remains comparable to that of state-of-the-art models.
Key innovations include self-supervised classification learning, small-range subset predictions, and optimizing the prediction results and eliminate the unqualified ones without the need for additional model support. Although the algorithm still has room for improvement in automatic parameter tuning and efficiency, it demonstrates outstanding performance across multiple datasets. Future work will focus on optimizing the classification components to enhance robustness and adaptability. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2502_18891 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Dynamic Classification: Leveraging Self-Supervised Classification to Enhance Prediction Performance Zhong, Ziyuan Zhou, Junyang Machine Learning Artificial Intelligence J.0; I.0 In this study, we propose an innovative dynamic classification algorithm aimed at achieving zero missed detections and minimal false positives,acritical in safety-critical domains (e.g., medical diagnostics) where undetected cases risk severe outcomes. The algorithm partitions data in a self-supervised learning-generated way, which allows the model to learn from the training set to understand the data distribution and thereby divides training set and test set into N different subareas. The training and test subsets in the same subarea will have nearly the same boundary. For each subarea, there will be the same type of model, such as linear or random forest model, to predict the results of that subareas. In addition, the algorithm uses subareas boundary to refine predictions results and filter out substandard results without requiring additional models. This approach allows each model to operate within a smaller data range and remove the inaccurate prediction results, thereby improving overall accuracy. Experimental results show that, with minimal data partitioning errors, the algorithm achieves exceptional performance with zero missed detections and minimal false positives, outperforming existing ensembles like XGBoost or LGBM model. Even with larger classification errors, it remains comparable to that of state-of-the-art models. Key innovations include self-supervised classification learning, small-range subset predictions, and optimizing the prediction results and eliminate the unqualified ones without the need for additional model support. Although the algorithm still has room for improvement in automatic parameter tuning and efficiency, it demonstrates outstanding performance across multiple datasets. Future work will focus on optimizing the classification components to enhance robustness and adaptability. |
| title | Dynamic Classification: Leveraging Self-Supervised Classification to Enhance Prediction Performance |
| topic | Machine Learning Artificial Intelligence J.0; I.0 |
| url | https://arxiv.org/abs/2502.18891 |